Great - Volatility and Correlation in The Pricing of Equity, FX, and Interest-Rate Options (Wiley, 2005) 2

This Page Intentionally Left Blank
In the ~ r i of~ Equity, ~ nFX~ and ~
Options ~ r ~
Risk ~ a n a g e ~ eand n t Analysis. Vol. l : ~easuring and ~ o d e l l i n g Financial Risk

Carol Alexander (ed.)
Risk ~ a n a g e ~ eand n t Analysis. Vol. 2: New ~ a r k e t and s Products CaroE Alexander (ed.) I~plem~ntin Value g at Risk
Philip Best
Derivat~ves D e ~ y s t ~ e Using d : Structured Financial Products

John C. Braddock
Imple~enting Derivatives ~ o d e l s
Les Clewlow and Chris Strickland
Advanced Credit Risk Analysis: ~inancial Approaches and ~ a t h e ~ a t i c~ ao l d e l to s Asses Price and an age Credit Risk
Didier Cossin and Hugues Pirotte
Derivatives for Decision ~ a k e r s : S t r a t e ~ i c ~ a n aIssues ge~ent

George Crawford and Bidyut Sen David F. DeRosa
Currency Derivatives: Pricing T h e o ~Exotic , Options, and edging plications Options on Foreign Exch~nge (rev~sed edition)
David F. DeRosa
The andb boo^ of Equity ~ e r i ~ a t i v(revised es edition) Jack Francis, William Toy and J .Gregg Whittaker Interest-Rate ~ o d e l l i n g
Jessica James and Nick Webber John F. Marshall
Dictiona~ o f Financial Eng~neering

andb book o f ~ybrid ~nstruments: Convertible Bonds, Preferred Shares, Lyon~~, ELKS, DECS and Other ~ a n d a t o ~ ~ o n v e Notes rti~le
Izzy Nelken (ed.)
Interest- ate Option ~ o d e l s :understanding^ Analysing and Using ~ o d e lfor s Exotic Interest-Rate Options (second edition)
Riccardo Rebonato
Volatility and Correlati~n in the Pricing o f E q u i ~FX , and ~nterest-Rate Opt~~ns

Riccardo Rebonato
Derivatives andb book: Risk ~anagement and Control Robert J . Schwartz and Clifford W. Smith. Jr D y n a ~ i cedging: an aging Vanilla and Exotic Options
Nassim Taleb
Credit Derivatives: A Guide to instru~ents and Ap~licat~ons

Janet Tavakoli
Pricing Finan~ial Derivatives: The Finite Di~erence et hod Doming0 A.Tavella and Art Owen
Chichester * New York * ~ e i n h e * i Brisbane ~
Singapore * Toronto
Published in l999 by John Wiley & SonsLtd, Baffins Lane, Chichester, West Sussex PO1 9 IUD, England
~ ~ t i ~ n a01 243 l 779777 I ~ t e r ~ a t i ~ (+44) n a l 1243 779777 e-mail (for orders and customer service enquiries): cs-books~wi1ey.co.uk Visit our Home Page on http://www.wiley.co.uk or http://www.wiley.com
Copyright 0 1999 Riccardo Rebonato All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording, scanning or otherwise, except under the terns of the Copyright, Designs and Patents Act 1988 or under the terms of a licence issued by the Copyright Licensing Agency, 90 Tottenham Court Road, London W1P 9HE, UK, without the permission in writing of the publisher or the copyright owner. Riccardo Rebonato has asserted his right under the Copyright, Designs and Patents Act 1988, to be identified as the author of this work. John Wiley & Sons, Inc., 605 Third Avenue, New York, NY 10158-0012, USA WILEY-VCH Verlag GmbH, Pappelallee 3, D-69469 ~einheim, ~er~any Jacaranda Wiley Ltd, 33 Park Road, Milton, ~ueensland 4064, Australia John Wiley & Sons (Asia) Pte Ltd, 2 Clementi Loop #02-01, Jin Xing Distripark, Singapore 129809 John Wiley & Sons (Canada) Ltd, 22 ~orcester Road, Rexdale, Ontario M9W lL1, Canada
~ i ~o frCongress a ~ C a t a L o g i n ~ i n ~ ~ u ~ L iDatu ca~on Rebonato, Riccardo, Volatility and correlation in the pricing of equity, FX and interest-rateoptions/~iccardo Rebonato. cm. p. Includes bibliographical references and index. ISBN 0-47 1-89998-4 (alk. paper) l. Options (~inance)-~athematical models. 2. Interest rate futures-Mathematical models. 3. Securities-Prices- ath he ma tical models. I. Title. 11. Title: Volatility and correlation. HG6024.A3R431999 332.6323-dc21 CIP ~ ~~ i sbh rC a a~ taLog~~ in~Publicu~on g Datu
A catalogue record for this book is available from the British Library ISBN 0-47 1-89998-4 Typeset in 10112pt Times by Laser Words, Madras, India Printed and bound in Great Britain by Biddles Ltd, Guildford and Kings Lynn. ~anufactur~d fromsustainableforestation, This book isprintedonacid-freepaperresponsibly for which at least two trees are planted for each one used for paper production.
1.1 1.2 1.3 1.4 1. S 1.6 1.7 1.8
~ntroductionand Plan of the Chapter 3 Funda~ental Concepts and Definitions 4 6 Hedging Forward Contracts Using Spot ~uantities Hedging Options: Volatilities of Spot and Forward Processes Definitions 14 A Series of Options on Futures Contracts 18 a Forward-Setting Strike l8 Hedging an Option with S~itching from the Real World to the Pricing Measure 22
2.1 ~ntroductionand Plan of the Chapter 29 2.2 Hedging a Plain-Vanilla Option in the Presence of Constant V ~ l a t i l i t ~ 30 2.3 Hedging a Plain-Vanilla Option in the Presence of T i ~ e - ~ e p e n d e n t Volatility 34 2.3.3 First View 35 2.3.2 Second View 36 2.3.3 Third View 36 2.4 Hedging a Plain-Vanilla Option When the Real-World Process is Mean Reverting 41
viii
2.5 Hedging a Plain-Vanilla Option With Finite Re-Hedging Intervals 44
stant~neo~s and Termina~ ~orre~ations

3.1 Introduction 51 3.2 The Stochastic Evolution of Imperfectly Correlated Variables 52 3.3 The Role of Terminal Correlation in the Joint Evolution of Stochastic 57 Variables Option, One Underlying Asset 58 3.3.1 Case 1:European 3.3.2 Case 2: Path-Dependent Option, One Asset 61 3.3.3 Case 3:Path-DependentOption,TwoAssets 3.4 Generalising the Results 68
4.1 Introduction 73 4.2 HedgingWith a CompensatedProcess:Plain-VanillaandBinary Options 74 4.3 Smile Tale 1: Sticky Smiles 78 4.4 Smile Tale 2: Floating Smiles 80 Facts About Smiles 83 4.5 Stylised Empirical 83 4.5.1 Equities 4.5.2 Interest Rates 85 Exchange Rates 87 4.5.3 Foreign. Features of the Smile-Modelling Approaches 87 4.6 General 4.6.1 Fully Stochastic Volatility Models 88 4.6.2 Complete-M~kets Jump-Diffusion Models 89 4.6.3 Random~Amplitude Jump -Diffusion Models 90 4.6.4 Stochastic Volatility Functionally Dependent on the Underlying (Restricted-Stochasti~~Volatility) Models 9l 4.7 Risk Derivatives for Plain-Vanilla Options in the Presence of 93Smiles
ethodolo~ies for Smiley

5.1 5.2 5.3 5.4 Introduction 97 General Considerations on Stochastic~olati1ity Models The Dupire, Rubinstein and Deman and Kani Approaches Greens Function (Arrow-De~reu Prices) in theDK Const~ction 101 97
100
5.5 The Peman and Kani Tree Construction 104 5.6 Numerical Aspects of the Implementation of the PK Construction 109 5.7 Implementation Results 1 13
6.1 Introduction 129 6.2 The Computational Framework 130 6.3 Computational Results 135 6.4 The Link Between Implied and Local Volatility Surfaces 139 6.4.1 S y ~ m e t (FX) ~c Smiles 140 6.4.2 Asymmetric (Equity) Smile Surface 144 6.4.3 Monotonic (Interest-Rate) Smile Surface 150 6.5 ~ a i n i n g an Intuitive Understanding 153 6.6 No-Arbitrage Conditions on the ImpliedVolatility Smile Surface 16 1 6.7 A ~ o r ~ e d - oExample: ut Pricing Continuous Pouble Barriers in the 174 Presence of Smiles 6.8Analysisof the Cost of UnwindingandRelatedConsiderations About Option Pricing in the Presence of Smiles182 Appendix 6.1: Proofthat Call($,, K ,T, t ) / a K 2 = @ ( S ~ ) l l t : 186
a2
189 7.1 ~ntroduction 7.2 Estimatin~the Risk-Neutral Density Function 195 7.3 Perivation Analytic of Formulae 199 7.4 Results and Applications 206 Range of Possible Applications 213 7.5 on cl us ions and Appendix 7.1 Obtaining the Pensity of the Underlying from Quoted Option 214Prices
5 Introduction 8.1 21 8.2 The Financial Model: Smile Tale 2 Revisited 216 8.3 Analytic Pescription of Mixed Jump-Pi~usionProcesses 220 8.4 A General Framework for Option Pricing in Complete or Incomplete Markets 229 8.5 Finding the Optimal Hedge 235
~ o ~ t e ~ t
8.6 NumericalImplementation of the Britten-Jones-~euberger Methodology 236 Computational 8.7 Results 243 8.8 Discussion of the Results and Possible Developments 249
9. 1 Introduction: Why Mean Reversion Matters in the Case of InterestRate Models 253 9-2 The BDT Mean-Reversion Paradox 256 9.3 The ~nconditionalVlziance of the Short Rate inBDT-The Discrete Case 259 9.4 The ~nconditionalVkriance of the Short Rate inBDT-The ~Ontinuous-Ti~e Equivalent 26 1 9.5 Mean Reversion inShort-RateLattices: The Equi-Probable Binomial Versus the Bushy-Tree Approach 263 9.6 Extension to More General Interest-Rate Models: The True Role Mean of Reversion 267 Appendix 9.1:Evaluation of the Variance of the Logarithm of the Instantaneous Short Rate 269
10.1 Introduction and Statement of the Problem 271 10.2 Constructing the Most General BGM (Market) Model 273 10.3 A ~orked-Out Example: Caplets and a Two-Period Swaption 278 10.4 A ~ o r ~ e d - OExample: ut Serial Options 280 281 10.5 Reducing the Di~ensionality of the BGM Model 10.6 Numerical Results 286 10.6.1 Fitting the Correlation Surface with a Three-Factor Model 286 10.6.2Fittingthe Correlation Surface witha Four-Factor Model 287 10.7 Conclusions 298
11.1 The LinkBetweenInstantaneousVolatilityandthe Future Term Structure of Volatilities 303 l 1.2 A Functional F o m for theInstantaneousVolatilityFunction 306
xi
11.3 FittingtheInstantaneousVolatility Function: ImposingTimeHomo~eneity of the Term S ~ r u ~ t uof r e Volatilities31 1 1l .4 Fitting the Instantaneous Volatility Function: Information from the Swaption ~ a r ~ 3~ 18 t 11 .S Conclusions 327
33
Even a cursory visit to the financial section of a good bookshop in the City should probably be more than enough to test the resolve and optimism of any would-be author of a financial book: the flowof new titles in virtually every area of finance can be described by few adjectives other than unprecedented,phenomenal, stagge~ng and the like. And if the prospective author had considered writing a book in the option area, her optimism would verge on hubris: with so many topics, angles and perspectives covered, can one hope to add anything new and ~eaningful? Yet, after several visits to my local bookshop I decided, over a year ago, that a different book on volatility and correl~tion still needed writing. How could Ijustify such a claim? In order to answer this question, let me begin by explaining what this book does not attempt to be. First of all, the book does not cover the e l e m e n t a ~ aspects of option pricing: the reader is expected to have gathered, from any of the excellent books available, a simple but clear understanding of the (and Scholes) formula. The level of the representative reader I had in mind when sitting at my word processor was such that she could understand with confidence the funda~ental ideas conveyed by a text such as, for instance, Hulls. Second, this is not a book on econometric estimation techniques: there are already literally dozens of works in this area, ranging in quality from the poor to the excellent; furthermore, statistical esti~ation techniques have never been my main professional area of expertise and there would therefore be precious little, besides all the sins I have conmitted in this field over the years, that I could bring to the table. Finally, this is not a book on stochastic calculus applied to finance: the topic is fascinating and can be of great practical relevance, and I have indeed tinkered in the recent past with the idea of trying to write something in this area; but a few very good books that have recently appeared have convinced me that there was very little I could have said differently, and that little certainly not any better. Despite all of the above, I still think that a different book on volatility and correlation needs writing. The reasons for this conviction are manifold: to begin with I believe that some fundamental concepts about correlation and
xiv
~or~wor
(time-dependent) volatility, despite being implied or alluded to in so many of the conference talks one might chance to attend, have still not been fully (or at all) appreciated by a large number of practitioners. What is at stake is not some obscure academic point, but lies at the very heart of interest-rate option pricing, and will inform the users decisions in so far as the choice and dimensionality of pricing model are concerned. This sunders stood heart of darkness of interest-rate option pricing is p ~ i c u l a r l yrelevant today, when the ace-Gaterek-~usiela modelling framework seems to have, deservedly, won day for path-dependent options, but, due to technical difficulties, is still only reluctantly employed in the case of compound and American options. I will try to show that the key to addressing the topic in a theoretically sound and practically useful way is to be found in a careful analysis of the time dependence of the volatility of forward rates. ~nderstandingthese issues will then furnish the reader with the tools to make an informed choice about the trade-offs between the complexities of a high model dimensionality and the accompanying real increase in pricing accuracy. To pre-empt future conclusions, and simplify matters more than a bit, my final thoughts onthe matter will be that, just as money cant buy you love, so high dimensionality is not a good mechanism to buy yourself de-correlation. This discussion is then linked to my second message: I will try to argue that there is a strong parallel between option pricing in equities and FX in the presence of smiles on the one hand, and interest-rate option trading on the other. A lot has already been said about smiles, but, as I shall argue, attention has so far not been of ~ the smile surface, and on ensuring that what a model focused on the e v o Z ~ t i o implies about this feature is at least c o n g ~ e nwith, t if not identical to, the traders t implied ~ volatility ~ e curves. Much as we all expectations about the shape of ~ hope to find it, the Holy Grail of a truly risk-less model arbitrage is very rarely, if ever, found. I will argue that, in reality, a trader can hope to make money from a no^-plain-vanilla options strategy if her view about the future evolution of some ~ ~ - t r key ~ quantities, ~ ~ ~ Z on e which herhedge is based, turns out to be similar to what she assumed when pricing the option. Trying to match a snapshot of prices as observed in the market today is obviously easier than producing a convincing model of how these prices will evolve in the future. Nobody, however, ever said that life had to be simple, and one can draw on the experience accumulated since the early 1990s in interest-rate option pricing in order to avoid falling into the same pitfalls when pricing equity or FX options in the early years of the new ~ l l e n n i u mMore . precisely, a close analogy can be draw model such as, say, the Black, Derman and Toy (1990) day models of Dupire or Derman andKani. The analogy is relevant because the EZDT model can easily reproduce on a given day not only the prices of an arbitrary set of discount bonds but also a set of caplets, and thereby gives the impression that it can explain all that there is to know about these two markets. When first introduced, the EZDT model was a huge improvement on what was
available (the Ho and Lee model could just fit exactly the bond prices; the Vasicek and ~ o ~ - I n g e r s o l l - ~ o models ss could not be guaranteed to price exactly even the discount bonds, let alone the caplets). Despite the fact that, in its days, the BDT was indeed a breakthrough, one soon began to realise that, if the yield curve i n model, and the tern structure of volatilities were evolved over time ~ i t ~ the then the predicted future shape of the volatility universe did not look at all like the world we observe today. ~nfortunately, this discovery was sometimes made at the expense of significant trading losses (and year-end bonuses) by some of the more naive players. Similarly, the smile models of Dupire and Derman and Kani have truly been a breakthough, since, for the first time in the short but intense life of option pricing, they have provided a tool capable of reproducing exactly, if used with some loving care, todays prices of all plain-vanilla options (and, therefore, an arbitrary smile surface). This undeniable strength should not obscure the fact that, very often, the.@ture smile surface they predict is strongly at odds with what is observed today, and has been observed for a long time in the past. I will try to argue that this has important hedging and pricing implications and that the naive user is ignoring this fact at his own peril. In a way (and here I come to my third message), more sophisticated models can be more dangerous than cruder approaches, inthatthey can engender a false sense of security in the inexpe~encedtrader, whowouldprobably treat with more circumspection models that appear clearly deficient from the start (perhaps because they do not account for smiles at all). The problem, if this is the case, clearly lies with the user and not with the pricing approach: better trading results can often be obtained with a poor model usedwith care and common sense, rather thanwith a better modelused as a black box. I shall therefore constantly encourage the reader to take a close look at what lurks inside the box. In particular, I shall argue, forcefully and repeatedly, that fitting prices today, even exactly, is no panacea, and the trader must always have inmind the ~ ~ a ~ c i rather a l , than mathematical or numerical, mechanism giving rise to the observed smile. The requirement that the future smile surfaces, or the future term structure of volatilities, produced by the model should be congruent with the traders expectations clearly imposes an extra burden, both on the modelling and on the parameterisation. But I will show that, in some areas, the goal can indeed be achieved; and in others, where the success is, so far, more limited, the user should at least be extremely vigilant about those option s t ~ c t u r e or s pay-offs that might expose the weaknesses of a modelling approach. I shall present several case studies of complex options (e.g. options with forward-setting strikes, continuous double barriers, etc.) to illustrate these points in practice. My encouragement always to look inside the black box is not prompted by the dubious pleasure of criticising what colleagues have done, but in order to alert the users to the unavoidable hidden dangers and shortcomings of any modelling approach, no matter how elegant, clever and convincing it mightbe. Whenever I suggest new or different ways of tackling some problems, or offer new solutions (as I
often do in the book), I hope that I will treat with the same degree of constructive criticism my own ideas. I have organised the book in three main parts. In Part One I lay the foundations for the subsequent treatment by defining as clearly as possible the fundamental ~uantities (spot, future, implied, average volatilities, etc.); by clarifying the essentially different types of volatility encountered in equity/% and interest-rate option pricing; by hi~hlighting the differences between forward-process and spotprocess option pricing, and their relative strengths and weaknesses; by drawing two f~ndamental distinctions: one between instantaneous and t e r ~ n acorrelation l and the other between volatility and variance; and, finally, by giving a first treatment of the concept of mean reversion, both in the real and in the risk-adjusted worlds. In Part Two I move on to the problem of smiles. I shall distin~uishcarefully between the different m e c h ~ i s m s responsible for the occurrence of strikedependent volatility on the one hand, and the mathematical tools that can explain the phenomenon on the other. I shall argue that the financial analysis should always precede and inform the analytic or numerical approach. In the context of smiles, great emphasis will be laid on option pricing when markets are not complete. Existing models are reviewed, and several new o~ginal approac~es presented. Part Three is then devoted to the case of interest-rate options. I shall re-visit the issue of mean reversion in the risk-adjusted world (first tackled in Part One), and will make the bold claim that its usefulness and correct role is not to provide a dynamics of the yield curve similar to whatwe observe in the real world, but to ensure the time homo~eneity of the volatility structure. I shall show why this latter quantity is crucial in option pricing, and why the real-world v ~ r i ~ n c e is irrelevant. I shall then analyse in great detail the issues of the ~imultaneous c~ibration of interest-rate models in general, and of the race-~aterk-Musiela approach in particular, to cap and swaption prices. I shall present a very general methodology to achieve simultaneous fitting to an arbitrary set of caplet prices and to an exogenous correlation matrix. Some of the material presented in this book has already appeared in academic journals that are likely to have been overlooked by market professionals. I have always avoided, however, a cut-and-paste-job, and I have done my best to weave the relevant topics into a coherent structure. Finally, I have kept the level of assumed mathematical and financial knowledge as simple as possible, given the subtlety and comple~ity of the topic. If the reader does know about change of measures, martingales, Girsanovs theorem and the adon-Nikod~m derivative she will be better placed to apprec ut I have not assumed that the reader does possess any further than a solid foundation in elementary calculus and a clear understa neutral valuation. This choice has made the treatment someti~es it might have been, and sometimes more cumberso~e,but I have preferred to
bore (slightly) a few readers rather than lose many on the way. Finally, as in my previous book, I have found the use of helshe, hisher too cumbersome, and I have therefore written a macro that r a n d o ~ s e s the use of the masculine and feminine pronouns throughout the text. Much as I am proud of this achievement, and eagerly await Bill Gates telephone call to purchase the comme~cial rights over such a fine piece of work, I hope this macro will not be remembered as the most outst~ding (or only) cont~bution of my book, Riccardo Rebon~to London, 26 June 1999
It is a pleasure to acknowledge the help, encouragement and suggestions from friends and colleagues. In particular, I would like to thank Mr Emanuele Amerio for stimu~ating discussions, and Dr James Pfeffer and Mr Sunil Bowry for useful suggestions and for computational assistance. Dr Dariush Mirfenderes~ has coauthored with me a paper on which one of the chapters is based. This paper has been accepted for pu~lication, but has not appeared in the literature yet. I a m therefore grateful to Dariush for his permission to publish this material before it is in the public domain. I would also like to give special thanks to Mr Mike Sherring for useful discussions and ideas regarding the material in Chapter 10, and to Dr Soraya Kazziah for p e ~ o r m i n some ~ calculations related to the same matter. Needless to say, I a m solely responsible for all the errors in the text. It is a pleasure to thank the staff at JohnWiley for the enthusiasm they have shown for this project, and the help they have provided. In p ~ i c u l a r , David Wilsons assistance and John Halls careful and thoughtful copy-editing are acknowledged with gratitude. Finally, I hope 1will be as supportive to my wife in her current book project(s) as she has been to me d u r i n ~ the course of this work. My hea~felt thanks for her valuable encouragement and support.
1.1 1.2 1.3 1.4 1.5 4.1 4.2 6.1 6.2 10.1 10.2,
A Series of Options on Futures Contracts Hedging an Option with a Forward-Setting Strike Evaluating the Present Value of a Forward Rate Contract e~s Evaluating the Present Value of a ~ ~ B O R - i n - A r r FRA Evaluating the Present Value of a Quanto (Dim FRA Process: Plain-Vanilla ~ p t i o n s Hedging with a Co~pensated with a Co~pensated Process: Binary Options A Degenerate Hedging Strategy for a Call Option Static Replication of a Continuous Double Barrier Caplets and a Two-Period Swaption Serial Options
Section 1.6 1.7 1.8 1.8 1.8 4.2 4.2 6.2 6.7 10.3 10.4
The purpose of this chapter is threefold: first, I intend to explain the f u n d ~ e n t a l difference between the treatment of volatilities and co~elationsin the case of equities and FX on the one hand, and of interest rates on the other. In order to do so it will prove necessary to clarify in detail the terminology necessary for future discussion. Second, I intend to show that volatilities and correlations enter not only the stochastic part of the evolution of the financial ~uantities of relevance for option pricing, but also their drift ( d e ~ e ~ i n i s tcomponent) ic if bitr rage is to be avoided. Therefore, volatilities and correlations are (almost) all that matter in so far as option pricing' is concerned. Lastly, I will point out that pricing options by no-arbitrage argu~ents involves changing, when possible, the (unknown) process of the underlying in the real world into a fundamentallynewprocess.Inthis transfo~ation thedeterministic (drift) part may be altered beyond reco~nition, but the stochastic part is left unc~an~ed . a consequence, the estimation in the real world (the only As one, after all, to which we have statistical and empirical access!) of certain q u ~ t i t i e s , namely volatilities and co~elations,is possible and ~eaningful, but of certain others, whichmightwellbedeceptively similar (suchasvariances), is often useless. These concepts will, hopefully, clarify what quantities must be recovered in the simulations of future rates and prices in the g e ~ a n but e different contexts of option pricing and risk management. The technical tools needed to d e t e ~ n the e nature of the ~ a n s f o ~ a t i o from n the real to the pricin relatively complex, involving as they do measure theory, and an understan~ing of the Girsanov theorem and the R a d o n - ~ i k o d ~ m derivative. I have decided to
The statement is correct for option pricing in complete markets. The cases when markets are not complete, or where continuous hedging is not possible, are tackled in Chapter 8.
avoid proving the results needed to introduce these concepts in a mathematically satisfactory manner (a relatively simple exposition can be found in the Appendices of Rebonato (1998a));the reader is referred for more thorough treatments to a variety of texts such as, in approximate order of complexity, Baxter and Rennie (1996), Pliska (1997), Bjork (1998), and Dothan (1990). I have therefore chosen to sacrifice rigor and generality in favour of intuition and simplicity. Even when the results presented are arrived at following mathematically somewhat heuristic uments, it is, however, always thecase, unless explicitly stated, thatthe proofs could be made rigorous if one wished to do so.
The analysis of those aspects of option pricing and risk management that are connectedwiththestochasticbehaviouroffinancial quantities is plagued by imprecise, confusing and often contradictory t e ~ n o l o g y : thetermsforward volatility, future volatility, term structure of volatility, volatility ofa forward or instantaneousvolatility,toname just a few, oftenmeandifferentthings to different practitioners and, sometimes, mean very little tout court. Even the apparentlyuncontroversialtermcorrelation is less clearly understoodthan it should be, since the fundamental distinction between i ~ s t a ~ t ~ and ~ e toe~~si ~ a 2 correlation is often overlooked. This confusion is of more than academicinterest, since it still prompts some practitioners to claim, for instance, that multi-factor models ~ u s be t employed in order to account simultaneously for the observed prices of caps and swaptions; or, more generally, to argue that, for a model to capture imperfect correlation between rates, it cannot be single-factor. In reality, as will be argued at length in what follows, imperfect i ~ s t ~ ~ t ~ correlation ~eous is but one of the ~echanisms that can produce t e ~ ~ de-correlation i ~ ~ Z amongst rates or prices, and the relative merits and shortcomings of using multi-factor models against employing time-dependent volatilities must be carefully weighed and compared. One reason for this rather confusedstate of affairs is that the underlying traded quantities in the equity and FX market on the one hand, and the interest-rate market on the other, are fundamentally different; not surprisingly, this gives rise to different links between the volatilities of the corresponding spot or forward quantities. In order to clarify these issues is it necessary to define the tools of the trade as precisely as possible. We shall therefore start by defining the volatility of a forward rate or price the future volatility of a spot rate or price the future volatility of a forward rate or price. In order to define their meaning, one must, in turn, define theconcept of forward (rate or price). This is done as follows:
a~ental Conce~ts an
. The T-maturityforward price today in a givencurrency of a given security is the strike that gives zero value to a forward contract whereby that particular security is delivered at time T in exchange for strike units of the same currency. . The T-maturityforward exchange rate (expressed as units of currency A perunit of currency B) is the strike that giveszerovaluetoday . to the forward contract for deliveryattime T of strikeunits of currency A against one unit of currency B.
rate today is the strike that gives zero value to a forward contract paying at time T z the difference between the reset at time 1" of the rate spanning the period [T, T z] and the strike itself.
. The T-expiry-
(7"
+ z)-maturity forward
Therefore, for instance, the T-expiry forward price of an equity is the quantity that gives zero valuetodaytothe contract to deliver that stock at time T in exchange for the strike. Note that inthe definitions there is no mention of expectations (either in a mathematical or subjective sense) about the underlying. As a consequence, distributional assumptions in general, and volatilities and correlations in particular, about the underlying do not enter the definition of a forward. W e n m a r ~ e t are s complete, forward prices and rates are uniquely d e t e ~ n e d by no-arbitrageconsiderations. (See Hull(1993) or Rebonato (1998a) for an example in the interest-rate case.) Anyintroductorytext (see, for example, Hull (1993)) showsthat, bynoarbitrage, the forward rates and prices are related to the spot quantities by
FWt, T)= S(t>exp[(r(t,T ) d)(T t)]
equity case
interest-rate case where is the spot price of the equity stock at time t , is the time-t forward price of the equity stock for delivery at time T, is the (T-$)-period spot interest rate from time t to time T is the dividend yield, is the (T-t)-period spot domestic (foreign) interest rate from time t to time T is the spot FX rate expressed in units of domestic currency per units of foreign currency at time t ,
FFX(t,T )
FR(t, T I ,T2)
p@, T )
is the time-t forward FX rate expressed in units of domestic currency per units of foreign currency for delivery at time T , is the time-t value of a forward rate expiring at time T I and spanning the period [Ti ,2721, is the time-t value of a discount bond maturing at time T .
Note that, in the definitions above, r(t, T ) is not the short rate, but the yield of the maturity bond, i.e. exp[--r(t, T ) ( T t ) ]ZE P(t, T ) Whilst Definitions 1 and 2 are fundamentally identical in the equity and FX cases (as one can readily see by thinking of one unit of currency I3 in Definition 2 as the security in Definition l), the similarities virtually stop moving to Definition 3. In order to appreciate more clearly the nature of this difference, in the next section we shall contrast the case of spot hedging forward contracts in equities with the case of hedging forward contracts with spot quantities in the interest-rate case. The fundamental differences that will become apparent will provide a useful framework to extend the analysis to options in Section 1.4, and to explore the added complexity in the specification of the volatility structure in moving from equitiesFX to interest-rate products.
In order to appreciate as clearly as possible the nature of the differences referred to above, let us put ourselves, for the sake of simplicity, in a deterministicdiscounting framework,2 and let us begin by considering how to hedge it^ spot it^ ti ties a series of equity forward contracts on the one hand, and a series of interest-rate forward contracts on the other. More specifically, let us begin from the first case, and assume that we have entered at time to N forward contracts, Xz, . , ., XN, for delivery at times T I ,T2, ...,TN of A I , A2, .. ., AN struck at XI, amounts of the given stock. The present value at time to of each contract, PVi, is given by
In the case of equities and FX, a deterrninistic-discounting framework is often referred to as the constant-interest-rate case. In the case of interest-rates products, however, if all rates were constant (or, rather, deterministic), there would be no uncertainty about the terminal pay-off of a forward contract or of an option. The stochastic changes in value of a derivative contract on an interest rate depend on the future realisation of rates both via the terminal pay-off and via the discounting. The impact on the present value of a derivative contractof the variability in the discounting, however, is very often much smaller than the effect corning from the variability in the future rate or price whose reset determines the pay-off. The term deterministic discounting is therefore used both for equities and in the interest-rate case to describe the situation when the volatility of the discount factor that present-values the terminal pay-off is so much smaller than the volatility of the underlying asset that it can be neglected.
PVi = Ai[FS(to,T i ) XilP(t0, Ti)

= z :
Ai[S(to)exp(r(t0, ri)(Ti to)) XiIP(to, Ti)
= A i [ ~ ( t O ) / P ( ~T Oi,) XilP(t0, T i )
= Ai[S(to)xif'(to,
In order to spot hedge3 this series of positions one can work out the net sensitivity of the forward contracts to the spot price and to the discount bonds. For each contract one can write ~ ~ V ~ / ~= S Ai ( t o ) and ~ P V i / ~ ~T (t )= o , -AiXi
Therefore the net position, Ahedge, to spot hedge the exposure to the stock price of aZZ the forward contracts is simply given by
i=1,N
This is not surprising since, in the equity/FX case, ~ i ~equity e ~ forward e ~prices ~ are the strikes that give zero value to a series of contracts for delivery of the s a ~ e underlying at different points in time. Therefore, if one enters a si~ultaneous position in a series of equity- (or FX-)based forward contracts, inorder to execute a spot hedge one simply has to take a suitable position in the s a ~ e spot underlying (plus the hedges needed for the discounting). Let us now consider the super~ciallysimilar situation where N forward contracts, struck at X I ,X 2 , ,.., XN, and with notional amounts A I , A2, ..., A N , are TI entered at time to on a series of forward rates spanni~gthe periods [ T I , TI], [T2,T2 Q ] , . ..,[TN,TN z ~ ] (Note . that we are not necessarily a s s u ~ n g that Ti 3 -zi = Ti+l, i.e. the payment time of the ith forward contract does not necessarily coincide with the expiry of the (i 1)th; in other words, the forward rates are not necessarily 'spanning' in the sense of Rebonato (1998a).)The present value of each contract, PVi, is then given by
PVi = Ai[FR(to,Ti, Ti
+~
i) Xil~iP(t0, Ti
~ i )
By'spothedging'astrategy underlying asset are taken.
is meantwhereby spot (asopposedtoforward)positionsinthe
In workingout the spot hedges for each of the contracts in the series one therefore finds: ~ ~ V i / ~ Ti) ~ (= tO Ai ,
Note that, unlike the equitylFX case, the partial derivatives are taken with respect to ~ i ~ e rspot e ~ discount t bond prices, Therefore, in order to spot hedge the whole series of N forward contracts, the trader must in general4 take a spot position e~t forwards are the strikes in 2N assets (discount bonds): ~ i ~ e r interest-rate that give zero value to a series of contracts for delivery of ~ i ~ e r underlying e~t inst~ments at different points in time. Even neglecting stochastic discounting, in order to hedge a simultaneous position in N interest-rate forwards contracts one must t&e positions in 2N underlying spot assets (bonds).
So far we have considered simple forward contracts, for the pricing of which volatilities and correlations, as mentioned above, are irrelevant. If one moved to considering series of options on equity or FX forwards, however, then the reasoning would remain very similar. We show below that the only change, as above is concerned, is that the notional of the spot hedge far as the ~ g u m e n t delta e ~ amount e ~ ~ e of ~ tthe underlying*The will simply become a V o Z ~ t i Z i ~ ~ ~ exact amount of delta hedge will in turn depend on the model used to price the option, but it should already be intuitively plausible from the discussion above that, provided the discounting is deterministic, only the volatility of the same spot process should matter for equities or FX. Furthermore, if we assume that the market and dis~i~utional assu~ptions unde~inning the lack (and Scholes) model are correct, we shall show that since the options on the forwards will, in general, have different expiries, the average volatilities out to different horizons will determine the correct hedge. These, in turn, will be shown to give access, under suitable conditions, to the future volatility of the underlying process, Therefore, in order to price options on equity stock or FX in a deter~nistic-discounting regime, what matters is the t i ~ e - ~ e ~ volatili~ e ~ ~ e o f~ the t spot process. In the case of interest rates, on the other hand, the volatilities of (and, possibly, depend in^ on the type of option, the correlations amongst) 2N processes determine the required delta amounts of spot discount bonds. Specifying volatilities for interest-rate problems is therefore intrinsically more complex, and we shall show
If it so happens, as is often the case, that Ti + zi + 1 discount bonds. The conclusions do not change.
==
Ti+l, then one can simply take positions in
amental ~ o n c e ~ and ts
that, in general, it involves speci~cationof the full ~ t i ~ e - ~ e ovar ~ eria^^^ ~ ~ e ~ ~atrix. Finally, it is importa~tto qualify these p r e l i ~ n a r yremarks with a caveat: we have assumed so far that we wanted to hedge forward contracts with spot transactions. We have only done so in order to highlight as clearly as possible the intrinsic difference between the interest-rate and the equity/FX cases. ~e have not implied, however, that these hedges would, in practice, be either the best or even desirable. We shall argue in what follows that forward pricing and forward hedging actually provide a theoretically and practically superior strategy. With the de~nitions provided in the previous sections clearly in mind, we are in a position to clarify and prove the results mentioned above. To this end we shall move the analysis to the simplest possible setting, i.e. we shall consider firstanequity stock (or FX) spotprocess,withtime-dependentvolatility (no smiles) and deter~nisticrates. We shall assume that we are given a series of plain-vanilla European options prices on the same underlying for different strikes and maturities. In a Black and Scholesfra~ework, the stochastic evolution of the underlying ~ ~ process o t in the real world is given by a Stochastic Differential Equation (SDE) of the form: dS(t)/S(t) = p(S,t)dt
+ u(t)dz(t)
(1.1)
where S(t) denotes the stock price, p its real-world drift, CT the instantaneous percentagevolatility of thespotprocessand dz(t) is theusualincrement of a standard Brownian motion. A SDE of the form (1.1) is often referred to in what follows as a diffusive equation, or, more simply, as a diffusion. A s is well known,inmoving from therealworldtothepricing (risk-neut~al)measure, Equation (1.1) becomes dS(t)/S(t) = r(t) dt
+ a(t) &(t)
(14
where all the symbols have the same meaningas in Equation (1.l), and r(t) is the constant short (risk-less) rate at time t. The solution to this SDE is well known to be given by
S(T) = S(0)exp (FT +:vg~
+ rr,,JTc.) +
(1-3)
with o&gT = o(u)2du, and E E N(0, l).5 Using the above, Equation (1.3) can be re-written as
de~nitionsintroduced (1.3')
S(T) = FS(0,T)ertp --$uZVgT ~ ~ v g J T ~ )
and, since at expiry, F ( T , T) = S(T), one can write
F S ( T , T) = FS(0, T)exp (--&FZ~~T oavgJTc.)

See Chapter 3 for a discussion and proof of this result.
(l. 3 9
omp paring Equation (1.3") with Equation (1.3) one can recognise that it implies that the SDE for the forward price displays no drift, i.e. it can be written as
d F S ( t , T ) / F S ( t ,T) = a(t) dz(t)
(1.4)
This observation will become very impo~ant in what follows when we discuss when it is possible to move freely from variance to volatility and vice versa. In the meantime one should note that the percentage volatility of the forward rate, given the assumption of deterministic discounting that we are enforcing so far, is the same as the volatility of the spot process. (This can be seen immediately by applying Ito's lemma to F S = f (S) = S I P , and treating P as a constant.) Therefore, if we denote by a"(t) the volatility at time t of the forward price, in a deterrninistic-disco~nting setting we can put a" = a. We can then draw on these results to say that the present value of each of the options in our basket can be written in either of the following equivalent ways:
~ ~ ( = ~ E[S(T) ~ t j Xi]+P(to, ) Ti)
E[FS(T, T ) Xi]+P(to,Ti)
where E denotes the expectation operator, the notation [ab]+ means rnax [ab, 0 1 , the quantity $ : a is defined analogously to the integral a:vg,and we have used the definitions introduced at the beginning of this section. Looking at Equation (1.5) we can therefore see that the present value can be expressed either as an expectation over the terminal values of the forward or the spot price. ath he ma tic ally, a treatment in terms of the spot or in terms of the forward process is perfectly equivalent. To evolve a forward rate out to expiry, as we have seen, no drift is needed, but we have only been able to arrive at the volatility of the forward price, a", starting from our assumed knowledge of the volatility of the process for the stock because we made the assumption that the discount or was deterministic (in which case, as stressed above, the volatility of the forward price is identical to the volatility of the spot price: a*= a). As an important aside, this way of looking at the pricin of options, althou common and 'pedagogically' more straightforward, is somewhat lopsided in that it assumes that we start from the knowledge of the volatility of the spot process for the stock and from this we derive the volatility of the forward price. To see why this wayof looking at thematter is rather 'perverse' we can begin by noting that, if the discounting were not deterrninistic, then the volatility of the forward price would include a component arising from the volatility of the discount bond that connects the spot and the forward price: FS = SIP. If, for simplicity, we assumed that the risk-neutral SDE for the spot bond were also of
the log-normal form

dP(t, T)/P(~, T ) = r dt
+ v(t, T)dz(t)
(with v(t, T)denoting the percentage volatility at time t of a maturity bond),6 then a straightforward application of the two-dimensional Itos l e ~ m a a~plied ~, = ~ ( ( t P(t, ) , T ) ) would ive for the volati~ityof the to the function ~ ( 17) forw~d price, G,
p is the correlation between the discount bond and the spot process. on (1.G) one can see that in the sto~hastic-discou~ting case the v01 forward price no longer coincides with the volatility of the spot process. an option price, however, by for the implied volatility, we can t a s i ~ ~ number, Z ~ i.e. the CO on of the spot price volatility, of the discount bond vo~atilityand of the co~elationbetween the two ~ r o c e ~ s e s k s formula haswider quation (1-6). The crucial point i s that than the deter~inistic-discountingcase, discount bond is stochastic. (See, for example in this re
the stochastic nature of discoLlnting and embed this knowledge in their quotes of the vanilla option prices. If a naive user were er~oneouslyto believe that these quotes had been made a s s u ~ i n g deterl~inistic discount in^, hewould be implyin from each option price a volatility that he would identify with the volatility of the spot process. If, therefore, this ed himself using a spot position in the underlying, then he would lling the wrong delta amoL~nt of u ~ d e r l y i, ~since he would be misinte~reting a part of the volatility arising from the discount factor. If the the other hand, r e c o ~ n i s e that ~ the option quote is made in the the stochastic nature of dis~ounting into full accoLmt, he would interpret the implied volatility as e volatility of the forward price. efore put in place a ~ e l t a bed in a forward contract withthe same strike expiry date. This st rate^^ W id auto~aticallytake the stochastic natur~ bond into account^ and would therefore provi ge, but also a theoretic all^ more satisfactory on say, for short EIaturities the two volatilities virtua~lycoincide, but becomes f a from trivial for long-dated equity and, even more s his long, but i~portant, di~ressio~ has highlight^^ the impo~antfact that, in general, each plain-vanilla option can, and in practice should, be regarded as a call or a put on a ~ o price, ~ and, ~ as ~ such, r ~ it gives direct infor~ation on the
Note that the ris~-neutral drift of P ( t , T) must also be equal to r because P(t, ? ) is an asset as well.
r a ~ ~ o d e rnuEIeraire n approach.) Traders
average volatilities of the forward prices, not on the average volatility of the spot process. With this important proviso in mind we can return to the case of purely deterministic isc counting and make the last easy step, i.e. establish the link between the quoted prices of plain-vanilla options and the time dependence of the instantaneous volatility of the underlying. We know that, for equities, all the forwards are a function of the same underlying. Therefore, if discounting is deterrninistic, then a series of option prices can give direct inforrnation about the ~ t u ~ s spot process, ~ ~ This e can be~seen ~ more preciselyas follows, volatili~ oft^^ ~ From Equation (1.5) and the de~nitionbelow Equation (1.3) one can see that each optionin our portfolio can bewritten as some function, f, of thetime integral of the (a priori u ~ o w n time-dependent ) volatility of the spot price:
o ~=tf(~a"g(O ~ -+N ) ) EE f (var[S(O -+ N ) ] )
(~~
os(u)2du)
ince we h o w that the solution to the Black pricing formula is a function of the average volatility of the forward rate, one can impute (imply) from the market price the implied Black volatilityof the forward rate. By so doing one can obtain all these variances, and therefore one can construct var[S(l
4
2)] = var[S(O + 2)] var[S(O " + l)]
In the limit as time 2 becomes closer and closer to time l(2 = l write
os(L 1
E)
one can
El2 =
[l
1+E
os(u)2du . h 1 oS(u)"du]
/E
ut we also know that
and therefore
This equation gives the ~ t vol ~~ t i l~ i of ~ the e spot process from time l to time 2 = l E implied by the option prices. Under the assumption of deterministic disco~nting, this volatility is then also the future volatility of the forward process. From the results above one can immediately deduce that if we live in a dete~inistic-discounting Black world (a world, therefore, inwhich, amongst other things, smiles do not exist) and the quantity
is not a strictly increasing function of the equity of the FX option expiry, T , then there is an arbitrage. Finding the arbitrage is very easy: if the quantity gaVg (T2)?Z, is smaller than oavs ( TI )2T1, then we can buy the long-expiry option (option 2) for less money than we receive by selling the shorter expiry option with the same strike (option 3). Let us invest this difference from time to to time T I . Notice that if I include the lending of the cash origi~ating from the p r e ~ u m differential in my overall portfolio, the strategy is self-financing and has zero set-up cost. At the first expiry time, T I ,there are then two possible cases:
l. Option l ends out of the money; in this case I owe nothing to my counterparty, but I still own option 2, which has no intrinsic but some time value left. I can sell this option at time T1 for a positive amount of cash which I add to the rolled-up cash differential. My total balance is strictly positive. 2. Option 1 ends in the money; the terminal pay-out owed to the c o u n t e ~ ~ y must be smaller than or equal to the value of option 2, which I still own and which has the same intrinsic value plus some extra time value. (The estra time value would be zero only if the future volatility from time TI to time T2 had collapse^ to zero.) I still own an option (the one expiring at time T2), plus the cash in the rolled-up money market account in which the premium cash di~erential had been invested at time to.
Therefore, in both cases I have started with zero (including the lending) and I end up with a strictly positive pay-off the result is actually stronger than an arbitrage (which simply requires that I should start with zero, never lose money, and make a strictly positive a ~ o u n in t some-not necessarily all, as in this case-states of the world; see liska (1997)). Note, in passing, that thisreasonin apply to the case of interest rates: caplet volatility, for instance, need not be strictly increasing. ~e shall discuss this point at great length in Chapter 3. In the meantime, let us now compare the situation just described with the case of interest rates. As mentioned be for^, N forward interest-rate contracts truly represent positions in 2N diRerent u ~ d e r l y i n spot ~ assets. Therefore one must specify the volatility of each of these assets, In addition, each of these assets can have a time-dependent a one-varia~le fu~ction (the ~ o l a t i l i tThe ~ . problem is no longer how to spe~ify volatility of the underlying spot process as a function of calendar time), but how
to specify a surface (the volatilities of different forward rates-which cannot be obtained from each other-at different points in time). In addition one also has to specify the (possibly time-dependent) correlation am t the forward rates. The problem therefore grows in complexity: the user ively has to supply a full ti~e-de~endent covariance matrix. We shall see in Chapter 3 that for a discrete-look option ~ r o b ~where e ~ , the pay-off is determined by N realisations forward rates (i.e. of different assets), 0 ( N 3 ) covariances must be (N) covariances for each of the N times when price-sensitive events occur. As discussed inwhat follows, this is, ultimately, the reasonwhy lowdimensionality (where low someti~es means one or two) who1~-yield-curve els have been introduced. clarified these points we are finally in a position to present in the next of the definitions presented so far, and to introduce some section a s~~mmary i~portant new concepts.
1. When, at time t , we speak about a price process (typically, for equities or FX) described by a diffusive equation of the form
(if t thevolatility that appears inthe equation above is called the ~ r e s e ~ t = toj or ~ t ~ (if r te > to j v o ~ ~ t i o f li the ~ spot process. Therefore the future or present volatility of a spot price = a($,t )
In the equities and FX cases the future or present volati~ityof the spot price, especial~ywhen it depends on the future value of the underlying, is often referred to in the literature as the local volatility,. ~e shall m&e frequent use of this term when dealing with smiles. The term instantaneous volatility can also be used, but, in this book, this latter term will ~ a i n l y be used in the context of interest rates. See the de~nition in point 2. When, at time t , we consider a forward price process (typically, for equities or FX) described by a diffusive equation of the form
dFS(t, T ) / F $ ( t , T j = ~ ( F S t ,,T )dt
+ a ( F S , t , T j dz(t)
(1.8)
the volatility that appears inthe equation above is called the present (if t = to) or f u t ~ r e(if t >to) v o ~ ~ t o fi the ~ i~~ o ~price a process. r ~ Note that we are not using the (meaningless and confusing, but nonetheless c o m ~ o n j expression forward volatility:a volatility can be either present or future, not forward; nor shall we ever use the term forward-forward volatility, which probably is usedtomean, if anythin ,the ~uturevolatility of a f o r ~ ~ ~uantity.Note also that an additional a r ~ ~ m e n T t ,, enters the de~nition
of the volatility of a forward price in order to emphasise that forwards of different maturities will not, in general, have the same volatility at time t. Therefore the future or present volatility of a forward price = a ( F S , t , 7 ' )
3, Just as in the case of spot processes, another term that is so~etimes used in the context of forward price processes is 'instantaneous volatility': it is simply another way to denote the time-dependent volatilities in Equations (1. " 7 ) and (13). See point 5 below for its more c o m m o ~ usage.
4. If the volatility is at most time dependent, then the time integral of the square of the volatility of a forward or spot price between time T I and time 7'2 is often ca~led the total v a r i ~ n cvar(7'1, ~, T2), of that forward or spot price:
.If the variance is referred to a spot process, then the volatility in the equatio~ above would not have the T subscript. The term 'total variance' is c o ~ m o n , butonly appropriate if the drift on time, Whilst of the spot or forward price process is at most depe~dent for most choices of numeraires this is indeed the case for prices, it is not ~utomaticallytrue for interest rates. See ~hapters 2 and 9. 5. he square root of the total variance from time 0 to time T of the forward price of expii.ation T,divided by I" is the lac^ ~ ~ r~ ~ ~ ez ~t tai~ z~ i t~~~, k ( for the forward price:
his quantity is also called the average v o l ~ t i l j tthe ~ ~ j ~ ~ l~ io lea t~ i l j t or ~, the ~1~~~ volatjlj~. Note, however, that the average or Black volatilities, as defined above, are only defined when the time-dependent volatility which is integrated is at mast dependent on time, but the term 'implied volatility' can wider context, and simply indicates the number that must be lack formula to obtain the correct price, ~mplied volatilities are discussed at length in Part Two, which deals with smiles. Finally, note carefully that the term 'average volatility' is somewhat misleading, since the qua~tity defined above does not in general coincide with : ' l a ~ ( udu]/T, ) as one would normally expect from the name. the value Despite being imprecise, this usage is so common that it would be pedantic not to use it for the impossibly cumbersome e~pression'square root of the total variance from time 0 to time I" of the forward price of e~pirationT , divided by T ' , The caveat should, however, be kept in mind.
,in what follows, we are

of the form
with a f o r ~ a r d rat
the vol~tilityexperienced by this f o r w ~ rate r ~ of ex e t is always referred to as the ~ r e (if~t = ~ ~ t t ~ ~ vo el ~ o t i l i t~ y o f~ the f~~~~~ ~ rate. In the vo~~tility is s o ~ ~ t ienco~nt ~es
in~tantaneousvolatility of a f o r ~ a r d rate = a(
n what follows the i
term which depended on th
a The disc~ssionin the next section and in ~ h a ~ t e2 rs the dangers in glossing over this i m ~ o r t a n t ~ i s t i ~ ~ t i o n ,
Serial options are options expiring at time t on a forward rate expiring at time T and spanning the period IT, T t], with t T . They are relatively u n ~ o and ~ illiquid ~ ~ n beyond short expiries. They are discussed in detail in Chapter 10, Section 10.4.
id not require to be made in the case of equities or by far the most c o m ~ o n choices of numeraire always give a drift which is at most dependent on time. Therefore, for price processes, what we have r total variance, t u ~ c to~~t, called stochastic variance and variance, o coincide, from time 0 to time T of the are root of the st hastic v ~ a n c e rate of expiration divided by 1 is the ~ l a ~ ~ akr k v e~ ~ l a t i l for i~ the forward rate aBlack([J7) = ~~stoChvar(0, T)/T1
his quantity is also often called the ~ v ~ v o rZ ~ ~t i , l, ~ , , ia ~ ,e or the ~ ~ ~ Z ote carefully that the m a r ~ e t ives information, via the price of caplets, only about th stochastic varianc from time 0 (today) to the ex of the forward rates. o (~iquid) options* are traded with expiry befor expiry of the forward rate. This et fact is funda~ental: it determines to what ~xtent the prices of caplets, in conjunc with the prices of swaptions, can pindown the full covariance matrix. shall examine this aspect epth in C~apters 3, 9 and 10. It is also essential to point out the identification between implied volatility and an be made only if we assume that we truly live in a instantaneous volatility at most time dependent. If the prices o f plain-vanil~a options were to inco~orate effects c o ~ from n ~ more com one can still obtai~, via inversion of the (now inap~rop~ate) an implied volatility-which will, in general, be strike dependent-but one is no longer authorised to a s s u ~ e that this latter tity has any precise relationship with the quantities entering a simple See the discussion in Chapter 4, Section 4 2 e fu~ction that associates with a forward rate of a given m a lack market volatility is called the t e s t ~r ~~ c ~ ~~voZatilit ~re the term s o m ~ t i ~ has e s a ~ i f ~ ~mea~ing r e ~ t in the equity an where it is used to indicate the time dependence of the volatility of the spot or forward price process. This usage is widespread, but, to avoid ~ o ~ f u s i o n , we shall never use the term term structure of volatilities in this second ~ e a n i n For ~ . equities (and spot processes in general) we shall talk instead f ~ ve u~ ~a c t i lei ~ . about the time ~ e ~ e ~ o
Having clarified these concepts, the next two sections will ~ ~ e s e in n t et ail two case studies that illustrate some of the more subtle points in the d e ~ ~ i t i o ~ just presented,
* Serial options, mentioned in the previous footnote, and dealt are the exception to this statement.
with in Chapter 10, Section 10.4,
For the purpose of this example the rate implied by a futures contract (given by (100 Futures price)/lOO) is used interch~ngeably with the corresponding forward rate. This i s not correct, but in this context the sleight of hand does not affect any of the conclusions
he final task unde~aken in this chapter is to tackle the p r o b l e ~ of the transformation from the real world to the pricing ~ e a s u r e and, , in par~icular,of the effect this trans~or~ation might have on drifts and volatilities. This topic is dealt with in the next section.
he first and most funda~ental result in option pricing, when the process for the ~nderlyin is ~a diffusion and markets are complete, is that neither the drift in the ea1 world, nor the investor,^ aversion to risk, if any, enter the valuation formula. f interest rates are ~ e t e r ~ i n i s t ithen c , the ell- own result is that, in so far as option pricing is concerned, asset prices can be a s s u ~ e d to grow at the risk-less rate, and investors to be risk neutral.
When one moves from deterministic to stochastic interest rates these results, a~though in essence still valid, require some qualifications and can be more profitably presented within the context of an equivalent but more complex framewor~. There exists, in fact, one pa~icular unit of account (the so-called money-market account) in terms of which asset prices still grow at the instantaneous riskifferent choices of numeraires, however, such as the pure discount rise to different drifts for the underlying price processes. In the stochastic-interest-rate context it therefore becomes much more convenient to work in a standardised framework that provides very general and 'framereference-free' results for any choice of the unit of account. Such a stand ised framework can bemuch more easily cast interms of the evolution of forward, rather than spot, prices and rates. en this is done a stri laity of results can be obtained for a va of option problems. self-contained treatment of the topic is beyond the scope of thisbook (see ato (l998a) for a concise self-contained derivation, or any of the refermentioned in the earlier sections of this chapter for lengthier and more co~prehensive treat~ents). One can, however, state some fundamental results that have a direct bearing on the esti~ationof volatilities. These ~esultswill also showthat volatilities and correlations are truly all that matter for option pricing in that they describe fully not only the stochastic component of the evolution of the relevant financial ~uantities,as one would expect, but also the no-arbitrage drift conditions (at least provided one is dealing with ~omplete markets and d~ffusive processes for the un~~rlyings). The logical reasoning as follows. Let us consider the price, S, in a given currency of an asset, and let (always strictly positive) price of another asset in the same currency. Then The ratio of price S to price B is called a reZ~tive price,and is
t
. The asset B is called a ~ ~ ~ e r ~ i r e . . The ~uantityl / is called the ~ i sfac~or. ~ o
= &/B,
~ssuming that a whole host of technical conditions are met, one can then state:
. To prevent arbitrage: the time-t expectation of the relative price at a future time T must be set equal to the value of the relative price at time t :
(1.12) ~alculatin an ~ expectation means associating a pro~abilityweight to each of the possible outcomes. One does not know this probability distri~ution a priori; rather, in order to prevent arbitrage, one must make sure to choose the probability
dist~butionsuch that, for the given choice of numeraire, Equation (1.12) holds true. For this reason, i.e. in order to ~mphasise the dependence on the numeraire of the p r o ~ a ~ i l idensity, ty Q, the symbol Q ( B )has been appended to the expectation operator, ELcB,[.] .Finally, the superscript t indicates that the expectation is taken on the basis of the information available at time t ,
. If the usual technical conditions are met, then any quantity x such that its future expectation satisfies property (1.12) above, i.e. such that
is called a ~ a r t i ~ g a ~ ~ .
. Given a non-traded quantity, R, if there exists a traded asset, N , such that the product R *N is a (linear combin~tion of) traded asset@),then the asset N is called the natura^ p ~ y of- R. ~ ~
Consider a forward rate f ( t ,T i ,Ti t )= ( P ( ~ i ) / P ( T t ~ )1)/t = (P(Tj)/ P(Ti+l)1)/t, denoted in what follows for brevity as f = (Pi/Pi+l 1)/t.. eing a rate, fi it is not, in itself, a directly traded ~uantity.However, if one multiplies fi by Pi+l one obtains:
The ri~ht-hand side is a linear combination of traded assets, and therefore f i P i + l is traded, and Pi+] is the natural pay-off of f i .
. Given Result l
andDefinition 5, one can write
Even if extremely reduced in scope and in degree of generality, the results presented above can already provide some powerful insight. The first o~servation is that, by virtue of their being very abstract, they are independent of the particular numeraire chosen (in this sense the approach hasbeen referred to before as 'frame-of-reference-free,). In order to carry out explicit calculations one always has to make some specific choices for f , N and B. Our next goal is to make use of De~nitions 1-5 and of Results 1 and 2 to derive the general expression for the no-ar~itragedrifts of the financial quantities that enter the valuation of an option pay-off, and to show that these expressions purely and only depend on volatilities and co~elations. Note that this i m ~ o ~ a conclusion nt would not be apparent if one worked in lack and Scholes framework, where the risk-less rate appears to be the no-ar~itrage drift of the process for the spot price of a traded asset. shifting attention from spot to forward processes, we shall show below by means of three i~creasinly complex examples the common underlying s t ~ ~ t uof r ethe no-arbitrage drifts for all forward quantities.
It is also i ~ p o ~ ato n note t that the conditions of no arbitra~e entailed the modification of the drift of the underlying financial quantity in rnovin world to the pricin ~ e a s u r eThe . nature of the drift in the real WO have beenpurely me-dependent, ~ e a n - r e v e ~ i nconstant, g, etc.) was nowbe mentioned, and was c o ~ ~ l e t ereplaced ly by the d e t e ~ i n i s t i c d y n a ~ obtain ics above. At the same time, the conditions of no arbitrage that we ed the deter~inistic (drift), and not the stochastic, cornpon the scope of this book, it is not possible to prove that this ally the case, but I hope to have provided at least a plausi~le the state~ent that the ~ o - ~ b i tconditio~s r ~ ~ e alter the drift, b process. on the volatility, of the un~erlying ( If the reader accepts the validity of this state~ent the ~revious results, then the f o l l o ~ i n ~ f ~ n d a ~conclusion ental
1. A s far as complete markets and diffusiv processes are c o n c e ~ ~ ethe d , norward rates or prices can be expressnditions for the und n terms of volatilities 2, Unlike drifts, volatilities and c o ~ e ~ a t i o n are s not altered byth from the real to the pricing measure.
ecause of point 2, esti~ationof these volatilities and co~elationsin the real world (the only one, after all, to which we have statistical access) has a aring on, and is essential for, option pricing. timationofvariances (as opposedtovolatilities) in the real world can e ofnouse for option pricing if the real-world drift contained the state variables, as is the case, for instance, for ~ean-revertingprocesses. however, we know precious little about real-world drifts, it is never safe to use st~tist~cal estimates of ~ ~ a n for ~ option e s pricing purposes, but we should always try to estimate volatilities and correlations. This last point is as subtle as it is important (and ~isunderstood); it is therefore treate~ in detail in the next chapter.
in this chapter three related topics will be explored: first, the hedging of options will be analysed in the presence of constant and time-dependent volatility; second, the differences between the real-world variance of a quantity and (the square of) its Volatility will be examined; and lastly, a first treatment will be given of the topic of mean reversion, both in the real and in the risk-adjusted worlds. This latter topic will be explored at much greater depth in later chapters in the context of interest-rate models (see, in particular, Chapter 9), but the first analysis presented below will provide the framewor~ for the later discussion. As ~entioned in the closing section of the previous chapter, the relevance of the disc~ssion presented below is f~ndamental in several respects: it will indicate which ~uantities observable in the real worldare of relevance are concerned; this will have in so far as option pricing and risk ~anagement on the econometric methodologies needed for the esti~ation e different contexts of option pricing and risk manage it will present a hedging paradox reported in the literature (Crouhy and (1995)) regarding the interplay between instantaneous and aver~ge volatilitiesinthe hedging of o ons; a further discussion of this paradoxwill be presented in Chapter 8, ction 8.8, after introducing the ~ritten- ones and ~euberger approach; it will highlight that the concept of meanreversion requires particul careful hand~ing even in the simpler case of options on equity stock or this will provide a backdrop to the fundamental discussion of ti~e-dependent volatility and mean reversion in the ris~-n~utral world in the case of interestrate options (see Chapter 9). The conclusions drawn in the discussion of this
latter topic will actually constitute one of the fundamental messages that this book tries to convey.
This section establishes the framewor~ for the s u ~ s e ~ u eanalysis nt to be found in the remainder of the chapter. We shall consider an arbitrary diffusive real-world process for a stock price:
S in Chapter l , p(S, t ) is the real-world drift for the asset price, S, and a(S, t ) its volatility. Note that within the set of diffusion processes where the volatilities are not affected by the realisation of any stochastic process other than S, the speci~cation above is the most general, since we have retained the possi~ility of ry dependence on the stock price itselfboththevolatilityand the jumps, however, are allowed(atthis stage). rice the volatility can the depend on the stock price itself, and since the latter is a stochastic ~uantity, volatility in (2.1) is a ~ e t e ~ i ~ i s function tic of a s t ~ ~ ~ ~ quantity s t i c (and time): given a future point in time t , one cannot know today what value the volatility will assume, since the precise realisation of S at time t is unknown. E n this sense thevolatilityitself is therefore stochastic. ont tin gent, however, on the future a certain value, say ST, the volatility can value of the stock price at time t havin~ n this sense the volatility in (2.1) is only assume one particular value, cr(ST, t). E a ~ e t e ~ ~ i n function. isti~ In order to distin~uish this particular origin of the stochastic behaviour for the volatility from a more genera e volatility itself could be a l ) will be referred to as the diffusio~),in what follows the spe resti~ted-stochastic-volatilitycase. ,Section 4.6, for a discussion of the various stochastic volatility first case we analyse is the constant-percentage-volatility case. Therefo~e on (2.1) specialises to
*
dS/S = p(S, t ) dt
+ cr dz(t)
(2.2)
If all the usual relevant assumptions about the financial market are fulfilled (no restri~tions on short sales, no co~missions, no taxes and frictions, etc.), then we are therefore exactly in the original Black and tholes (1973) framewor~. We shall to hedge a plain-vanilla option (a call) of expiry T on the stock itself. ividing the time period [0, T] into an arbitrary large number of time steps, we want to carry out a Monte Carlo simulation of the stock price process. however s7 (see, for what is done in normal7Monte Carlo s i ~ ~ l a t i o n oyle (1977)), we do not assume that the deterministic part of the price
dyna~cs is described by a risk-neutral drift. We are instead goin e ~ to ~ the e ~stock t price evolution (i.e. p(S, t ) = @ ( l ) ) , arbitrary t i ~ e - ~ e ~drift and rebalance, at each time step, a delta-neutral portfolio made up of ack) delta amount of stock and or lent) to make the portfolio zero valued. equired cash (bo~owed We shall call a position in a lack delta amount of stock plus the ending 'the portfolio'. The rule we have chosen, i.e. to hold a nt of stock (plus the invested cash), ch~acterises our tradin require this strategy to be self-financi~~, and therefore we cann subtract ~ o n e y from our portfolio during the life of the option. As w set-up cost of this portfolio is exactly equal to the cost of the option. It is irn ortant to emphasise again that, despite superficial similarities, we ar not out a direct valuation of the option value, As we said, we are goin to s an arbitrary possible version of>the real world, and there for^ we are not invoking risk-neutral valuation. ~e shall instead compare the the option pay-off at expiry time 1" after the arbitrary real-world the stock price has taken place. Ifit indeed turns out to be the case that these two quantities (i.e. the option pay-off and the value of the portfolio at 7") are exactly the same, not only on average, the end of every single path, thenwe shall conclude that the proposed he strategy was indeed correct, and that the set-up cost of the strate t coincide with the 'fair' option price. The crucial condition for the second state~ent to be true is that no money s h o u l ~ be injected into or removed from the strategy during the life of the option; we can, however, rest assured that this will be the case because all the str propose will be self-financing, i.e. will simply entail redistribution option life of the current wealth across different assets. The purpose exercise is therefore to determine which strategies succeed in br -by-path replication of the final pay-off. ince our si~ulations will obviously take place in discrete time, i.e. by evolvin rocess (2.2) over small but finite time steps, we cannot expect the exac coincidence of the terminal option pay-off and of the portfolio value. In order to obtain useful information about the co~ectness of the hedging strategy we shall therefore focus attention on a p~ticularstochastic quantity, x, defined as the difference between the hope full^) replica tin^ portfolio and the option pay-off* The first moment of this variable, i.e. its expectation, will tell us whether we he option pay-offand the terminal value of the are achieving a match bet portfolio, at least on aver examining how the variance of x behaves as a function of the time-step size we shall be able to obtain the more impo~ant basis.I information as to whether the portfolio is replicating on a path-by-~ath
Despite the fact that we are concentratillg our attention on the mean and variance of the quantityx, we are not assuming a mean-variance investor (i.e. an investor with a quadratic utility unction). The
st beused in t la to obtain the correct delta (an that the variance s u of ~ the portfolio and the nitude smaller than the option pri had taken more time steps (as we shall show later on), and it is the tity to oni it or in these si~ulations, since we want to be i~different ortfolio at the end o f every ~ i ~ ~ - ~ y - ~ acondition th is to zero with the number of time then shows (not sur~risin o, which, as one can readily check, is indeed oles value for the option. o ~ s i d e r above e ~ showed little more than the fact that our si~ulation rly, and produced results that s ~ have o s u r~ ~ ~ s no e ~ dreader. ~ owever, in ~ l a ~ i fthe y i~~ e~ t h o ~ o l used o ~ y in themore interest in^ cases analysed in the next section.
1
analysis could be extended to higher ~oments, and indeed the coincidence of the whole distribution of the option pay-off and of the replicati~g portfolio could be analysed, but the topic is not pursued (1990) on convergence in dist~butions. here. See, for example, Nelson and R a ~ a s ~ a m y
et us now consider the case where the process is still a pure random walk, but now the volatility is time-dependent:
dS = p(S, t ) dt
+ o(S, t ) dz(t)
a(t)= a~exp[-utJ
The choice of the simple exponential functional form for the instantaneous volatility is motivated by little more than ~esibility and analytic tractabil~ty. It is ive a particularly realistic representation of market time-dependent far as edging i s concerned, the obvious question is: nstruct the delta hedge? Unlike the constant arenow several plausible candidates: should one use Q? 0-0 expF-uir]? The The square root of the ave~age variance? In order to answer shall a s s u ~ e again that the chosen volatility will be e ~ ~ c t Z y the life of the option (is. durin the sim~lation) and that Ita hedge dyna~icallythe option ~ositionb r the gamma profit from tradin~ the resulti g a m ~ aand , therefo delta-neutral we ble, since to r e ~ a i n stock whenever its price falls and sellin when its price rises. If, on the other hand, the stock moves over the time interval At by less than a certain critical unt (the brea~"even point) we cay (theta) more than we because of its movement; see the first time step the profit or loss is therefore given by amma of the option, atrtradi s the implied the option, and the esperienced variance, G$,, that enters on (2.3) is given by the square of the percentage realised price change: = (AS/Sj2. Note that if we have traded at a 'fair' price in a everal time steps in succession, the overall by the sum of the many (in the limit, infinite) terms of the same form as above:
See Chapter ll for a discussion of flexible and realistic functional forms for the instantaneous volatility of forward rates.
The P& L profile of a delta-hedged long option position. If the lo~arithm of the stock price moves by more than a 4 A t (below 4.47 or above 4.72), then the strategy w i l l yield a positive profit. The stock values S1 = exp [4.47] and S, = exp 14.701 are called the break-even points. If there i s no price move at all over the time step At, then , , , a = 0 and the option w i l l lose the maximum amount (28 in the figure) in time value
If, for a moment, one could assume that the quantity rS2were a constant, one iately conclude that the trade will have been struck at a 'fair' if the two quantities
ince the right-hand side is just the total v ~ i a n c e of the (lo stock price, it would therefore seem that, in the limit as the time ste in order to price and hedge the option all that ~ a t t e r s is the total variance of thestock price, or, more precisely, the quantity assumptio~ that rS2is a constant is, however, patently wro~g. the ~ ~ a n t i rS2 t y is not only time-~e~endent, but also stochastic, does on the path of the stock. This o ~ s e ~ a t i has o n given rise to at least three different 'views' about the nat~re, and the very existence, of an exact delta-hedging strategy, even when continuou~ tra~in is~ allowed, Since these views are frequently p u t forward by practitioners, ~ made t to or presented at professio~al conferences, in what follows an a t t e ~ is explain the u n ~ e r l y i n reasoning ~ in as fair and unbiased a way as then to present the results of tests that will indicate which views are correct,
From the o ~ s e r v a t i o that ~ the quantity rS2is stochastic, one is led to conclude that the total P&L, is the weighted sum of individua~terms of the form [(a2),,,-
were the case, if one hedged the
h to e l i ~ n a t the e fears by the proponents of this view is that of a simple call esence of a strongly i n ~ r e ~ s i n g t i ~ e - d e p e n volatility. ~ent
do matter.
nother possible wayof looking at the problem is to sub~ivideeach trading interval At into further sub-intervals which are small and numerous e n o u ~ h for the stock price (which describes a c o ~ ~ ~path) ~ i ~ u the ~ s gamm~ to be approxi~ately constant over each of these su~-intervals, could then carry out re-hedgin~trades on each of the sub-intervals. If these were suf~ciently numerous, then one could hope that the discrepancy betwee and average volatility c o ~ l d be made as small as we wished. r succession of draws, for the rownia~ motion could only matter freak draw, o ven sub-inte~al, we could only trade a small number of times; if we each sub-interval, a very large numberof times, we could rest assured, accord in^ to the second view, that no abnormal path will occur, and that enough L, will be captured over each. At to make the total P&L path-insensitive.
A third party looks at the problem in yet a diBerent light. To the above-mentioned example of the stock ending deeply out of the money because of the unlucky path, this third camp would reply by noting that if the stock has ended so far out
of the moneyinthe st part of its life, then the trader must have ha l u c k ( h i ~ h volatilit path when the volatility was sup~osed to be lo h-volatility, but low that we can trade E r
elow:
or
f the option with a ( d i ~ f e ~ e same vol~tilityas when the trade
realised volatility) than in
variances as a function of the size of ponents of the first theory we neither hedging s ~ a t e ~ will y an asy~ptotically zero-v~riance portfolio. If this were the case, then investors would e ~ p e r i e ~ c a e te v a ~ a n c e of return in the presence of t i ~ e - d e ~ e n ~ e n aversion would i n ~ ~ e n the c e price they would pay for onclude that the first view is correct if we obs~rveda portfolio variance, ~ e f i n as e ~above, that fails to converge towards zero with the very large. ~e shall apply the same c number of steps b e c o ~ i n g to the other two views by choosi~g the delta amount based on and (2.49, res~ectivel~.
Tdbles 2.2-2.4 and Figures 2.2 and 2.3 show, somewhat surprisingly (see in articular the crucial column VarPort), that the conclusions of the second and the third camps were correct: as the trading interval goes to zero, the difference between the final value of the portfolio and the option pay-off indeed shrinks to zero on a pat^-by-path basis irres~ective of whet~er the option has been edged using a volatility based on the current total residual variance, or the same initial un~onditionalvariance. Since in both cases conver~ence of the variance to zero
~~~~
Portfolio variance vs. number of steps. The conv~rgence of the portfolio variance towards zero for two different values of the average volati~ities (24.65% and 49.30%),corresponding to the cases inTables 2.2 and 2.3
0.1
0.05
I 0 0
20
60
40
80
100
120
140
160
1 I80
Convergence of theportfolio variance. Co~parison in thespeed of convergence obtainedusing Equation (2.4) or Equation (2.4') to determine the delta
ddddddc;
0000000
$?$?$?$?$?g$?
0000000 0000000
0000000 d-d-"d-d-d-d-
0000000
d-vd-d-d-d-d-
0000000
0000000
z. 2 z. z. z. z. 2
0000000 0000000
z. 2 z. z. z. z. z.
0000000 0000000
0000 0000 NVcO\L) NVooQ

T "
F"
$?$?g$?
0000 0000
n h h h
N N N N
0000 V V V V
2 z. z. z.
0000 0000
is achieved, the first theory is proved wrong. Furthermore, as shown in Table 2 5 . 3 ,the second view does not provide a statistically significantlyfaster and Figure 2 convergence than the third. This might be somewhat su~risingsince intuition probably would have su~gested that hedging on the basis of the residual variance (second view) was somehow financially more appealing: with the third approach to delta hedging, in fact, we somehow have to keep track ofwhenwebegan trading each pa~icular option, and compensate over the life of the option excessive profits with insufficient gains, Note also that, at all points in time, the two winning strategies recommend a different delta amount, but that the compensation between the two different strategies takes place, in the limit, along each and every path. We seem to have reached the conclusion that all that matters for pricing an option is the total variance of a process out to expiry, However, since the drift of the stock has been chosen to be independent of the stock itself (and actually, for simplicity, zero), the quantity :J o ( u ) ~ du happens to coincide with the total vari~ce of the ( l o g ~ t of h~ the) stock. This would no longer be true, however, if the stock price drift contained the stock price itself, as would be the case if the process were meanreverting. In the real world we can estimate both the volatility of a time series, and its variance. If the drift is at most time-dependent we h o w o quantities are linked by a simple relationship like Equation ( ,if the stock price were to follow a mean-reve~ingevolution, would expect its spread to be reduced after a finite time, If this is indeed the case, which of the two quantities that we can estimate econometrically using real-world data is the relevant input for option pricing, the (square root of the) unconditional variance, or the volatility? This is the question addressed in the next section.
N
In order to answer the ~uestionposed at the end of the previous section, let us now assume that the stock follows a mean-reve~ing process:
LS(t)) ew dt dS(t) = ~ e v S ~ ~ ( ~ e v
+ oS(t) dz(t)
( 2 . 5 )
where rev^^^ and RevLev indicate the reversion speed and level of the process, and all the other symbols have the usual meaning. By Equation (2.5) the stock pull towards the reversion level whenever it price will esperience a d e t e ~ n i s t i c is above or below it, with a force proportional to the (positive) reversion speed (see Figure 2.4). Since we have alreadyanalysedthe case of time-de~e~dent volatility in the previous section, we shall hold the volatility constant in this set of si~ulations. Note that the total unconditional variance of the stock process is now smaller than it would have been had the drift been purely time dependent. What is the
cocococo
0000 00
000
3000
p$?$?$?
3000 300
$?S$?$?
0000 Q 0 0 0
c=lddd
"MM
z. z. z.z.
0000 O O Q O
3
60
50
+ ~ e aRev n
~
Rando~~ai~ RevLevel
20
10 0
20
40
60
80
100
for the same draws of random numbers. Note how the dispersion of the stock price is more limited for the mean-revert in^ path. The ~nconditionalvariance out to a finite time horizon is correspondjn~ly smaller
me~n-revert in^ process (dS = ~ e v S ~ ~ ( ~ e S)dt v ~ e +vaSdz(t))
The realisation of
a random walk
(&/S
= cr(t)dz(f)) and of a
correct volatility to use in order to price the option? Is it still the total variance r is it the instantaneous volatility? To answer these questions let us set up the hedging portfolio neglecting the total variance, and simply usingthe (constant) volatility o-as the implied volatility to derive the delta amount of stock. Note that since the ti~e-dependent volatility case has already been analysed in the previous section, for the sake of clarity we now keep the volatility constant. Also, since we have already ascertained that the portfolio variance does go to zero with the number of time steps in each simulation, we keep the number of steps constant at a relatively small value. The results are showninTables 2.5 and 2.6. The interesting result is that the correct replication is obtained when the instantaneous volatility, rather than the square root of the variance, is used to determine the delta a m o u ~ of t stock, Since the difference between the integral of the square of the volatility and the unconditional v ~ ~ can c only e be due to the presence of the state variable in the drift term, wecan conclude that the existence of mean reversion (in the real world!) of a stock price process does not affect the value of an option on the stock. This result can be somewhat counter-intuitive, and many practitioners find it dif~cult to accept. The matter is made even more complex because mean reversion (in the ris~~adjusted world!) of the sho~-rate process does affect the value of an option on the short rate. This paradox will be fully solved later on, but, in the meantime, one can begin by noting that the short rate, unlike the stock price, is not a traded asset, and that, even in the int~rest-rate case, it is only mean reversion in the ris~-adjusted world that matters. This topic is addressed in detail
in Chapter 9, but, in the meantime, it is worthwhile repeating the main result obtained so far: the real-world unconditional variance of a financial ~uantity does not necessarily have any bearing on the price of an option3 whose pay-off depends on the terminal realisation. If it so happened that, in the real world, the drift of the underlying quantity were constant, or purely time-dependent, then the statistical estimation of the finite variance or of the volatility would provide e~uivalentinformation. In reality, however, very little is ~ n o w n about real-world drifts, and, as far as option pricing is concerned, the safe option is always to estimate voZ~t~Zi~ies, not variances. Needless to say, if the estimation s, were carried out for risk management, rather than option pricing, ~ u ~ o s ethe full real-world (drift, plus volatility, plus, possibly, jumps) becomes of relevance, but this is not the topic addressed here. Finally, given the results resented in the final section of Chapter l , one can rest assured that the estimation of the volatility (which cannot but be carried out in the real world) will not be altered by the tra~sformationto the pricing measure.
Thus far we have considered the case where re-hedging could be carried out as fre~uently as we wanted, without incu~ing any transaction costs. The reader might surmise that if the hedging intervals are finite, then the trader might experience a non-zero variance of portfolio returns (as indeed was observed in the examples in the previous sections), but that the optimal hedging strategy would still be the same, i.e. would be based only on the value ofthe ~uantity :J ~ ( udu, ) ~ irrespective of whether the volatility is constant or time-dependent. In the light of the discussion in the previous section we should refrain from calling this integral the variance tout court, but, if we restrict our attention to the ris~-neutral world then , the total variance does indeed coincide with (where the drift is purely ~ ( t ) ) the time integral of the square of the volatility. In this section, therefore, and with this proviso clearly in mind, we shall use the term total variance interchan~eably with the expression the time integral of the square of the volatility. A very interesting, and somewhat counter-intuitive, result can however be proven (see Crouhy and Galai (1995)): if the re-hedging interval is fixed, then the o~timal hedge depends neither only on the total variance nor purely on the instantaneous volatility, but on a combination of the two. To understand this result is repeated below, both because of more clearly, the Crouhy-Galai const~ction the intrinsic interest of the problem, and because it is often claimed that the
Wheneverwespeakaboutoptions in this chapter, reference is always made, unless explicitly of path-dependent or co~pound options otherwise stated, to plain-vanilla European options. The case is much more complex, and is treated in Chapters 3, 10 and 11.
conclusions they reach can be of great practical relevance for the choice of volatility with which to compute the optimal hedge. Let us consider a two-period problem in two separate universes: in the first, the instantaneous volatility is 10% over the first period, and 20% over the second; in the second universe the reverse is true. The total variance in both universes is therefore the same. If we could trade (and therefore re-hedge) continuously, we know that the o p t i ~ ahedge l would s i ~ p l y depend on the ~ u a ~ t i ~ ty o= t (0.102 0.202)At, where A t is the length of each time period. Let us impose, however, that we can only trade once in each time period (at the beginning), ~ be ~ ~~ t e ~ Let ~ us e~~ then . t construct and that the length of each time period c a bushy (i.e. non-recombi~ing)tree that locally matches the two ~ o ~ e nof ts the d i s t ~ ~ u to if othe ~ underlying. For simplicity, we shall assume zero interest rates; see Figures 2.5-2.8.
110.51 71
30.48374
Time 0
Time 1
Time 2
.5 The stock price in universe a, where the volatility is equal to 10% over thefirstperiod, and 20% over the second. The stock moves up to Su and d ) / ( ud), down to Sd, with U = Factor, d = 1/ U , Factor = exp(odAt), p(up) = (1 p(down) = 1 p(up)
1 22.1 403
134.9859
110.5171
81.87308
90.~8374 74.081 82
Time 0
Time 1
Time 2
The stock price in universe b, where the volatilityis equal to 20% over the first period, and 10% over the second. All the symbols have the same meaning as in Figure 2.5
Figures 2.5 and 2.6 show the stock prices in the two universes, and Figures 2.7 and 2.8 the option values and the delta amount of stock calculated as (Opt (up) O ~ ~ ( d o w n ) ) / ( ~(u $(down)). ~) Note that the delta amounts do not depend on the probabilities for the up and down jumps (see Rebonato (19!98a), Chapter 6, for a de~vationand discussion), and that they are vastly different in the two universes. For the particular value of the strike chosen, the delta amount at the origin happens to be exactly the same, but, as Crouhy and Galai (CC in what follows) point out, for any other value of the strike it is in general different. (The actual numerical values are somewhat different from CG7spaper because we chose zero interest rates, and a different discretisation scheme.) A. few comments are in order: to begin with, when one lets the time step approach zero, as inmost applications one would certainly want to do, the actual ~iscretisation scheme becomes conceptually irrelevant, and only affects the speed of convergence. Therefore, as At goes to zero, the up state can
34.98588
15.74945 0.?$~1~2
10.51 709
4.734437
0.2~$~~1
Time 0
Time 1
Time 2
igure 2.7 The call option value for strike 100 in universe a and the delta amount Opt(down))/(S(up) S(down)) shown in bold italics (Opt(up) -
22.14028
1.00000
34.98588 10.51 709
9.966799
0
0
0
Time 0
Time 1
Time 2
.8 The call option value for strike 100 in universe b and the delta amount Opt(down))/(S(up) S(down)) shown in bold italics (Op~(up)
be equivalently modelled as S,, = Sold( 1 al/At), S , = Sold exp(az/At), S,, = Sold( 1 az/At $a2At), or in a variety of other asymptotically equivalent ways. In CGs approach, however, the time step cannot be allowed to be reduced at will. Different values will therefore be obtained for the state variable, the option price, and its delta, depending on which of the (only asymptotically equivalent) discretisation schemes is arbitrarily chosen. Given that CG impose a fixed re-hedging interval, these differences cannot simply be swept under the carpet and forgotten. Note that the fixed-re-hedging assumption is crucial and is justified by CG (1995) as follows:
...because of transaction costs and other execution problems, hedges are readjusted discretely, often once a day, or even once a week. The issue of the ap~ropriate volatility measure becomes important in such a trading environ~ent.
With this observation in mind, in Figure 2.9we plot the delta in the two universes for a variety of strikes between the values where the deltas are exactly equal to 1 or 0 for both volatility regimes. Note that, in this figure, there are three distinct linear segments for the delta amount, with diEerent slopes and three intersection points: the two degenerate levels a t which all the nodes in the tree are in or out of the money, giving a delta of 1 or 0, and the at-the-~oney level. If we had subdivided the same trading interval into more and more steps, then the two different delta curves would have crossed at correspondin~lymore and more points, and would have begun progressively to merge into each other. ote again, however, that, given the constant-re-hedging-intervalassumption, this limiting process cannot be undertaken. Therefore, as a first observation, one
1
0.9 0.8
0.6
0.7
c,
0.4
0.3
0.2
0.1
84
74
Strike
igure 2.9 Delta (a)and delta (b). The delta amount of stock to hold at the origin (today) in the two universes
can say that, for the CG effect to be significant, the length o f each t i ~ step e must f~ ~ ethe residual time to expiry. be o ~ t ~ e order s ao ~ e a g n i t u as The second necessary condition for the CG effect to be strong is that the volatility should be significantly non-constant in both universes over the (short) residual life. In the examples above, it either doubled or halved, depending on the universe, in going from the first to the second and last step. This, however, is hardly realistic if the residual maturity is short. If coupled with the first necessary condition ( few, steps), the second requires for the effect to be so important that the trading environment should be one of strongly non-constant volatility, the option-maturit~-to-time-step ratio small, and the re-hedging periods few. This obse~ation gives a first indication that the effect presented by C certainly i~teresting, is perhaps not as relevant for practical hedging purposes. To quote CCs num~ers, if one considers the case of only two subdivision periods the deltas for the two universes are indeed vastly different: 0.90 and 0.68 for a K / S ratio of 0.90, and 0.43 and 0.67 for a K / S ratio of 1.1, Simply adding two more steps, however, already makes the deltas very similar (0.85 versus 0.84 for K / S = 0.90, and 0.65 versus 0.63 for K / S = l. l). To look at the problem in a different light, one could ask two related questions: Would a trader really keep her re-hedging interval constant if she is within a day or two of option expiry, and the spot level is roughly at the money? and Would the same trader really l y over such expect the volatility to be not only time-dependent, but p r e ~ i c t ~ bso, a short trading period? The answers to both questions point, in my opinion, to the conclusion that the CG paradox is very interesting, and surprisingly had not been appreciated until very recently, but poses less of a practical challenge than it is often thought to. The topic will be addressed again in Chapter 8 from a more general angle. There is, however, a second, and, in my opinion, fundamental objection to the CC reasoning. One of the apparent strengths of the CG approach is that the construction of their deltas does not rely on a limiting process, or onan asymptotically correct matching of moments: the delta amounts they obtain in the two universes are exactly the amounts needed to replicate in discrete time the final and i n t e ~ e d i a t e option pay-offs at the various nodes, given the knowledge of the possible realisations of the underlying in the different up and down states. See Rebonato (19984 for a thorough discussion of this construction and, in particular, of its l i ~ t i n g properties. CGs observations are certainly correct. The real interest of the two-state, discrete-time dynamics, however, does not lie in the fact that we truly believe that over each time step only two realisations will be possible for the stock price. After all, if three, instead of two, states were reachable, a possibility that makes as much, or as little, financial sense as the binomial branching assumption, then one would obtain the very ugly result that another asset would have to be added in order to replicate the pay-off with certainty. The true interest of the binomial construction therefore lies not t A t goes to Zero, two in its descriptive realism, but in the fact that, in the l i ~ ias
suitably chosen states are all that is needed to discretise a continuous Caussian process. This justification, however, cannot be invoked in CGs setting, since it is crucial to their argument that the trading interval should remain fixed as the option maturity approaches. This being the case, no special meaning can be attached to the requirement that each move in the bushy tree should only lead to an up or a down state. Needless to say, if three, or more, states had been allowed, then one would have had to introduce correspondingly more securities depending on the same underlyi~g (presumably other options) to replicate exactly all the pay-offs, and the resulting delta amounts of stock would have been different. Given the arbitra~ness, for a fixed time-step size, of the choice of two as the number of states to which the stock price can migrate, and the considerations about the actual tra~ingfrequency close to expiry put forward in the previous paragraph, it seems fair to say that the effect presented by CC, whilst interesting, should not be of serious concern to traders in their hedging practice.
The importance of correlation has often been emphasised, both in the academic literature and by practitioners, in the context of the pricing of derivatives instruments whose pay-offs depend on the joint realisations of several prices or rates. Examples of such derivative products are: basket options: often calls orputsonsome linear combinations of equity indices; swaptions: calls or puts on a swap rate, where the latter is seen as a linear combination of imperfectly correlated forward rates; spread options: calls or puts on the difference between two reference assets (e.g. equity indices) or rates (typical examples could be the spread between, say, the 2- and 10-year swap rates in currency A, or the spread between the 10-year swap rate in currency A and the 10-year swap rates in currency The o~servation that correlation is impo~ant in the pricing of these types of option is uncontroversial. The next logical step, however, which identi~esthe mechanism capable of producing the terminal de-correlation between the rates or prices as the coefficient of ~nstantaneu~s correlation, needs serious qualification. Imposing a correlation coefficient less than unity can certainly produce imperfect correlation amongst the underlyin~variables. But this is neither the only, nor, very often, the most efficient tool by means of which t e ~ i n a Zcorrelation can beachieved. ~ntroducingtime-dependentvolatilities, as willbeshownbelow, constitutes a powerful, and often more realistic, alternative ~ e c h a n i to s~ produce
Needless to say, if one is simply interested in pricing a swaption, working directly with the swap rate, and its volatility, provides a much more efficient route: the (generally lower) volatility of the If, however, one wants to appreciate the relative swap rate indirectly embodies the correlation effect. dearness or cheapness of swaptions vis-&vis the underlying caplets, then the correlation amongst the underlying forward rates does become relevant; see the discussion in Chapter 3 0 , Section 10.3.
the same effect, and the relative effectiveness and realism of the two approaches must be carefully weighed in each individual application. The following sections will therefore show how a non-constant instantaneous volatility can give rise to de- orr relation amongst the underlying variables; how a 'naive' Monte Carlo simulation can be carried out in the presence of time-dependent volatilities and instantaneous correlation; how a much more efficient Monte Carlo simulation can be carried out. The last point will be of intrinsic interest, but, more importantly, it will show what quantities fully describe the stochastic evolution of sets of financial quantities (rates or prices) whenever the finalpay-off of the contin ent claim depends on the realisations of a finite number of price-sensitive events. European, pathdependent, compound and erm mud an options therefore all fall into this category.
Let us consider the evolution of two log-~ormally dist~buted q~antities (rates or prices), which we shall denote by x1 and x2, respectively:
rownian increments, dzl and dzz, can-but need not-be correlated, and, in complete generality, we shall therefore posit
imperfectly
Note that we have e ~ p l i c i t l allowed ~ for the possibility of time dependence in the two volatilities. Also, we have appended an index (l or 2) to the volatility symbol to emphasise that we are dealing with the volatility of ~ i ~ e rprices e ~ t or assets at the same time (not to the volatility of the same spot process at different times). Let us choose a time horizon, T, and let us impose that the unconditional variance of each variable over this time horizon should be exactly the same:
LT
a ~ ( udu )~ =
02(u)~ du
= aX2T
Finally, for simplicity, let us assume that both variables start at time zero from the same value:
Xl(0) = x2(0) = x"(0)
Under these assumptions we can now run two simulations, one with a constant (and therefore, given the assumptions, identical) volatility for both variables and
with imperfect correlation, and the other with different instantaneous volatilities (albeit constrained by Equation (3.1)) but perfect correlation. One can then evaluate the correlation, calculated along the path, between the changes in the (log of the) two variables in the two cases. For the time-dependent volatility case, the same functional form used in Chapter 2, i.e. ~ ( t = )00 exp(-Ut), is employed again. The results of these trials are shown in Figures 3.1 and 3.2. As is apparent from the two figures, the same sample correlation can be obtained despite the fact that the two de-co~elation-generatingmechanisms are very different. Note that, in this very stylised case, where, for clarity of exposition, a very strongly time-varying volatility has been assigned to the two variables, it is quite easy to tell which figure is produced by which mechanism. Note in Figure 3.1 how the changes for Series 1 are large at the beginning and small at the end (and vice versa for Series 2), whilst they have roughly the same magnitude for the two series in Figure 3.2. In a more realistic case, however, where the correlation is high but not perfect and the decay factor U not as pronounced, it becomes difficult to distinguish the two cases by inspection; see Figure 3.3. In theory, it is of course possible to analyse the two time series sep~ately beforehand in order to establish the possible existence of time dependence in the volatility function. Armed with this information, the trader could, in principle, then analyse the joint dynamics of the two variables and estimate an instantaneous correlation coefficient.In practice, however, these statistical studies are
0.14
Time Step
Figure 3.1 Changes in the variables x 1 and x 2 .The two variables were subjected to the same random shocks (instantaneous correlation = 1).The first variable (Series1) had an instantaneous volatilitygiven by o l ( t ) = ooexp(-vt), 0 5 t 5 7, with oo= 20% and U = 0.64 and7- = 4 years. The second variable (Series 2) had an instantaneous volatility ) 00 exp(--v (7- t)),0 5 t 5 T. The empiricalsample correlation turned given by ~ ( t = out to be 34.89%
0.15 0.1
0.05 0 -0.05
-0.1
-0.15
Time Step
re 3.2 Changes in the variables x1 and x 2 .The two variables were subjected to different random shocks (instantaneous correlation= 35.00%). Both variables had the same constant instantaneous volatility of GO = 20%. The empirical sample correlation turned outto be 34.89%
-0.1
'
re 3.3 Changes in the variables x1 and x2. Can the reader guess, before looking below, whether this realisationwas obtained with constant volatilityand a correlation of 85%, or with a correlation of 90% and a decay constant U of 0.2? The sample correlation turned out to be 85%. (See footnote2 for the answer)
fraught with difficulties, and, especially if the instantaneous volatility is mildly time-dependent and the correlation relatively high, the task of dise~tan~ling the two effects is extremely difficult.
The two series were obtained with a Correlationof 90% and a decay constant
U of
0.2.
~nfortunately, the case of mildly varying instantaneous volatility and of relatively high instantaneous correlation is the n o m rather than the exception when we are dealing with the dynamics of forward rates belonging to the same yield curve. The combined effects of the two de-correlating mechanisms are priced in the relative implied volatilities of caps and swaptions (see the discussion in Chapter 10, Section 10.3), and even relatively stressed but still realistic assumptions for the correlation and volatility produce rather fine differences in the relative prices (of the order of one to three percentage points-~appas-in implied volatility). In order to study the relative i m p o r t ~ c e of the two possible mechanisms to produce de-correlation, Table3.1and Figure 3.4 show the sample correlation between log-changes in the two time series obtained by running many times the simulation experiment described in the captions to Figures 3.1 and 3.2, with the volatility decay constant ( U ) and the instantaneous correlation shown in the table. More precisely, the first column displays the sample correlation obtained for a series of simulations conducted using perfect instantaneous correlation and more and more strongly time-dependent ~olatilities (decay constants U of 0.2, 0.4, 0.6 and 0.8); the second row displays the sample correlation obtained with the same time-dependent volatilities and an instantaneous correlation of 0.8; and so on. The important conclusion that can be drawn from these data is that a nonconstant instantaneous volatility brings about a relatively more pronounced decorrelation when the instantaneous correlation is high. In p ~ i c u l a r ,when this latter quantity is zero, a non-constant instantaneous volatility does not bring about any further reduction in the sample correlation (apart from adding some noise), From these observations we are therefore led to draw the important conclusion that the volatility~based de-correlation mechanism should be of greater relevance in the case of same-currency forward rates, than inthe case of equities or FX rates. This section has atte~ptedto convey a qualitative feel for the impact on the sample co~elation of the joint influences of a time-dependent instantaneous volatility, and a less-than-perfect instantaneous correlation. The discussion has been kept at a very qualitative level; in particular, it is not obvious at this point
Table 3.1 The data used to produce Figure 3.4. Note the greater decrease in sample correlation produced by the non-constant volatility when the instantaneous correlation is high
Decay constant Instantaneous correlation 0.4
0.2 0.901 0.973944 0.705876 0.470971 0.330294 0.1 72208 -0.1 2066 0.675758 0.475828 0.332757 0.1 73 1 78 -0.09665 0.573349 0.338425 0.285563 0.1 29632 0.091 779
01
0.7201 0.837356 1 83 0.8 0.6 0.4 0.2 0
0.062323
why the sample, rather then the instantaneous, correlation should be of relevance for optionpricing. The discussion in Chapters 1 and 2 about volatilitiesand variances should justifiably make us rather cautious before deciding on the basis of intuition whichquantities matter when it comes to the price of an option. The purpose of the next section is therefore to identify in a more precise manner what ~uantities affect the joint stochas~ic evolution of correlated ~nancial variables, in so far as option pricing is concerned. The analysis will be carried out by considering a thought Monte Carlo experiment, but the main focus is more on the conceptual part, rather than on the desc~ptionof a numerical techni~ue.Since I havenotassumedan a priori knowledge of stochastic calculus, I have reported some results without proof, and provided very sketchy and formal proofs for others. The references provided are the s i ~ p l e s t (rather than necess~i1ythe best); in particular, Appendix X of Rebonato (l998a) provides probably the quickest self-contained crashcourse of which I am aware to acquire the ~ d i m e n t s of axiomatic probability theory and stoc~astic calculus. For a t r e a t ~ e n t i n t e r ~ e ~ in iat length e between the 4.0 pages pendix and a slim book, I can recommend the course notes of andard, book-length references are then Oksendal (1995), L a and Lapeyre (1991), Neftci (1996) and Baxter and Rennie (1996). If the reader were to fall in love with stochastic calculus, Karatzas and Shreve (199 I) i s properly the bible, but the amount of work required is substantial. Finally, good and simple treatments of selected topics can be found in Bjork (l 998) (for stochastic integrals) and Pliska (1997) (for ~ltrations).
In what follows we shall place ourselves in a purely Black (and Scholes) world; in particular, in addition to the usual assumptions about the market structure, we shall require that the spot or forward underlying variables should be lo~-normally distributed. A s a consequence, we shall ignore at this stage the possibility of any smilesintheimpliedvolatility. These effects are discussedinPartTwo.Our purpose is toidentifywhat quantities are essential in order to carryoutthe stochastic part of the evolution of the underlyin~variables. We shall obtain the fundamental result that, in addition to volatilities, a quantity that weshall call terminal correlation will play a crucial role. This quantity will be shown to be in general distinct from the instantaneous correlation; in p ~ i c u l a rit , can a s s u ~ e very small values even if the instantaneous correlation is perfect. In this sense, the treat~ent to be found in this section formalises and justifies the discussion in the previous sections of this chapter, and in the first two chapters. On the way to obtaining these results we shall also indicate how efficient Monte Carlo simulations can be carried out.
. Let us consider a function o ( t ) that, over an interval [a,b],satisfies the usual suitable regularity and integrability conditions. Let us divide the interval [a, b] into y1 sub-intervals, with to = a, ..., t,-1 = b. Let us further impose that the function a should be piecewise constant over each interval [ t k , tk+l],with k = 0, n l. This being the case, it can more simply be denoted by the symbol a k .The elementary Ito integral of the function o,(t), In(a,6) is then defined by
[,(a, b) =
b
on@)&(S) =
k=O,n-l
4 t k ) [ l W k S d;7(tk)l
where dz(s) is the increment of a Brownian process at time S , and Z ( t ) is the realisation of a Brownian process at time t. If there is an appropriate limit for the sequence of f~nctions 0% ( t ) as the integer y1 tends to infinity, then the It0 integral of this limit function, o(t), is defined as
(For a simple heuristic derivation of the Ito integral see, for example, Rebonato (1998a) or Bjork (1998). For a mathematically more satisfactory treatment, see ksendal (1995).) Once the Ito integral has been defined, a wealth of i~portant results can be obtained. For our purposes, however, we are simply going to need two results:
1 . For each n , the elementary stochastic integral j" , 0
( S ) dz(s) is a well~de~ned stochastic variable, W,. 2, The sequence of stochastic variables W, converges, as n goes to infinity, to a well-defined stochastic variable, W. Therefore, the It0 i ~ t e g ~ a l i t s e ~ be can ~ e g a r as ~ ea~ st0chastic variable, W. We may not know yet the distribution of this variable, but the important result for our i ~ e d i a t e purposes is to know that it exists.
In the next logical step we shall impose an important restriction on the function o(t), which will allow us to obtain some further useful results.
. Let o(s) be a ~ e t e ~ i ~ i sfunction tic of time. Then one can prove that the Ito integral, I ( t ) = $o(s) dz(s), is a stochastic variable which. is normally dist~buted with mean zero, and variance equal to
Note carefully that the second integral is a simple (non-stochastic) time integral.
Then, integrating both sides, we obtain
ut, from the result in Step 2, we know that the distribution of the It0 integral ~ ( s ds. ) ~ Therefore in Equation (3.3) is normal with zero mean and variance it follows that
1 ;
x ( T ) is normally distributed with mean x(0) and variance
1
T
~ - ( s ds )~
. Let us consider
a process of the form
Applying Ito's lemma to the function h ( f ) one readily obtains
Integrating Equation (3.4) from time 0 to time i ' one obtains
Letusassumethatone was interested, perhapsinordertoprice a European option, in obtainingthevalue of theprocess f at time T. Clearly, wecould produce a simulation of the process ln(f) by sub~ividing the interval C O , T]into small sub-inte~als, and evolving the quantity ln(f) according to Equation (3.4) O ) ] . Since the procedure is and subject to the initial condition ln(f)lo = ln[f ( obviously quite laborious (after evolving the process over y1 steps we have simply obtained a single r~alisation at time i ' ) one would be very tempted to make use quation (3.5) in order to be able to obtain directly the value of f at y , looking at Equation (3.5) time (7') we are interested in. ~ n f o ~ u n a t e lby see that, even if the time integral of the drift could be evaluated, one would still of the esponents. have to deal with a stochastic integral as one of the argu~ents b~te-force solution of evaluating this integral by means of a step-b~-step nte Carlo s i ~ u ~ a t iwhilst o ~ , obviously correct, would bring no computational i~provement over the direct simulation of quat ti on (3.4). However, we shall show below that this laborious approach is not necessary, and that we can make use of the results in Steps 1-3 to carry out a single long jump.
identify the terms in Equation (3.4) as
qua ti on (3.3j with the co~espondingterms in

1
rt
rt
Then, after noting that the integrals $ $Q ( u ) du ~ (U j du are purely a function of time and after making use of the results in Steps 1-4, we can conclude that the quantity ln[f(Tj] is normally distributed with v a ~ a n c e equal (r(uj2du. If we call this latter quantity ~ " ~(i.e. 7 . cP2T J $~ ( u jduj, 2 it to therefore follows that
J i
J c
tells LIS that we can simulate a realisation of the variable f all the wa to time T simply by drawing a single normally distributed standard ns of Equation (3.6j we have therefore succeeded in evolving outto time T in one single step, withan obvious reduction in the co~putational bu n, and in the di~ensionalityof the integral w h i ~ h is implicitly solved by the nte Carlo simulation. As ~ e n t i o ~ e above, d we are using the Monte Carlo frame~ork in this discussion more as a conceptual device than necessarily as a practical computational tool. in practice, it might very well be the case that, for certain pay-offs, we do not have to solve the option problem nume~cally, or that other methods (finite differences, computational lattices, trees, etc.) might provide a more efficient computational route. The reasoning has been cast in the simulation context, because it provides, in my opinion, the clearest way to gain an appreciation of the quantities that really affect the value of an option. In particular, we have establish^^ with the example above that the only quantity that matters in so far as the stochastic part o(u)~ du. ~e had already of this E~ropeanoption problem is concerned is establis~ed this result in a more imprecise way in Chapters 1 and 2, but the approach followed here will allow us to generalise the conclusion to a multi-look setting, and, most importantly, to the case of several underlyin~ variables.
se!
The second case that we analyse is that of an option, still on a single price or rate, but whose final pay-off will depend on the realisation of the underlying variable at several discrete points in time. For simplicity we consider a two-look case, the generalisation to more price-sensitive events being straightforw~d. Let us call the two look times T 1 and 772. Once again we could carry out a brute-force Monte Carlo evolution of the process in many small steps from time 0 to time T1 ,and from here continue similarly to 72, after storing in m e m o ~ , the pay-off-sensitive condition encountered on the way at time T I . We already h o w , however, how to evolve the process f from time 0 to time 71:
and, therefore, we ce~ainly want to make use of this equation to carry out at least jump out to time 2 1 , where the price-sensitive condition can be remembered. The remaining question, since we are not interested out a in any event for any i n t e ~ e d i a t e time between T I and T2, is: can W second long jump between the times of the two price-sensitive event rning to the reasoning we followed before, it is clear that the answer is yes, provided that the draw for the second jump is totally independent of the draw for the first:
I
= f(0)exp
i.r {
7.2
IUJp(u) ~ ~ ( u ) dl 2 } ~ ; , p h z l
+ 4 2 J ( T 2 -T d Z 2
where the de~nition of
CThas been extended in an obvious way to
Therefore, in order to value the two-look, path-dependent option on a single underlying asset, we simply have to draw two independent random draws, normally distributed with zero mean and variance G : , ; T I and ; , a (T2 TI).
that if the underlying process is a spot price process, then we can obtain t-implied values for the two variances from the traded prices of plainvanilla options expiring at times T1 and 7'2, as described in Chapter 1. However, if f ( t ) represented a forward-rate process, then the only plain-vanilla option that gives direct i n f o ~ a t i o n about one of the variances above would be the caplet iring at time T2 (assuming, that is, that the option expiry, 7'2, coincides with the expiry of the forward rate). No liquid i n s t ~ m e n t exists to provide information about the variance of theT2-expiry forward rate at time .1'2 A s emphasised in Chapters 1 and 2, the price of the caplet expiring at time T1 provides, in fact, i n f o ~ a t i o n about the variance of a d i ~ e r e forward ~t rate (i.e. the T~-expiry forward). This quantity does not provide any direct information about the variance of the T2-expiry forward rate from time To to time T I . When dealing with the processes for forward rates, the user must therefore make some assumptions about the time-behaviour of the volatility. A thorough discussion of this extremely important aspect is presented in the final section of this chapter and in Chapter 10.
now move on to the more complex case of a path-dependent option, whose nal pay-off is a function of the joint realisations of two state variables at two different price-sensitive look times. If changes in the two prices or rates were totally ind~pendent of each other, then clearly the problem could be solved trivially by applying twice the approach shown in Case 2 above. We shall instead require in what follows that the increments of the two processes should not be instantaneously uncorrelated. The processes of the two underlying variables are
nce again, the brute-force approach, by means of which the two processes are evolved over many small steps from time 0 to time T I(where the price-sensitive condition is recorded) and then to time T2, is still possible. If one wanted to proceed in this manner, one would simply have to carry out the simulation by drawing, at each step, a first draw, 21, used to shock the first variable; and then to make sure that the second draw, 22, is given by
(0,l)-distributed variable drawn independently of 21. We want to explore, however, whether the long jump procedure is still applicable. More
precisely, in the light of the results obtained above, we would be very tempted to make an identification similar to the one made in Step 5 of Case l , and write
with
;52
= P*ZI
+ -J( l -p*2)W
and W independentlydrawn from a N(0,l) normaldistribution.What is not " should be equal to: the ti~e-average obvious, however, is what the quantity p of the instantaneous correlation definedby (3.8)? The value TI)? It will actually turn out that, in general, neither answer is correct. To obtain the correct answer, let us pursue the following reasoning.
. Let us re-write Equations (3.7) and (3.7') in such a way that the same correlation structure between the variablesx1 and x2 is retained, but the equations are expressed in terms of two i n ~ e ~ Brownian ~ ~ ~ motions, ~ n t dz1 and dw, This can be easily accomplished by writing
with E[dz(t), dw(t)] = 0 It is straightforward to verify that, indeed,
(See, for example, Rebonato (1998a)J
. Integrate Equations (3.9) and
(3. IO) to obtain (3.l l )
(3.12)
that the From the discussion above we already know
stochastic integrals . 1 ( 4 dz1 (U), Sor,a2(U)P(w) dZlf4 and a2(4dU P ( W 2d ~ ( are ~ ) three stochastic variables with means and variances given by
l ?
so?"
Given the independence of dw(t) and dzl ( t ) it also follows that meanixl ( T ) ] = x1(0) mean[x2(7')] = x2(0) (3.13) (3.13') (3.14) (3.14') Therefore the variables x1( T I ) and x2(T1) are jointly nomally distributed with means and variances as above. What r e ~ a i n s to be evaluated is the correlation between the two variables.
. The correlation we seek to evaluate is given by
From the de~nition of covariance one can write
where in the last line use has been made of E~uations (3.1 l), (3.12), (3.13) and (3.13'). One can now return to the de~nition of the Ito integral in Step 1 of Case 1 ; write each integral as the limit of a discrete sum; recall that dz and dw are independent, and therefore E[dz dw] = 0; recall the formal rules of stochastic calculus, namelyE[dz dt] = 0 and E[dzdz] = dt; return to the limit as the partition becomes in~nitely fine. When all the algebra is done ( ebonato (1998a) provides a less sketchy formal de~vation), and the i~enti~cations analogous to those in Step 5 of Case 1 are carried out, one finally obtains the fundam~ntal result
(3.15) with
and
Equation (3.15) is (note carefully that the integral only extends out to time TI). all we need to carry out the simulation by using two long jumps per variable. The prescription for the evolution of the two variables is therefore as follows: 1. sample two independ~nt mean-zero, unit-variance normal random variables, Z1 and W; 2. calculate ~ " ( 7 ' 1 ) as
4.
evolve f 1 and
f 2
according to
At this point the outcome of the price-sensitive event can be recorded and the two variables can be further evolved to time 7'2 using an identical procedure. The crucial quantity p"(7'1) to be used to produce the joint evolution of the variables is therefore neither the correlation at time 7'1, nor the simple average of the instantaneous correlation out to time 7'1, but a weighted average of the instantaneous correlation
P*(Tl) =
~
h(U)~(U)du
with the time-dependent weight, h(t), given by
This funda~ental quantity p "is impo~tant enough to be given a name of its own, i.e. the t e r ~ i n a l c ~ r r e l ~ t i ~ n . ~ u ~ ~ ~ i s ini order n g :to evolve the two stochastic variables from time 0 to time 7'1 all that is needed, in so far as the stochastic part of the evolution is concerned, is knowledge of the quantities
~~' crl ( u ) du ~
These are the elements of a 2 x 2 covariance matrix. uring the following computational step, one can distinguish two cases: if the un~erlying variables were the spot prices of two distinct assets, then the estension is obvious, and the quantities above clearly become
'I%e four prices of plain-vanilla options on the two underlyin~s expirin T Iand T2 allow the recovery of the market assumption for the tim of the (squares of the) individual volatilities. The trader would 'only' have to provide a 'guess, for the instantaneous correlation p(t). If the variables f 1 and f 2 described forward rates, however, then the problem is more subtle, andtwo further distinctions must be drawn: if T I happens to be the expiry time for the forward rate f l, then this forward rate would reset will be degenerate (1 x l) at time T I , over the next step the covariance ~ a t r i x and the only quantity of relevance will be : J C F ~ (du. U )If, ~ on the other hand, T I is not the expiry time for either forward rate, then they will both have to be evolved over the next time step, and all the above quantities in the second (2 x 2) covariance matrix will be relevant. It is also impo~ant to point out that no liquid market instrument will give any information about 02(u)~ du or about : J C F ~ (du U (the ) ~ price of a T2-maturity caplet will provide information about : J 02(u)~du);if T Iis the expiry of the T ~ - ~ a t u r iforward ty rate, then the price of the ~1-maturity caplet will give direct information about 01 ( u ) du. ~ Otherwise, no liquid market instrument will give any information about the value of this integral either. The trader, much as in the one-for~ard-rate, path-dependent case, will have to make suitable ass~mptions about the time-behaviour of the instantaneous volatility. These are discussed in Chapter I 1. Before discussing and general is in^ these findings, it is useful to say a little more about the properties of the terminal correlation. First, the followin limiting cases should be noted:
J t '
l. If the volatilities are constant, then the terminal correlation simply coincides
with the average correlation over the period. 2, If the first volatility is exactly zero whenever the second volatility is nonzero, and vice versa, then and only then can the terminal correlation be zero with a strictly positive instantaneous correlation. 3. If the instantaneous correlation is positive, then the terminal correlation cannot be negative.
In general, the terminal correlation can be lower than one, or, aswe have seen in the limit in^ case 2, even zero, even if the instantaneous correlation is perfect. Since, as we have seen, what matters for pricing is not the instantaneous correlation in isolation, an effective de-correlation can be effectively obtained by allowing for a time-dep~ndent volatility. A s we saw in a qualitative manner in Section 3.2, this mechanism is particularly effective when the instantaneous correlation is high, and becomes progressively less important for lower and lower values of the instantaneous correlation coefficient. Therefore, as mentioned above, non-constant instantaneous volatilities are likely to play a
more impo~antrole in the case of same-cu~encyforward rates. A s we shall see in Chapters 10 and 11, they actually provide an essential ingredient, in this latter context, for a coherent simultaneous description of the cap and swaption markets.
he case studies presented in the previous section have dealt with at most two underlying variables, evolved over at most two steps. It is, however, not difficult extend and generalise the results to the case of N. variables, evolved over rice-sensitive events. In particular, if one is dealing with forward rates, there are likely to be exactly as many forward rates in the problem as there rice-sensitive events. In a very common situation encountered in the pay-off of path-dependent L I ~ ~ R - b a s derivatives, ed each of the price-sensitive events is dete~ined by the reset value of one forward rate. Since this setting is pa~icularly common, in what follows we shall deal with the case of n forward rates, evolved over n look times. The modi~cationsneeded to deal with the case where the umber of events is not equal to the number of variables are trivial. Consider a collection of n forward rates, fi, i = 1,2, .., n. For each forward rate let us make the log-normal assumption:
dfi/fi
= pi dt
+ ai(t)dzi
Let us label the times of the price-sensitive events by t(l), t(2),...,t(n), and t y defined as today's time by t(0).Consider then the ~ ~ a n t i CZ,
Therefore
c;j=
J( r ) l) t(r-
d u ) q (u)P~j(u) du,
r5 n
etc. Notethat all the elements of Cfi are, in general,3 non-zero, and that the elements C:j can be arranged in an n x n r e a l - s y ~ e t r i c matrix, which we shall
For strictly positive volatilities, the only case when they could be zero between two forward rates was exactly zero.
is
when the correlation
At time t(1) we have assumed that one forward rate, which we shall conveniently label f l , will reset. Therefore, over the next step only ( n 1) forward rates will have to be evolved. The real-s~mmetricmatrix 1) x (n 1). At the last time step there wil be of dimension (n ~ giving rise to a degenerate 1 x element C i n = t(n1 on( u ) du, This succession of ti~e-dependentmatrices of decreasing size, fully describes the stochastic evolution of the n forward rates. By generalising what we saw in the previous section, this succession of matrices is all that is needed for the stochastic part of the Monte Carlo simulation of the problem. Other techniques, perhaps even closed-formed solutions, might be able to reduce the computational burden, but, whatever nu rical or analytical approach is chosen, the relevant quantities, i.e. the matrices Indeed, if one wanted to solve the problem by a series of long jumps in a Monte Carlo framework, one would be faced with the problem of generalising E~uations (3.9) and (3.10), i.e. with the task of expressing the dynamics of k correlated variables in terms of k indepen~ent Brownian increments. The general n for solution to this problem is known as the Cholesky d e c o ~ p o s i ~ o(see, example, Kreyszig (1993)). The algorithm is both fast and simple to implement. Alternatively, one could orthogonalise each covariance matrix, and work with the resulting eigenvalues and (now orthogonal) eigenvectors. Let us gain a deeper understa~ding of the structure of these matrices. To this effect, let us define new matrices of constant size ( E x n ) by adding zero values co~es~ondin togthe forward rates that have reset, and arranging the elements as follows:
1 ............
0 0 0 0 0 0 ............
* * * * * *
If we call all the matrices framed with zeros C", then we can form the (n x n ) matrix C i d d simply by adding Therefore the (ik)th element of the j t
+ +
t(j>
W
~
8t.i-1)
~ ~ ( ~ ) c T du ~ ( ~ ) P ~ ~ ( ~ )
= ~~) CT~(u)CTk(u)~~~(u) du
matrix contains the total terminal covariance Therefore each element of the from time t(0) to time t ( j ) between variables i and k. In particular, if i = j = k, then the element of the matrix simply gives the total variance of the kth forward rate. Only in this case, therefore, can it be directly accessed from the market price of the caplet on the kth forward rate: in general, even in a perfect lack world, the user must assign O(n3) covariance elements, andhas ~ i ~ e access from the market to only O(n) values. The prices of swaptions do add some information, but, as discussed more precisely in Chapters 10 and 11, not enough to pin down uniquely the O(n3)quantities that describe the evolution of the forward rates. n the other hand, as we have argued, all that matters in option pricing (for discrete-look securities) are just the elements of these covariance matrices. One might therefore be tempted to dispense altogether with the a priori specification of a model, i.e. of a black box that fills in the gaps in the matrices, and to estimate, either econometrically or on the basis of trading views, the necessary (time-dependent) volatilities and correlations. Much as it would be tempt in^ to define a model by directly assigni~g the elements of the n covariances, the size of the problem, and the dispro~ortionbetween the directly-market-accessible quantities and the number of unknowns, is such that, unless one did so in some highly structured and systematic way, one could virtually be certain to introduce inconsistencies in the overall d y n a ~ c s The . highly structured and syste~atic procedure for assigning the covariance elements is, however, nothing more than a model in disguise, or, more correctly, a modern specificatio~ of an interest-rate model. These are exactly the topics that we shall investigate at greater length in Chapters 10 and 1l. The first three chapters (Part One) of this book have dealt with the most im~ortant ~e~nitions and concepts regarding varian~es,volatilities and correlations inwhat has so far been assumed to be f~ndamentallya Black world,i.e. a world of log-normal underlying assets and rates, complete and friction-less markets, and volatilities at most time-dependent. The real world is, ~un)fo~unately, considerably more complex. This new dimension of complexity, su~sumed under the blanket term of volatility smiles, is dealt with in the next four chapters, which constitute the body of Part Two.
idd
for all its impe~ections and inadequacies by means of a skilful doctoring of the only relatively opaque parameter, i.e. the volatility of the underlying. This is perhaps where the luck of the modelplayed a significant part by turning this possible w e a ~ e s into s a strength: had several inputs being similarly opaque, or, at the other extreme, had all the inputs been totally transparent and unquestionable, there would have been no easy way for traders to express their market views by quoting a single number, i.e. the implied Black volatility. trader does not think of himself as making a marke the exotic or pl~n-vanilla the funding rate (despite the fact that interest rates are stochastic), or to have taken a position in the fatness o r the curve (despite the fact that real returns are more leptoku~ic than a log-normal dist~bution would imply) but can straightforwardly see himself as m a ~ n a g market in volatility.The accompanying folklore of breakeven points, normal-point equivalents, risk-reversals, etc. directly stems from this conceptualisation of the option problem. For these reasons it has therefore become established market practice to quote the price of plain-vanilla options using the implied lack volatility. This practice is followed both in the equity, FX and interest-rat as. The market chooses to retain the simplicity and convenience of a Black-like quote by assignin different implied volatilities for options with identical underlying and expir tion, but ~ifferent strikes. The dependence on the strike of the implied volatility for options of the same matu~ty is referred to in what follows as the smile, irrespective of the precise shape of the volatility/strike curve, Terms such as smirk, frown, etc. are sometimes used in the literature, but in what follows all these different shapes of the curve are simply called smiles. When the collection of s ~ l curves e for different maturities is considered, we shall often S smile surface. It is essential to stress from the beginning that this practice is purely a shortcut to quote a price, and does not directly imply anything about the process of the un~erlying (or about an equivalent or compensated process). To make the point absolutely clear, the next section will present two simple but important studies, i.e. the delta and kappa hedging of a plai~-vanilla option, and the ng of a cash-or-nothing option, both in the presence of smiles.
Theexpressiona30-deltaoptionindicatesaplain-vanillaoptionwithastrikesuchthatthe Black option delta is +30% of the notional if the option is a call, and c-30% if the option is a put.
77
A second example can make the same point even more clearly,
ht of these comments, one can provide the following a p h o ~ s m as the best description of what a smiley i ~ p l i volatility e~ is (and is not):
clarifiedthese points aboutthe ~eaning of a smile quote, in the following sections 1 presenttwostylised tradi stories to intro~ucefrom a ncially relevant angle the possible mechani s responsible for the smiles. ese tales of two s iles are not meant to be taken too literally but, if pla~sible,would indicate that ~ p a on ~ toption pric different economic s~ile-gener mechanisms can be which ~ o d ~ l l i n approach g mi
n option trader considers establishing a long and short position in plain-vanilla options on forward rates spannin~ different portions of the same steep yield curve. The forward rates in the low-maturity spectrum trade at around 3% and the long end of the curve has forward rates in the 6% area. The trader wants to deltae dynamically his option ~ositions through time, and trade the g a ~ m a i.e. , pes that the realised volatility will be greater than the implied volatility;2 more precisely, since he is going to be long gamma, every time he re-hedges he will be selli~g when prices rise and buying when prices fall. He hopes that the positive cumulative effect of these re-hedging transactions will more than compensate him for the p r e ~ u m he has paid up-front. The two options that the trader is considering buying trade at the same implied volatility. He observes that, when the yield curve moves, the forward rates in the h po~ion of the curve move by more basis points than the forward rates in
This is the same situatio~ that was discussed in Chapter 2, Section 2.3.
rici
the low-maturity portion. Given the posited shape of the yield curve, this is to be expected since, in a lack world, the absolute one-standard-deviation move should be
~~(t)fT ( t, T ,
where c ~ ( tis ) the percentage volatility at time t of the forward rate, f ( t , T , T z ) , expiring at time T and spanning the period IT, T c ] . Since the ~ e ~ c e volatility for thetwo options is the same, the long-maturity forward rate (at 6%) should move, on average, roughly twice as much as the front-en~-mat~rity forward rate (at 3%). The trader, however, observes that the basis point moves of the two forward rates do not ~ u i t e scale in the ratio 2 1 predicted by their i ~ p l i e ~ ~ e r c e n t a volatilities. The 3% forward rate moves on average more than half as much as the 6% forward. After running a regression, he concludes that the 'low' forward by as much. rate moves, a p ~ r o x i l ~ a t e l ~ 75% , Th~refore the trader forms the belief that, ceteris ~ ~ r as ithe~ level ~ of ~ a~ pa~icularforward rate increases, its percentage volatility tends to be reduce^; si~ilarly, when forward rates fall, their percenta~e volatility seems to increase. to ~which In other words, perhaps the trader subscribes to a world view ~ c c o r d i n
?c l. If the trader's view were that the curve moves strictly in para~lel, with / purely ote that, in the expression above, the term a has r e ~ a i n e d a f~nction of time. This p a l ~ i c ~ l a r a s s u l ~ pabout t i o ~ the link betweenthe percenta and the level of the ~nder~ying rates is r e f e ~ e d to as the Gonst can ore generally, these distribution^^ assu~ptions
df/f = EA, dt
+ f ~ ~ (dzf ,
2)
lack implied volatilities. Also, he trader, h o ~ ~ v elikes r , to t h i in ~ term ~ system does not inco~ol.ate the model, let alone the even more so~histicated stochastic volatility approach~s, and can only price The trader knows that a process like
e to a distri~lltiondifferent from the log-normal, and prices different options. heref fore Ize ~ ~ ~ his ~ s ack prices for ~ot-at-the-~oney prices tu ~ e ~ etc ~te s ~ e e ~ i e and fs q ~ ~ t e s i~ ~p i l ~ i volutilities e e ~ e ~ t for u ~ ~ i o ~ wit^ the s u ~ e e ~ b ~~ ~ t i ~ i e stri~es. ~ r ~ ~Ift one plots the implied volatility against the strike, then one would probably obtain a curve of the shape shown in Figure 4.3.
0.31 I
0.29 0.27 0.25

0.23
0.21 0.19 0.17

0.15
Absolute and squareroot smiles. Thesmilesquoted believed in a CEV model for interest rates with = f or = 0
by a trader who
The most impo~dntfeature of Smile Tale 1 is that the trader in the story believes in a deviation from the log-normal behaviour for the forward rates which is a function of the absolute level of the rates themselves. F u r t h e ~ o r e if , rates tomorrow were to move say, up, then the trader in our story would still use the same graph to read the quote for the implied average volatility for his options. In other words, his beliefs about the relative movement (in basis points per day) rate will of a 3% forward rate with respect to the movement of a 6% f o r w ~ d not be different according to whether rates today have moved up or down. implied volatility versus strike curve would therefore not move with the level of rates. In what f o l l ~ w this s ~ r o ~isereferred r ~ to asa stickys ~ i Z eIt . is essential to note that it is a property of implied, not instantaneous, ~olatilities.
A fund manager has been heavily invested in equities and has ~ a n a g e d via , a
cunning trading scheme of her own device, to outperform her peer group by 2%. he is, however, afraid of the possibility of a market crash, and, as insurance, decides to buy out-of-the-~oney puts. Since many of her colleagues are in a similar ~osition, they are all buying insurance protection, and thereby bid up the prices of out-of-the-money puts. rbitrageurs operating as traders for the invest~ent bank Short Term Capital scour the earth for arbitrage opportunities; in particular for mis-priced options. They cannot fail to notice that, owing to the systematic bidding up of the fund manager, out-of-the-money puts trade 4 kappas above the at-the-money options. If the universe indeed behaved as postulated by the Black model, then there can
rici
@S
be only one volatility for the underlying process, irrespective of the option we might choose to transact. Therefore there seems to be a killing to be made by selling the out-of-the-money puts and buying as a vega (kappa) hedge the cheap at-the-money options. enior manage~ent of hort Term ~ a ~ i t a however, l, is alerted by ment of this trade idea, and begin to ask some probing questions, would happenif the market crash the fund managers are seeking ins did ha en? earch is brought into the discussion, and she points out that, ent, even ignoring the drying up of liquid it^, one of the funda~ental y, i.e. continuous re-hedging, would fail. strategy of the traders, if a crash occurred, then id always lose out. Since over time re-hedging is i~possible, t calls would always lose ptions. A risk-averse is therefore confront ot only with a finite ring a gap), but also, mo
The trading opportunity becomes more and more en
The crucial point is that, despite the conventional wisdom that option prices do not depend on the investors utility f~nction,the de trade, at least in this s t o ~ has , re~e~te the d a ~ ~ e t i for t e risk of the ince the ris~-less po~folio cannot be ut in place in the presence of jumps of unkn itud it is no longer tru that option pricing is independent of risk es . market price for jump risk enters the valuation of the option. The mathe~atics IS rather complex, and will be touched upon later, but the intuition is simple: just as in the case of diffus~onprocesses in i~complete markets (i.e. when eff~ctive hedging is not possible) risk aversion makes the buyer of the option see a less positive expected return on the underlying, in the same way risk adjustment in the presence of unhedgeable gaps makes the jumps more frequent and more to the left. hile all of this has been going on the equity index has been ~ o v i n g all over the place: from ~ 0 0 0 it went to 4500 and back up to ~ ~ 0In 0 this , story, fund managers always face the problem of protecting their portfolio from the ~ ~ s i t where they are; at any point in time they buy puts that are out-of-the-money with respect to the then c ~ r r e ~ at-the-money t level. They do not get sentimentally attached to puts struck at 5000 (quite useful when the index used to be around
ity
~000 but , in the money when the index is at 4500, and producing rather cold comfort when it is at 7000). The bidding up always takes place towards a certain is the degree of out-of-the-monieness, which, in turn, migrates with spot: t ~ i s esse~~ o f ea ~ o ~ s ~ itZ e i . Unlike ~ ~ the caplet trader of Smile Story l, when the index is at 7000 the fund manager will not use the same imp lie^-volatilityversus-absolute-strike curve that she used to buy the puts when the index was 000. Once again, the floating nature of the smile is a feature of the implied volatilities, not of the instantaneous ones. Another way of characterising this assumed behaviour of the fund managers is to postulate that they always seek protection by buyin s i ~ i l ~ out-of-thely mo~ey puts. For this reason, this picture of the world is sometimes referred to as the sticky delta model. In order to avoid confusion, we shall not use this terminology in the future. The purpose of these two smile stories, clearly is not to dispense with careful econometric and statistical analysis of empirical data r e ~ ~ d i n the g dynamics of trading in different markets, but rather to show that plausible but radically ifferent explanations can be found for the same heno omen on, i.e. for the fact that the implied volatilities of options of the same maturity but different strikes are not the same. Reality, needless to say, can be considerably more subtle and com~licated: regimes of sticky and floating smiles, for instance, are not as neatly separated as the two stylised stories above seem to indicate. The crucial point, however, i s to disting~ish as clearly as possible the following two questio~s:
What ~nancial mechanisms can ive rise to smiles? What mathematical models can reproduce smiles? Several ~nancial mecha~isms can be responsible for smiles (some of which will be is cussed at length in what follows):
l. Fully stochastic volatilitymodels. 2. ~omplete-markets jump- iff fusion models. -amplitude jump-di~usion models. stic volatility functionally d e ~ e n ~ eon n t the underlying, CH volatility models.
~ e e d l e s to s say, these ~echanisms are not mutually exclusive. these models it is essential to have clearly in mind at least t (if somewhat stylised) e ~ p i r i c a l facts concerning smiles ab ~idespread consensus. The cases of equities, interest rates an separately; the more detailed properties of the various smile surfaces will also be analysed in greater detail when dealing with the various approaches used to model them.
This sectionreportssome essential empirical facts aboutsmilesindifferent markets that should always be kept in mind when assessing the results produced by the different models,As mentioned in the~ntroduction, it is not the purposeof this book to provide a statistical or econometric study.Therefore these e~pirical, facts are presented without the q~alifications and caveats that would be required if a careful discussion. of the estimation procedure were presented. In a sense, they represent the statistical results about which, in my 5 ~ i ~and i ue~ ~~~~ie an c e broad consensus could be obtained, if all the is were dotted and the ts crossed (with all my excuses to serious economet~ciansj.
. The magnitude of the smile as a function of a fixed money-strike tends to decrease for increasing option expiries: short ~ a t u ~ t i display es pronounced smiles, and distant ~ a t u ~ t i give e s rise to shallow smiles. hat constitutes a maturity, can, of course, change over time; see Figure 4.4.
. The magnitude of the smile as a functio~ of the degree of out-of-t~emo~ieness(as expressed by the delta) ismuchmoreconstantacrossdifferent option expiries.
n fixed expiry, the itude of the smile observed on. es depends on the d of out-of-the-~onieness,not on the absolute level of the strike; see igure 4.5 and the discussion. thereafter.
In the case of equity indices, there is a pronounced smile going from th ney level towards out-of-the-~oneyputs, but h ~ d any ~ y smile goin from the at-the-money level towards out-of-the-money calls; see Figure 4.4.
Figure 4.4 is quiteself-explanatory,and displays boththe ~ a t t e ~ i n of g the smile for increasing maturities, and its virtual disappe~ance for strikes towards the direction of out-of-the money calls. Figure 4.5 is in a way more interesting, and displays the nor~alised smile for two different levels of the spot index. The no~~alisation, achieve^ by dividi each quoted volati~ity by the asymptotic flat volatility co~espondingtodeep-the-moneyputs,attemptsto disentangle the effect due to the change in smile from the overall change in volatility level. If the s ~ l e s ~ ~ were f a c~e~ i cthe ~ ytwo , ~ o ~ a l i s curves ed shou~d approxi~ately be on top of each other. a floating smile, the surface would ~ i g r a t e with t at-the-moneylevel. The t index levels on thetwotrading dates consi~er dif~ered by approxi~ately500 points, which is also the x-axis shift i n the two
c-)
rici
.5 The normalised smiles on two trading days (11 July 1998 and 13 August 1998) for theFTSE index and expiry of one year. The spot levels of the index were S944 and 5460, respectively. The normalisation was obtainedby dividing each volatility by the deep-in-the-~oney-~ali asymptoti~ volatility in order to disentangle effects arising from overall changes in levels of the volatility curve from the smile itself. Note that the 500 two normalised smile curves are shifted by a p ~ r o x i ~ a t e l y points, which roughly corresponds to the difference in the spot level of the index. In other words, the smile floats and is not sticky
nor~alised curves. In other words, Figure 4.5 clearly displays the floating nature of the smile.
S
. The shape of the smile is monotonically decreasing from low to high strikes; see Figure 4.6.
The ma~nitudeof the smile does not systematically decrease with increasing option ~ a t u ~ tvery y ; often it does not decrease at all; see Figure 4.6. practice, in the run-up to the Euro conversion quote volatilities of caplets of relatively less liquid currencies expected to enter the EMU (such as ITL, or ES rking out the absolute basis point M or FRF); therefore, if a trader equivalent with the most liquid curren observed a REM forward rate for a given ~ a t u r i t y to trade at, say, 4%with a volatility of 12%, and was asked for a quote in the volatility of the samem a t u ~ t y f o r ~ arate r d in ITL trading at, say, 5.5%, then she would make a first guess of its volatility as a solution of the equation ~ D 2 ~IT-cTITL ~ (i.e. ~ CTITL2 : ~ D ~ ~ ~ D ~ ~ / ~ this ~ T practice L C T would ~ T ~ indicate ) ; that the trader, as in
. It was common market
irty
tory l , subscribes to a normal (rather than log-normal) view of the distributional prope~ies of forward rates. liest and most pronounced smiles in interest rates have been where interest rates have been lower than in all other major maturity caplet implied percentage volatilities are much higher than in all other markets, again indicating that the market seems to favour a / ? coef~cient in Equation. (4.1) less than l.
. The shape of the smile for mature market cu~ency/e~erging market currency frequently changes skewness over time (from smile to left to smile to right, to a~proximatelyflat, smirk, etc.). The shape of the smile is often synthetically desc~bed b means of two related quotes: the 25-delta risk reversal3 and the 25-delta s which are quoted as the differenc between the im~lied volatilities of options un~erlying the s t ~ c ~ r e so, . if the risk reversal is positive, then the out-of-the-~oneycall is more exp sive than the out-off , say, 2% would indicate that the i ~ p l i volatility e~ the-~oney put. A quote o of the call is 2% above the implied volatility of the put. The strangle is i~stead quoted as the average of the two out-of-the-money implied volatilities minus the at-the-money volatility. A positive quote for the strangle indicates that the outthe-~oney options are relatively more expensive than at-the-money options. ughly speaking, the risk reversal and strangle quotes therefore give in~ication.s about the curvature and the slope (second and first derivative) of the smile curve, respectiv~l~. ~nalyses of the quotes for 25-delta risk reversals over extended period^ show wide va~ations both in magnitude and in sign; 25-delta stran on the other hand, tend trade around variable but mainly positive values. 99) for some empirical data regardin ee ~ o o p e r and Talbot exchange rate.
. The shape of
the smile for mature market c u ~ e ~ c y / e m e an e~uity-likesmile in the direction of emerg
facts about s ~ l e s , aving briefly reviewed some of the most salient empi~cal we can begin. to look at the general features of some of the models that have been
A risk reversal i s created by purchasing of an out-of-s he-money call and then selling a similar call and a long out-of-the-~oney put. A strangle consists of a long position in an out-of-the-~oney position in a similar out-of-the-mon~y put.
proposed to account for them (or, sometimes, in total disregard of the above). As in the case of the empirical facts, the description has been kept at this stage rather s~~mmary, and is meant to give little more than a birds eye view of the topic. The assumptions underlying the various models and their implications are treated in detail in the following chapters.
With this description the dynamics of the underlying is given by
The term fully stochastic is introduced to differentiate these models from the stochastic volatility models (treated in Section 4.6.4 below) where the stochasticity only arises because of the functional dependence of the volatility on the underlying. In the more general fully stochastic case a second Brownian motion affects the evolution of the stock price. ~ i t h o uincluding t options in the set of fundamental trading i n s t ~ m e ~ tmars, kets are, in general, not complete when the volatility is fully stochastic. It is however suf~cient to introduce another option in the set of un~erlyin~ securities to remove the incompleteness. If, therefore, one wants to hedge using only the un~erlying stock, either one mustintroduce the market price of risk, or one is left with the task of implying the risk-aversion parameters from traded options. There is, however, one i ~ p o ~ aexception: nt if the incre~ents of the Browl~ianmotion describing the evolution of the volatility are unco~elated with the increments of the stock price diffusion, then one can show that risk-neutral valuation is still possible; in this case the volatility risk should therefore not be pricedinthe model. hile intellectually interesting, the main drawback of fully stochastic volatility models is that, for reasonable choices of coefficients, they tend to produce large smiles for long maturities, and shallow smiles for short maturities. In reality, as the empirical facts reported above remind us, one observes the opposite. Many studies (see, for example, Das and ~ u n d a ~(1997), ~ m Jorion (19971, and Bates (1996)), and my own experience, c o n ~ r m this observation. It is for this reason that fully stochastic volatility models are not treated in any great detail in this robably a combined jump-diffusion plus fully stochastic volatility model could, in theory, provide the best description of market observed option prices; (1997)) such a model (Das and ~undaram
would ( ... )be able to generate adequate kurtosis at short mdturities (via the jump component) and at moderate maturities (through the stochastic volatility co~ponent). Unfortunately, such a choice would not be a p ~ s i ~ o n i o u one. s
Finally, note that Equations (4.2) and (4.3) above represent a fairly general approach to fully stochastic volatility modelling. Even within a wel1-de~ned subclass of models, however, suchas the class described by ~quations (4.2) and (4.3), there is no such thing as the most general model. A very interesting model was presented by Naik (1993), for instance, where the volatility was assumed to undergo a jump process. For reasons of space this approach is not treated in this book, but the reader is encouraged to study the approach (perhaps after fa~iliarisin herself ~ with jump processes in Chapter 8).
In this model, in addition to Brownian diffusion, the stock price is also affected by j ~ ~ that p soccur at random times, but are assumed to be of known amplitude(s). The analytical treatment is given in Chapter 8, but, in order to gain an intuitive understanding, one can consider the po~fo~io-re~licating strategy shown in Figure 4.7. As is well known (see Rebonato (1998a) for a detailed discussion), if only the 'up' and 'down' states were reachable, then the pay-off of the contingent claim, C, in either state can be exactly replicated by a linear combination withweights a and /? of the stock and a second security (the 'bond'). These weights are the solution to the 2 x 2 linear system
Provided the associated deter~inantis different from zero, the system always has a unique solution. Let u 3ssurne now that, in addition to the ' n o r ~ a l ''up'
.7 The states reachable by theunderlying, S, the contingent ciaim, another security,P (the 'bond'), from the same parent state
C, and
and down9 moves, a jump of known amplitude is also possible~and that both the securities S and P are affected by the jump. ne could be tempted to write
and to follow the same reasoning as above. ow, however, there are three possible are no I states for the system and therefore, the two secu~ties provide a unique solution to the 3 x 3 syste which must be affected by the jump as well t, the system is now 3 x 3 and formally one still obt e same type of solution. Note carefully, to zero, S(up), P(up), S(down), down^, ever, that, as the time step, come closer and closer to S(O),P(O),C ( 0 ) up), C(down), Q(up) and Q( and P2(0), but Stjump), P(ju p) and ~ t j u m p ) do not, In other words, Poisson jumps do not scale The analytic and ~nancial properties of jump-di~usion models are analysed in detail in Chapter 8. In the meanti~e, one can already state if the amplitude of the jump is assumed to be of known a m p l i ~ d ethen 9 these approaches can reproduce many of the qualitative features of the prices of smiley the vanilla options observed in the FX and equity markets, In particul~, ve rise to large smiles for short matu~ties and shallower smiles for lon maturi~ies.Just as important~y,if the jump amplitude is assumed to be known a priori, then one can obtain these results while still employing a risk-neutral valuation in order to arrive at a unique option price. It should be kept in mind, however, that if one thought that Smile Tale 2 was the correct origin for the smile, then one would not be accounting for the mechanism that was assumed to be responsible for the smile itself, i.e. the risk aversion to jumps.
S
With thismodel the possible a~plitudes for the jumps span a continuu~ of values. ne can no longer set up the discrete replicatin portfolio with a large but finite number of s e c ~ r i t i ~~ s .b t a i a ~ unique n ~ price using ris~-neutral valuation is no longer possible. One should be careful with too mechanistic a transition from complete to incomplete markets, and with the a ~ ~ o m ~ a n y intro~uction ~ng of the market price of risk. F o ~ a l ~ if y ,the u~derlyingS can jump to, say, 10000000 000
In discrete time one has to be s o ~ e ~ hcareful at in distinguishing between moves originating from a ran do^ walk, and jumps. In short, one can say that the two types of moves can be distinguishe~ by observing their scaling properties when the time step is reduced.
possible states and one has 10000000002 securities (the stock, the bond and 10000000000 options) at ones disposal to hedge with,then one should be indifferent to jump risk, If, on the other hand, the same investor had an infinite number of possible jump amplitudes, then she should display her risk aversion in a price. In reality, if a trader could include in her set of hedgin~ securima~ng ties 10000 000000 options, then the prices of the latter would not be explai~ed by the model,butwould become an exogenous given. The model therefore ~ g h well t account for all these prices, but, since we can observe them in the precious little extra information. The only contrib from the pricing of the 10 000 000 001st option. is the case, would not any reasonable interpolating or extrapolatin furnish pretty much the same answer? What would be the ex the model? The reader should compare this situation with the case, wherebythe trader can obtain slightly wrong prices for all prices and matu~ties with a single number, the implied volatility.
As briefly explained in Chapter 2, these models describe the stochastic evolution of the underlying state variable by means of a volatility term that is a ~ ~ t e r ~ istic function of the s t o c ~ ~ s tunderlyin~ ic stock price: at time to the user does not know what volatility will prevail at a future time t , since the future realisation of the ~nderlyingis not known. Contingent on the knowledge of the value of the process at an arbitrary future time, however, the volatility is uniquely determined. Owing to the very special type of stochasticity that is permitted with these models,inwhat follows they are referred to as rest~cted-stochastic-volatility models. ~ i t h i n the framework provided by this specification, the dynamics of the underlying is therefore given by
dS = rSdt
+ OS(S,t)dz(t)
The first observation is that the market is still complete, since the (stochasti~) volatility functionally depends on the underlying S. One can set up the replicating po~folio with the underlying security and the bond without any problems~ and risk-neutral valuation will give rise to a unique option price. Since, however, the market is complete and one can use risk-neutrality, if the origin of the smile did lie in risk aversion to jumps, then the trader would be deluding himself that he can hedge where in reality he cannot. Since this point is very important, we a first presentation of the argument below, and return to it at greater leng It will be shown in Chapter 7 how, from a set of option prices, one can impute the underlying distributions of the underlying at diEerent matu~ties, Let
us assume that these distributions of stock prices o~tainable from options prices Poisson process; had been generated by, say, a general (i.e. a random jump-si~e) thenweknowthat a risk-neutral valuationwould no l o n ~ e r be possible. If, however, one simply observed the terminal dist~bution, one ~ g h still t be able to fit it via a suitably complex diffusive behaviour, and try to set up the repliportfolio implied by this process. Such a portfolio would, by necessity, be
This statement is quite obvious if the underlying jump process were such that the jump amplitude were random, since clearly we would be trying to set up a perfectly replicating hedge in a case when one cannot be created, due to market incompleteness. The statement above regarding the inco~ectness of the hedge is true, however, even if the true distribution had been generated by a (series of)Poisson jump(s) of known amplitudes, which would still allow, as discussed above, risk-neutral valuation. precisely, let us call equivalent diffusion the giving rise continuum of distributions for all possible ti by the true known-amplitude mixed jump-diffusion process.Under suitable conditions, a the ore^, to which we shall return again in Chapter 6, states that, if the drift is specified, then there is one and only one diffusion process generating a contin~~um of distributions. One can rest assured, however, to have found the correct gen~rating process only if the distrib~tions had indeed been ge~erated by a diffusion. In many cases it is unfo~unately possible to create a diffusion that can exactly reproduce a continuum of distributions created by a large class of esses, of which jum~-diffusionis one instance. iven that both processes permit 1-isk-neutralvaluation, one might still hope that this equivalent diffusioll would give rise to the correct hedging para~eters. Unfo~unately this is not the case. This can be seen most simply as follows: let us discretise a spot or forward process, assume that the ~eneratingprocess is ~rownian(i.e. assume, without loss of generality, that the underlyin move to two possible states over each discrete, but sufficiently small, and c o n s t ~ c t a replicating portfolio as indicated above. Note that the process under which this replication is possible is more general than is allowed by the lack formula, since the volatility can be state-dependent. Let us then consider the case where in reality the evolution of the stock price was generated by a mixed ~rownian-Poissonprocess with jumps of known amplitude; we have seen that, in this case, the risk-neutral valuation would still be possible, as long as we added a suitable number of inst~ments to our replicating portfolio (as many additional instruments as possible jump amplitudes). In the limit as the time step (which will, in general, display goes to zero, let us estimate the equivalent di~usion a stock-price-dependent volatility). In order to set up the replicating portfolio suggested by this eq~ivalent diffusion process we know that we would need two assets only. We also h o w , however, from the discussion above, that the correct replicating portfolio should be composed of as many assets as states reached over
the next time step. The amounts of these assets to be held to achieve replication is given by the solution of a linear system, and thissolution, if it exists, i s unique, The hedge sug~ested by the e~uivalent diffusion can be seen as a different allocation of hedging amounts; in particu~ar? the amounts suggested by the e~uivalent-diffusion hedge for the extra instrument(s) needed to solve the linear system in the jump case is zero. Since the correct solution is unique, the equivalent-diffusion hed must, by necessity, be wrong. Summ~ising: the fact that we can invoke ~sk-neutrality for valuation purposes does not imply that any modelling approach that produces the same final price distributions and is consistent with risk-neutrality will give e~uivalentresults, Once gain, it is essential to convince oneself thatt h e ~ n a n c ~ mechanism ~2 responsible for the smile can be accounted for by the model. As a final remark, the modelling approaches briefly described above (and more fully in the following chapters) to account for the existence of smiles are by no means the only ones that have been introduced in the literature. Some of the omitted models (such as the Arnin (1993)) are very interesting and would certainly deserve careful attention. My intention, however, has not been to provide an encyclopaedia of smile models, but to furnish the reader with the analytical tools to assess on her own the suitability of a modelling approach for the financial lem at hand. The following chapters should therefore be seen in this light. efore moving on to the analysis of specificmodels, however, it is useful to e x a ~ i some ~ e e ~ p r ~ s s i o for n s the delta and kappa statistics for plain-vanilla options in the presence of smiles often found in the literature. The a~companying analysis will provide some further clari~cation of the conceptual and practical distinction between sticky and floating smiles~ and will constitute a useful backdrop for the more general discussion to be found in Sections 6.7 and 6. Chapter 6.
It was pointed out in ection 4.2 that special care niust be taken when employing lack implied volatility function; in particular, we emphasised that it can be safely usedonly for the purpose of obtaining a price-asits de~nition implies-and that other uses, such as the calculation of a delta or a kappa via lack formula, are unwa~anted. If, however, we forget for a moment how the Black f o ~ u l was a originally obtained, and what the links are between this formula and the volatility of the underlying (at least, if the volatility is time-dependent)? we could regard the market practice of quoting a lai in-vanilla option price in the presence of smiles as made up of three co~ponents:
1. an a priori agreement of an arbitra~ formula, which is a function of several inputs; reement on the procedure needed in order to obtain from the market or the product specification all the inputs but one (i.e. todays stock price, the funding rate to maturity, the strike, the time to maturity); and . the price quote, expressed in terms of the remaining unobservable input.
If we look at price-ma~ngin this light, as far as this last input is concerned, despite its being called volatility,, no specific interpretation in terms of more fun~a~ental ~uantities is in general wa~anted, This inte~retationof the pricemaking process is obviously quite strained: even in the presence of srniles, the lack formula, for all its imperfections, is clearly not any arbitrary formula; and the implied volatility, despite the fact that it is not truly a time-dependent volatility, is in some imprecise way similar toits well-defined cousin that applies i s a pure diffusion driven by a time-dependent when the process for the underlyin~ volatility. The advantage to be reaped by looking at the p l ~ c e - ~ ~ process n g in this matter will become fully clear in Chapter 6, Section 6.2. In the meantime we can note that it allows us to write in a purely formal way:
t is the price of a plain-vani~la call or put; S, T and r are the stock strike, the time to expiry and the short rate, price-generating number chosen by the trader to make a -upon function to obtain a quoted price from this last input. As for the ~uantity crimp{, it is im~ortant to note that it can be a fu~ction of S, t i.e. q m p l = OimpI(S, t , K ) :in other words, the market agrees to keep the lack-box formula constant over time, but allows for the possibility that the priceenerating input crimp] might change both as a function of the price level, and as a function of time. Today, traders can directly observe only c r i ~ p l ( S to, ~, K). I fthe matter is looked at in this way, one is tempted to write for, say, the delta of the option an expression like
where Asjack and Black are just the usual lack delta and kappa, respectively. rom this o~servation it might seem that evaluation of the presented in ~ection 4.2 (see also Figure 41), was not wro plete, and that the addition of a simple term would have fixe seem, in other words, that, even in the ~resenceof s ~ i l e s the , delta statistic calls and puts can be calculated without the need for a sophisticated model. alters, u~fortunate~y, are not that simple.
rici
Let us focus attention on the two terms ~ ~ ~ t / ~ and a i ~ ~a p i ~ l p l The / ~ calcu~ * lation of the first, as we have seen, presents no problems, and it is the well-~nown kappa (vega) statistic; but, when we come to the second, the price-making nism described above gives the trader no indication as to how the implied volatility shouldbe changed when the underlying changes. ~bservation of today's smile surface tells the trader directly how the impliedvolatility changes as a function of matu~ty and s t r ~ not ~ ~as , a function of time and stock price. same observed prices, for instance, could be explained by either of the im volatilities below:
%npl
= I
{2 1
Clearly, in the first case the term in the derivative, ~ai~pl/~S, would simply h (independent of the option strike), whilst in the second it would at varies for options with different strikes). The distinction be ating stniles has therefore resurfaced again: in order to calculate the term ~ a i ~ the ~~ trader / ~ must S decide how the implied volatility will chan level of the market will have changed. The p~ce-generating black box (or ck box?) does notprovid y information about this choice, y prices. observation of yesterday,^ and t ervation of t u d ~'S combined with a subjective assumption about time homogeneity (or absence thereof) can provide the information necessary to com~ute the term ~ ~ i ~ ~ but this choice is clearly to some extent dependent on the j u d g e ~ e nof t the trader. Therefore, even if the market p ctly s u ~ s c ~ i b e to d the procedure outlined in still have to make a trading choice when it steps l to 2 above, the user w for her plain-vanilla option. i m i l a r term int out tlnat the deceptively s ut how the implied volatility cha perfectly evaluated from the of ~ u d ~ y market 's prices: the latter, in fact, directly provide infor erating input ~uantity aiql changes when the stri is left as the proverbial exercise for the reader to case for the cash-or-nothing ~uropean ~ i n option ~ y discussed in be expressed as a function of the term ~ ~ i and can ~ ~ ~ / ~ ~ independently of any assumption about the sticky or float surface. e ~undamentally, in 6 and 7 we shall discuss how, under sLlita~le ons, the probabilit~ of the stock price can be obtain~d fro of second ~ e ~ v a t i v ethe option price with respect to the strike. The fact t term ~ a i ~ p is l independent / ~ ~ of the sticky or floating nature of the smile will to be of funda~ental i ~ p o ~ a n in c ethis context. ome of the topics discussed in this section are touched upon in a very interworking paper by einer (1999, who embeds the discussion in the broader context of jump processes,
In the final section of the previous chapter it was mentioned that stochastic~olatility and jump-diff~sion models are amongst the most commonly invoked to model and account for smiles. In this chapter we shall analyse i the theoretical features of the former modelling approach (Sectio then focus attention on an importa~t sub-class of stochastic-~olatilitymodels, i.e. on the so-called restricted-volatility models, and analyse in detail one of their most popular nu~erical i~pleme~tations (the Derman and
means the most general) ap eneral ( a l t ~ o by ~ ~no ~h the dynamics of prices or rates is to assume a mixed j u ~ p - d i f f ~ ~ s iprocess on for the ~ n ~ e r ~ yof i nthe g form
In E .l), S ( t ) indicates the value of the price or r drift sted, if approp~ate) and a ( $ , ,t ) its volatility. the notation we have assumed that the stock price is shocked b nian motion, dz(t), but the extensio~ to several shocks would be straightforward. Note also that we have allowed for a depende~ce on the u ~ d e r l y i n in ~ the term cr. (To lighten ~otation, in what follows the dependence on time of the underlying when it appears as an argument of the function cr will be dropped.) Also present in E~uation (5.1) i s a jump component, described by means of a random
The calculations ~ e r f o r ~ e by d Sunil Bowry and D r James Pfeffer are gratefully acknowl~dge~.
jump amplitude, As, and the increment of a counting process, jump co~ponent is concerned, if the jump a~plitude is rand the previous chapter, and we shall prove in greater detail in Chapter 8, .3, that arriving at a u n i ~ u e (preference-fre option price u s i a ~ risk~ luation is in general no longer possible. wing to the considerable technical and conceptual complexities connected with jump processes a common roach has therefore often been to neglect the jump CO ) altogether. Since in this chapter we shall focus on volatility models, the jump component, which, despite its complexity7can indeed be very i ~ p o ~ ain n order t to account for smiles in a financially convi~cing way, is t e ~ p o r ~dropped ly from the a~alysis, but will be explored in detail in Chapter 8, S far as the volatility term, a($,, t ) , is concerned there are at least four main ternatives: in order of increasing complexity it can be desc~bed by ic function of time, by a deter~inistic function of time and of the ock price, by a diffusion process, or by a mixed jump- iff fusion latter possibility is not explored any further (see, however, (1993) for an interest in^ perspective), If one does not allow for the possibility of jumps in the volatility7the most general desc~ption of the time evolution of the volatility can be described by a process of the form do2($, CT,t ) = p , ( $ , a2, t ) dt quation (5.2) p , c~efficient U-, and the way that
+ v,($,
a2, t ) dw(t)
d v, denote the drift and the volatility of the d i ~ u s i o ~ ownian processes z ( t ) and w(t) are correlated in such a
E [dz(t)dw(t)] == pdt
(5.3
order to keep the notation as simple as possible, we have a s s u ~ e d rownian shock, dw(t), affects the volatility; the extension to the multi-factor case does not present conceptual difficulties(butobviously does ve an impact on the number of instruments needed to complete the market). that there can, in general, be two distinct sources of stochastic behaviour for the volatility a: the first sterns from the functional depen~ence of a on the underlying S, which, in itself, is a stoc~astic ~uantity; the se fact that the volatility is allowed to be shocked by a second W t on1 i ~ p e r f ~ ccorrelated, t~y if at all, with dz(t), as show splaying stochastic it^ in the volat origin at in^ both from the funcnianprocesswill be referred to ndence on S and from a separate inwhat follows as fully stochastic models. ~ o d e l for s hich the volatility is stochastic only because of its dependence on the underlyi will be described as ~est~cted-stochasticvolatility models. A restricted-sto stic-volatility model case ~ of the more general d e s ~ r i ~ t i o (n can be seen as a l i ~ t i n volatility term v goes to zero. The distinction is impo~ant, because, as discussed in the previous chapter, risk-neutral valuation cannot in general be used to obtain
a unique option price for fully stochastic models, whilst it is perfectly appropriate for rest~cted-stochastic-volatilitymodels. It is, in fact, importa~t to recall wnian increments dz that, for a generic correlation coefficient p between the and dw, markets are no c o ~ p l e t e unless one includes e options in the set of hedging, s~cu~ities. n other words, if p is not equal to zero, thenunique option prices can be obtained using risk-neutrality only if the possible prices of o ~known, ~ If one is allowed to hedge the underlyi~g of so^^ other o ~ t i are S, then no portfolio can be made to replicate th only with the underlyin~, exactly, options are not redundant secu~ties,and the ~nancial,mathe~atical, n~merical and practical complications introduced are substantial, A common, but strong, simpli~cation of Equation (5.2) is achieved by imposing that the volatility of the underlying, cr, should only exhibit the stochastic behaviour allowed by the functional dependence on S. This corresponds to averaging out all the sources of unce~ainty associated with the and is sometimes referred to as the effective volatility th ani (1998)). Therefore, if this assumption is made, then
~~~
.4) therefore describes the most general set-up that the case of a purely d e t e r ~ ~ i s t (time-dependent) ic volatility, and still allows ris~-neutralvaluation without introducing other hedging inst~ments (i.e. other options) apart from the underlying itself. d in the previous cha ,it is important to note that, whilst volatility implied by ation (5.4) displays a full functional dependence on the (random) realisati f the underlyin~;in other words, at a given future point in time, the value of the volatility is uniquely det by the value attain~dat that point in time by the stock price (or rate) (5.4 n should therefore not fo et that the model describe^ by ~ ~ u a t i o han a purely deterministic-volatility approach, still contains restrictions about the possible values that can be assumed As men~onedabove, we shall refer to this class of models as restricted-stochastic-volatilitymodels. As usual, the power of a m o d ~ l l i n ~ approach resides in its ability to predict the hedging costs that will be encounout the iife of the option. In partic r, an approach such as the one 4) makes a very stro statement about the ing costs, that will b countered at future time. The importance of this proviso will be ill~strated at length in the present foilo~il~ chapters. g espite the fact that the stochastic process for the underlying S d .4) is not unduly complex, a further simpli~cation is some one, in fact, sometimes assum~s that the function o ( S , t ) is separable in the form a(S,t ) = ~ ~ ( S ) ~ and ( ~that ) ; the function f (S) is of the form f (S) = Sp, with
B >_ 0. If this is the case, then

t) dS(t) = p , ~ ( dt
B = 0, $ or l simple explicit closed-forl~solutions exist. era1 solutions in terms of modified esse1 functions are also available an arbitrary coefficient / 3 (see, for exampl iner (1998)). The class of models descri~edby Equation (5.5) is often c l to as the Gonstantance (GEV) models. E~uation(5. quation (5.4), and the advanta~esthat c from edthis loss rality are purely comp~tationalinnatur moreis than no simpler rest~ctive nt ion (5.5), in what follows we shall discuss one i ~ p o ~ a case i~cation~ i ~ actually h t provide a ~ l a u s i ~ description le of an irn ortant fina~cial problem. For the oment, however, we shall retain the degree rded by Equation (
or
+~
~ d~(t) ~ ( t )
(5.5)
upire (1993,1994), hat follows) (see als extract from todays q
thattheusershouldprovidethepricesofoptions for the ~ i s c r e t e ~ a t u r i t i ~ and strikes that corres~ond to the nodes of their c o ~ ~ u t a t i o n lattices. al however, in order to obtain results of acceptable llul~ericalquality a ve space-time mesh is needed, I shall always assume in what follows that a co~tinuous price (or im~lied-volatility) fullction versus maturity and strike has already been obtained by some numerical means (least-squ~es fit, bi-cubic splines, or other). This latter exercise is often not at all trivial (a p o s s i ~ l e a ~ p r o a is c~ presented in Chapter 7), but it is assumed in what follows that a continuous2 of any maturity function that supplies the option implied volatilitiesas a functio~
is actually desirable that this unction should have at least continuous first and second derivatives with
We shall show in what follows (see, in particular, the discussion in Chapter 6, Section 6.2) that it
and strike is already available, so that from the latter one can obtain the prices of plain-vanilla options for any strike and maturity one might need. It i s important to note that the Rubinstein, upire and DK approaches provide, at the same time, a pricing engine (the calibrated tree), and the local volatility surface (i.e. the function cr(S, t ) ) . These two pricing ingredients, however, are i ~ g need not conceptually completely distinct, and an efficient ~ r i ~ ~ethodology provide the most accurate toolto extract the local volatility (or viceversa). T shall argue in what follows that this is indeed the case, and that a procedure conceptually similar to the one proposed by K could provide an efficient prici ~echanism once the function cr(S, t ) had b obtained, but that the canonic DK method is actual1 n u ~ e ~ c a relatively ll~ poor at obtaining the local volatility surface itself. atte~ptingthe de-coupling of the extraction of the local volatility func m the pricing engine, however, the DE( procedure will be analysed in detail. Since their approach depends more fundamentally on the Greens function formalism than the usual bi- or trinomial tree const~ction, this topic is examined in the following section in the discrete-time, discrete-space framework best suited to the
Let us place ourselves in a universe where trading only takes place at discrete time intervals, and where only a finite number of states at each possible time step are reachable. ~e impose no restrictions on the trading frequency and the number of states, other than they both have to be finite. As for the characte~satio~ of a state at a given time, we shall assume that it is fully and uniquelydefined by the realisation of the price at that point in time. Therefore, speakin ( j , k ) (with the firstindex referring to the time slice and the second to the state) i s exactly equivalent to speaking of the stock price having attained the value S ( k ) at available time t j . (In technical terns, the filtration that describes the i n f o ~ a t i o n to the trader at different points in time is therefore e~cZ~siveZy generated by the price evol~tion.) Given this set of prices and times, let us construct a re comb in in^ trinomial lattice withnodes located at the trading times and at the possible prices; see Figures 5.1 and 5.2, shall also assume that we have already obtained the set of all probabilities connecting each parent node at time j with its three offspri~gs7 at time j 1. An Arrow-Debreu price (also sometimes referred to
respect to the strike, and a first derivative with respect to the option ~aturity. Also, the price function can be fitted directly, or can be obtained by fitting the implied volatility function. If this second route is followed, the continuity requirement will still apply to the derivatives of the price function.
G(up)=exp[-r
* dtj * p(up)
~(mid)=~xp[-r * dt] * ~(rnid)
G(down)=exp[-r
* dt] * p(down)
root
The Arrow-Debreu prices for the three statesreachedat
time 1 from the
G(up)=exp[-r
* dt] * p (
G(rnid)=exp[-r* dt] * p(mid
~onstru~tio of n the price of an Arrow-Debreu security paying 21 i f the central state at time step 2 is reached, and fO otherwise
function) for state j at time k , G(j, k ) , is then d e ~ n e d to be the price curity that pays &l if state j is reached at time k , and 0 o t h e r ~ i s e . ~ al lattice and the set of probabilities we can easil breu prices. To begin with, by definition, G(0, U prlces of secu~ties paying &lin any of the nodes at step are almost as strai~htforwardto determine, as shown in Figure 5.1, where deterministic and constant interest rates have, for simplicity, been assumed: they iven by the discounted probabilities of reaching each of the thr nodes at time step 1 Note, in passing, that the discounting rate r in .lis actually the difference between the risk-less one-period borrowing rate and the (constant) dividend yield in the case of equities, or the difference between the domestic and foreign rate in the case of FX rates, The const~ction is sli more comp~ex for the second time step, and is shown in Figure 5.2. is modified in m o v i n ~ from the first to the second To see bow the const~ction time step, let us now place &l at, say, the central node at time step 2 (i.e. the u p jump, or by node reached after a d~wn jump followed byan jump and a clown jump, or after two mid jumps), and let us calculate mid,2); see Figure 5.2. y putting &lat node (mid-mid,2) one obtains a non-zero cont~butionto the upper t ~ ~ o ~sub-tree, i a l the middle trinomial sub-tree, and the bottom trinomial sub-tree (These sub-trees are simply the three branches o~ginatin from ~ the top, mid and bottom node at time step 1.) The contribution, for instance, to the node of the upper~ost sub-tree (state u p )is given by
This is the value at tree node (up,l) of &lplaced at (mid-mid,2). ~ i ~ i l a r l&l y, placed in the central node at time 2 will contribute to the central and bottom sub-trees as well. ~e can compute the value of &l placed at ( ~ i d - ~ d 9 at 2 tree ) at tree node (down,l) in node (mid,l) and the value of &l placed at (mid-~id,2) a similar way. from placin Let us call the three contributions to the thee sub-trees o~iginating &l at state ( ~ d - r n i d 7 2 X), Y and 25, Therefore
The ~ ~ o w - ~ e b prices r e u G(up,l), G(mid,1) and G(down,l), however, give, by de~nition, the value of a unitpayment in states u p , mid and down at
The const~ctionjust presented is muchmorerestrictivethan i s needed in order to define Arrow-Debreu prices; since in what follows use will be made of these prices purely in the context oftheDK ~ethodology, to whichthisset-up is perfectlytailored,noattempthasbeenmade to embed the derivation in a more general framework.
time l , respectively, Therefore, again by the de~nition,mid- mid,^) is equal to
imilar considerations obviously apply to later time steps. This simple c o n s ~ c tion can therefore be generalised, and one can draw the following conclusions.
. Once the Arrow-~ebreu prices out to time y1 are ~ n o w n ,one can c o n s t ~ c t all the Arrow-~enreu prices at time y1 l with O(y1) operations (in a recombining tree). One does not have to traverse the tree all the way back to the origin again.
there are (n) time steps, and obtai~ing the prices from a give ime slice takes O(n) operations, the construction of the whole tree of Arrow- ebreu prices requires computations, A naive const~ction whereby each payment was disco~nted all the way back to the root would require O(n3)operations. This would soon render the computationa~cost prohibitive. The A ~ o w - ~ e ~ r prices e u can be used to obtain the value of the origin (i.e. today), provided that the values of this security in all states of the world for which Arrow- ebreu prices are a v a i ~ a b are l ~ known. This property is central to the D
. Since
t is essential to keep in mind that assuming the existence of Arrowount to assuming market completeness, absence of arbitra~e, risketc. (see, for Duffie (1996) for a deep and thor treatment, o r (1998a) for a relatively short but question: Are contained isc cuss ion). One still has to ask oneself the f~ndamental these assumptions actually appropriate to describe the ~nancial problem at hand? This issue will be discussed at much greater length later. inally, note also that the tree recombines. This is computationally ex~edient, of nodes. This feature is since it avoids the exponential plosion of the nu not, however, without its own drawbacks, since it ly reduces the f l e ~ i ~ i l i t y in the possible values reached by a parentnode. T turn, is closely linked to the stability of the Explicit Finite ~ifferences methods-to which trinomial trees are closely related-and to the failure to obtain positive ~probabilitiesemanating from some of the nodes. T h e s ~ two closely related aspects will also be discussed reater length in what follows.
order to reproduce exactly the market prices of todays plain-vanilla options, rman and Kani (1998) recommend a computationa~procedure based on a
trinomial tree. The possible values attained by the stock4 price at the different levels and time steps are arbitrarily pre-chosen. Recombination is ensured by the geometry of the tree. In this respect the const~ction is similar to the Explicit Fi~ite Di~erences technique: at each node the user has to determine the three emanating pro~abilities connecting a parent node with its three offsprings. The fundamentaldifferenceisthatfinitedifferencesapproachesattempttofind a solution to a Partial Di~erentialquat ti on (~DE)-typically parabolic for option applications-given the drift and the volatility at each time-space node. In other words, the procedureis akin to a local m o ~ e nmatching, t and inthe discretisation of the various time and space derivatives the local volatility is assumed to be known. A s far as theExplicit Finite Differencesmethod is concerned, which a priori is only co~~ition~ZZy stable, it is therefore possible and easy to check whe~er a stability violation will be encountered at any of the nodes (see, for example, Arnes (l 977) and, Wilmott (1998)). In the case of the DK construction, however, the local volatilities are a by-product of the algorithm, and, therefore, it is not possible to check beforehand whether any of the nodes will give rise to negative probabilities. Note carefully that, in discrete-time financial theory (see, for example, Pliska (1997)), absence of arbitrage is linked tothepositivity of all thepseudoprobabilities. It is important to emphasise that the possibility of en~ounterin~ negative pro~abilities in the DK algorithmic const~ction has no such f~ndamental implications, andsimplyreflectsthefinancial arbitra~nessofthechosengrid points. In other words, a different length of the time step, a different position in^ of the nodes or a different choice of arrival points could either give rise to or eliminate negative probabilities, without any of the actually traded prices being affected. To see howtheDK c o ~ s t ~ c t i o isn carried out in practice, let us start from the firstnode as shownin Figure 5.3. Three probabilities (p(up), p(mid) and p(down)) have to be determined. The usual normalisation condition provides the first equation between them: p(up)
+ p(mid) + p(down) = l
(5.5)
The first moment of S (i.e. the expectation) given its value today is also easily obtainable: a simple no- bitr rage argument shows that it must be equal to the one-period f o r w ~ d value of the stock price given its price today. More generally, when the c o n s t ~ c t i o is~carried out from a node other than the origin, the firstmoment-~atching condition equates the expectation with the one-period forward price as seen from the parent state: S(up)p(up)
+ S ( ~ i d ) ~ ( m i+ d) S(down)~(down) = S ( 0 ) exp[r dt]
(5.6)
The variableevolved onthe treecauldbe a stock price, anequity index, an FX rate, etc. For si~plicity, it will be referred to as the stock inwhat follows.
~ ( d o ~ n
S(down)
The first step of the DK construction, with the stockprices at time1 denoted by S(up), S(mid) and S(down), and the probabilitiesconnect in^ todays stock price, S(O), with the three possible values at time 1 denoted by p(up), ~ ( ~ and ~ ptdown) d )
ote that, as mentioned before, in ~ q u a t i o (56) ~ above the ~uantityr denotes the di~erence between the short (deter~nistic)rate and the ck case, or the difference between domestic and the local volatility were known, then we could carry out the usual moment-matc~ingexercise using any of t e ~ ~ n i ~ (see, u e s for e~ample,Nelson and ~ a r n a s w ~ r n (1 y (1994)). h our case, however, we do not have access to the local va~ance, an S method is different from traditional approaches. In order to see how the problem has been overcome, it is important to remember that it has been assumedthat the prices of plain-va~illa calls and puts of all strikes and matu~ties are known. Therefore we know all the prices of all calls and expiring at time step, say, 1 . in p ~ i c ~ l awe r , know, for instance, the price of the call with strike equal to S(mid) as shown in Figure 5.4, The pay-offs of this option in the different states of the world at expiry (i.e. in states up, down and mid) are given by
in state (down) = 0
in state (up)= [S(up) S(mid)] The ~ ~ o ~ - ~ e price b r eatuthe origin is, trivially, given by G(0,O) fore the model value today of the chosen option is equal to
~ P ~ ( ~= oed xp ) [ - r d ~ ] { ~ ( d o wI ~0 )t p(mid) *0
1. There-
+ P(UP>*N u p ) S(mid)I~
0)
= ~(up)[S(up) S(mid)] exp[-r dt]
The pay-off of a call optionwithstrike equal to 5 ~ ~ drawn i ~next ) tothe tree c~nstructionfor the first step
t ~ ~ ~ ( n l= o~ d~ ) t ( m a r k e tone ) , obtains one equation in one

(5.7)
This formally closes the set of equations for time step l. i ~ p o ~ anumerical nt co~ments should be made about the quantity market). We have a s s u ~ e d i n the treat~ent above that the model value of the option was set equal to the true observed market price of the relative option, This would however introduce an u~wanted~ u ~ e r i c loss a l of in~ormation,whichcan be tood at the first time step since we are, after all, evaluatin only three sample points. As shown later on, this n u ~ e sappear or average out for later time steps, but would ive rise the local volatility function. y the e~uation above equal n ~ a r ~price, e t but to the btained with a traditional trinomial tree. ore precis el^' the quantity arket) et) above is o ned by pricing the relevant option using a traditional trinomial tree built U no smile, the same state space and a ~ o ~ svolatility t ~ ~ tequal to the i volatility of the relevant attempts to option, This procedure (similar in spirit to a contrava~ate techniq~e) purge out of the DK smiley tree construction the nume~~cal errors o r i g i ~ a t i n ~ from the coarsen~ss of the discrete tree, and which can therefore be expected t ~ n o ~tree, ia~ Note, to be e~countered also in the construction of a tra~itional that, unlike the DK tree, the latter is built starting from an assumed e of the volatilities. It is therefore not a priori obvious that numeric~l errors inthe constr~ctionof the two trees will be strongly correlated (which
is the necessary condition for a co~travariate approach to work effectively). As shown in Section 5.6, not only does this turn out to be the case, but the procedure actually proves essential in obtaining n~merically acceptable results. o r later time steps the idea is the same, although there is a slight complication to the fact that, as shown below, the model price will be a function not only of one u n ~ o w n probability, but also of other (already determined) probabilities. ore precisely, the algorithm is as follows:
onsider the first time slice for which probabilities have to be d e t e ~ n e d . Choose as many options as nodes -2. from the topmost sub-tree of the time slice, made up of three branches, se the first option to be a call and place the strike of the first option exactly at the second node from the top. time slice, only one node has a pay-off different from e, acrass the w~ ul e (this is the reason why a call with this ~ a ~ i c u lstrike a r was chosen), this will ~ e t e r ~ i n the e discounted value of the pay-off at the parent node in the time slice as a function of a single prob this discounted pay-off by the Arrowprice of the parent btain the model option price today. this quantity with the market95price of the first out-of-t~e-money ves one equation in one unknown. Together with the forward-price condition and the probability - ~ o ~ a ~ i s a t i o condition this equation uniquely determines the three ~rob~bilities from the rent node in the previous time slice. ove to the second sub-tree from the top, and add one (call) option with a e positioned at the middle node of this sub-tree. option also introduces non-zero pay-offs for the sub-tree above the one onsidered, but only adds one ~ ~ ~ probability. n a ~ n ly the probability-wei~hted nted pay-offs ori by the appropriate Arroweu prices at the p (which have already been determi n sa ~ ou ~ i ly one ~new~ ~rubabili~ is i n t r a ~ ~ c e ~ . Obtain the model option price, equate it with the market option value and solve for the u n ~ o w n probability. Repeat the same for all the nodes along the time slice. ove to the next time slice.
*
As one can appreciate from the description above, the procedure is conceptually
very easy. The only slightly cumbersome aspect is the fact that, as mentioned above, when one moves from the topmost node downwards, the model price r more becomes a function not only of one unknown probability, but also of one o
Again, the term market should be understood to mean the model price obtained using a traditional trinomial tree. as discussed above.
of the already determined probabilities for the same time slice. Simple algebraic ~anipulations give rise to the expressions reported in Derman and Kani (1998) (the symbols have been modified to conform with the notation used above):
+ l), t(n + l))G ( ~ ) (~S (( i~ + )1)) PG)= G(i)[S(i + 2) -S ( i + l)] F(i) ~ ( i ) ( S (+ i 2) S ( i + 1)) S(i + l ) di) =
exp[rAt]C(S(i
j=i+l,2n
S ( i )S(i
+l)
when calls are used, and

PG) =
q(i) =
F(i)
G(i)[S(i 1) S(i)] ~ ( i ) ( S ( i 1) S ( i ) )S ( i 1) S(i)S(i l )
j=X,i-l
when puts are used. In the e~pressions above, S ( i 2), S(i 1) and S(i) correspond to S(up), S(rnid) and S(down) in the previous figures; F(i) denotes the forward price; p ( i ) is the up p r o ~ a ~ i l i t y e ~ a n a t i from n g node i at time n ; q ( i ) is the down probability emanating from node i at time n ;C ( . ) and P(.) indicate the prices of calls or puts, respectively; and G(.), as before, are the
,in ~ ~ n c i p leither e, a call or a put could be used in the

purposes it is important to switch option types across the id-poi~t for each time slice so that at-the-money or out-of-the-money options are always used (calls in the top part of the tree and puts in the lower part): if that were not the case, then one would end up using very deeply-i~-the-money options, which contain virtually no volatility information. in a forward contract struck at K plus a long position in a pu e~uivalent to a long position in a K-strike call, the simultaneous correct recovery of the forward price and of the out-of-t~e-moneyoption ensures, by call-put parity, that the in-the-money option will also be automatically correctly priced. As is a ~ ~ a r eby n t this exposition, the method is conceptually very simple and elegant, easy to i~plement and quite ingenious. In its i ~ p ~ e ~ e n t a t ihowever, on, one can easily encounter rather serious technical problems. The main ones are discussed in the following section.
The previous section described the general idea behind the DE; construction. Its practical implementation, however, can be fraught with dif~culties,many
of which stem from the fact that the price-time grid is chosen inwhat is a substantially arbitrary fashion, without prior owle edge of the local volatility. This state of affairs gives rise to the first two, closely linked, nu~erical problems listed below, whilst the third is partly ~itigated by the use of the contrava~ate parallel tree.
This is the case depicted in Figure 5.5. When the forward price is greater than S(up) or lower than S(down) no linear co~binationof $(up), $(down)and S(mid) with positive weights can be equal to S(0) exp(rAt), This is one of the well- own stability problems en count ere^ in the case of the Explicit Finite ifferences technique (see Ames (1977)): the method become^ ~nstablewhenever the expectation is greater thanthenode above S(up) or the node below S(down) at the following time step. Note that, as ~entioned before, ~nancially this would imply a possibility of arbitrage; this is, however, only true if the arbitrarily chosen values of S(up), S(down) and S ( ~ ~ itruly d) W states reachable from the parent node, In the context of the ne~ative pro~abi~ities simply indicate that, given the one-p stock prices at the fo~lowingtime step do not nstitute a possible discretisation of the underlying continuous-time process. nce, as ~entioned above, the choice of the grid points is arbitrary, no special financial meani~g can be associated with this violation. Finally, note that it is relatively easy to check whether this condition is metatany point in the tree before actually car~ying out the constructio~.
$down)
The case when the forward price is not contained between
S(up) and
The maximum conditional variance that can be calculated from the three values S(up), S(mid) and S(down) is obtained when p(up) = p(down) = i.It is easy to understand why this should be the case: if there is any probability weight to the middle node this, by necessity, would reduce the overall dispersion, and therefore imilar~y? if the probability of reaching either S(up) or S(down) were greater than $, then one would also have a greater concentration of mass at either node, and therefore, again, a smaller dispersion. Thus no greater variance can be obtained than for the degenerate case of zero middle probabi~ity? and up9 and down probabilities equal to But, if this is the case, one can easily check that the conditional sample s t a n d ~ d deviation is simply given by 1S(up) S(down)l. If this standa~d deviation is not large enough to a c c o ~ ~ o d athe t e local volatility necessary to price the corresponding option correctly, then at least one of the probabilities must become negative. The result of this is the same as before, i.e. instability of the associated Explicit Finite Differences scheme. however, that, u n l i ~ e the situation with the usual discretisations of volatilities in the DK construction are not known a priori, i.e. the tree, and it is therefore not easy to check if the Explicit Finite Differences stability criterion i s met. The probability violation on the forward condition can be easily fixed using, for example, n o ~ - s y m ~ e t r or i c abnormal ~ ~ ~ n c h ias ng, first indicated byull and White. ~ i ~ i the n g varian~e violation is ~ u c h more cu~berso~e.
i.
The choice of the strike described above gives rise to a very simple and elegant algorit~m, which is ~otivated by computational rather than financial considerations. In pa~icular,if we want to place strikes on tree nodes, the choice (with its mirror image, whereby one starts from the bottom of the tree) is the only one that allows us to solve explicitly for one probability at a time. Any other ~o~itioning of the same number of strikes would in general require solving a series of 3 x 3 linear systems. It is also possible to retain the simplicity of a one-un~nown-one-e~uatio~ proble~ by placing the strike not exactly at the middle node, but at any (more out-of-the-money) position located for a put between the middle and the down nodes; see Figure 5.6, This simple observation gives further insi~ht into the role played by the r e p l a c e m ~of ~ tthe true market price with the correspond value obtained using the traditional trinomial tree. ~onsider, in fact, the effect of movin~ the strike progressively more out of the money without crossi
The trinomial tree, with different possible values for the strike, still giving rise to a single equation in a single unknown
node. The pay-off would change linearly, and multiplication of this discounted pay-off by the same ~ ~ o w - P ) e b r eprice u at the parent node would give a purely linear variation in the model option price. We know, however, that, in reality, a true (i.e. Black) one-period option is not a linear function of the strike. If one true call or put market price (as opposed fitted a model price to the co~esponding to the control market price), then one would pollute the estimation of the local volatilitysimply because of the numerical failure of the naive algorithm to capture the correct convexity. Therefore, if one used the naive Derman and ni approach with, say, a flat (no smile) volatility surface, and one moved e position of the strikes as described above, and shownin Figure 5.6, then one would obtain a non-flat local volatility. This would be a purely numerical artefact that will tend to become of small nd smaller magnitude as the number of time steps increases. If, however, the model is implemented inthis naive manner (which the authors do not recommend), then the speed of convergence can be empirically observed to be very slow, If, on the other hand, one carries out the same exercise, but equates the model price with the pseudo-m~r~et price obtained using the traditional trino~ial tree, then the effect almost completely disappears (see the tests below), showing the effectiveness of the recommended proce~ure to compensate for this convexity effect. u ~ m a ~ s i the n g discussion so far, in this section we have pres fundame~tal technical atures of the implementation of the The reader is referred to P) S papers (l998 and references therei out nume~calexamples of the construction, which are therefore not repeated here. Instead, in what follows we present some detailed and specific tests of the
implementation of the R approach in different input regimes (e.g. flat or timedependent implied volatilities, simple linear smiles, etc.), with a view to assess in^ the ~ e ~ i b i l i and t y ro~ustness of the model in realistic pricing conditions.
The first test to which the method can be submitted is the pricing of options in the absence of any srnil r time dependence for the implied volatility function, with and without the fitting to the control market prices (as described in th these inputs the derived local volatilities can only be everyequal to the implied volatility (the truth of this statement is proved later on), For the purpose of this test, the DK construction was carried out both by fitting to the prices produced by a traditional refere lattice as e~plained above, and by fitting to the true market prices. recom~ended procedure the agreement was found to be virtually the resulting local volatility surface, being exactly flat, is not very e~citing, it is not ~ i s p l a y e below. ~ Figure 5.7 instead shows what local volatility would be obtained if one instead naivelyfittedto the true m ~ prices. ~ To ~ begin t with, the resulting local volatility surface is far from flat on the W more noticeably, it also displays a large jump in the central section; sponds to the switch from call to puts. Note carefully that the ~agnitude of this effect does not die away for later and later time steps, but actually increases in ~ a g n ~ t u dNo e . such effect is ~ ~ e s e when n t option prices obtained from the Aatvolatility t ~ n o ~ itree a l are used instead of the true market price. Loo~ing at the poor quality of the results obtained using the naive, procedure, one can therefore im~ediately conclude that fitting to the proxy trinomial tree option prices, rather than to actual observed market prices, is a must, rather than a nice-to-have feature, in order to obtain reasonable local volatility surfaces. In what follows all the tests will therefore only deal with trees constructed using the r e c o ~ ~ e n d e d procedure. The next test is to explore how well the proach copes with an i implied volatility surface which displays pur dependence; see Figure As we saw in ~ h a p t e r 1, in the case of pure time dependence for the i ~ p l i e d volatility, one canuniquely obtain the corresponding instantaneous which is also stock-price i~dependent(see ~ h a p t e r l, Section 1.4). that this solution is unique. If the procedure worked perfectly, then there should be no dependence whatsoever on the stock level in the volatility surface. The input implied volatility fed into the D displayed a very mild (and decreasing, see the caption to Figure 5.8) gradient of the implied volatility function versus expiry: the overall increase in implie volatility is in fact little more than 1% over nine years. The numerical results for the local volatility surface are displayed in Figure 5.9, which shows the
u)
T-
g
I
c9
T-
LZ
- 6
The triangle indicates the boundaries ofthe trinomial tree. The local volatility in the area inside the rectangle is then shown in Figures 5.7, 5.9, 5.12, 5.14 and 5.16. Note that values of the local volatilities close to the top and bottom of the left-hand vertical side of the rectangle have a much lower probability of occurrence than values towards the centre, orthan values along the top or bottom hori~onta~ sides towards the right-hand part of the rectangle
local volatility for the portion of the t ~ n o ~ i tree a l schematically shownin Figure 5.10. It is apparent from Figure 5.9 that there exists a clear dependence of the local volatility function on the stock price. From the discussion above, this dependence must be a purely numerical artefact. This is not very worrisome at the very edges of the picture (which, as discussed in the caption to Figure 5.10, have very low probabilities of occu~ence), but more so in the middle and intermediate re For the verymild input impliedvolatility surface chosen, the effect appears rather small, but for more pronounced (and more realistic) time dependences it quickly becomes significantly greater. The twin test to which the DK methodology can be subjected then consists in giving as an input a purely stri~e-dependent implied volatilitysurface, as displayed in Figure 5.1 l. The corresponding local volatility is shown in Figure 5.12. Note that in order to obtain numerically stable results, a verymild smile surfixe (less than 0.50% in volatility from 0 to 90 delta) was again given as an input. The surface obtained is qualitatively very similar to the purely time-dependent case. Once again the whiskers at the top andbottom left-hand edges of the rectangle are neither very meaningful, nor partic~larlyworrisome. Apart from a small change in convexity between the two curves, the user would be rather hard-pressed in deciding which has been produced from a time-depe~dent,and which from a strike-dependent implied volatility input surface. Since we know
i ty
-a
N
that, for the purely time-dependent case, any price dependence is a pure artefact, and we find a dependence of the local volatility on the stock price of sinsilar magnitude in the pure-smile case, the user can have some legitimate doubts as to how much of the resulting structure is signal and how much numerical noise. The way to answer this question, of course, would be to repeat the tests after increasing the strike and time dependence of the inputs. It is not possible to do so, however, without encountering serious numerical problems. DK recommend (cumberso~e and not sure-proof) ways around these problems, but, rather than purs~ing this avenue, in the next chapter we present a much more efficient way to obtain the local volatility. Finally, the tests reported in Figures 5.13-5.1 8 deal with the case where the implied volatility surface depends on both strike and maturity, either for a shallower (Figures 5.13 and 5.14) or more pronounced case (Figures 5.15, 5.16 and 5.17). For Figures 5.13 and 5.14 the qualitative shape is, once again, quite similar, and the onlynoteworthy feature seems to be a change in the convexity of the resulting local volatility in going from the purely ti~e-dependentimplied volatility input (convex), to the s~le-dependent case (roughly flat) to the mixed time-de~enden~smiley case (concave). It is far from certain whether anything very meaningful should be read into this feature, since, as we pointed out, we h o w that in one case the effect is a pure artefact. In order to explore whether this trend in convexity is systematic, the results for the more pronounced mixed input case of Figure 5.15 are shown in Figures 5.16 and 5.17. These refer to the areas depicted in the two rectangles in Fi Note the change in conve~ity and the i the magnitude of the central trough in moving from shorter to longer times (Le. from the small to the large rectangle). For long maturities, this latter effect, which is certainly a nume~cal artefact, becomes of the same approximate magnitude as the gradient in the input volatility curve, these empirical observations it therefore seems fair to conclude that construction is conceptually very powerful and elegant, but of delicate ~umerical i~ple~entation. As the authors themselves indicate (D are indeed waystotry to obviate these problems. The fixes, however, are far from straightforward and cannot be guaranteed to work in realistic pricing environments. A completely different line of approach is therefore presented in the next chapter. Before e m b a r ~ n g onthis task, however, note thatwe have not even begunto assess to what extent the equations underpinning the DK procedure could provide, if implemented in a ~ u ~ e r i c stable al way, a viable and convincing explanation of the financial mechanisms responsible for smiles in different markets. In particular, we have not been in a position even to begin the tests that would allow one to d e t e r ~ whether ~e the future smile surface is of a sticky, floating or other nature,GThe difficulty of disentangling numerical
See Chapter 6 for a discussion of how this i n ~ o ~ a t i o isn obtained.
i 9 39L
l0928
2
LZ
-62
$ ? $ ?
0
o ! t
FT-
T-
F-
F-
FT " T "
e,
. -
Q)
. W
N T - T - c - "
The two areas of the trino~iai tree shown in Figures 5.1Q and 5.1 7
noise from the ~nancial signal has stood in the way of such an analysis. We shall show, however, that the method presented in the next chapter can produce results of suf~cient numerical quality that an assessment of the ~ ~ a n c icontent al of the rest~cted-stochastic-volatilityapproach does become possible. The taskahead is rather challenging: o~tainingprobability densities i s not verydifficult.After all, the quoted option prices, suit~blysmoothed, can give on the density. The reski -Rebonato (1999) approach 7 provides one such applic , for pricing exotics (e.g. in order e process using a Monte Carlo simulation), one needs local from a series of densities for different maturities today to the ity is indeed possible, but less straightforward than proceeding in the opposite direction (i.e. obtaining densities from a knowledge of drifts and in the previous chapter, lly, it is important to remernber that, as ~entioned oking at todays option prices one cannot say whether future i ~ p l i e d observed prices on any one volatilities are of a sticky or floating nature, since day are compatible withboth pictures of reality. servation in the real world of implied volatilities after movement f the underlying allow one to discriminate between the two world pictures ving empirically decided which world ~icture is a~propriate, one can then only tell whether a given model willproduce a sticky or floating future smile surface by ~ositioning oneself at a series of future ti~e/p~c states, e evaluating model prices from these new t i ~ e / p ~ c origins, e obtaining the future conditional implied volatilities, and c o m p ~ i n g them with
todaysobservablemarketsmilesurface.Onceagain,thedegree of n~merical noisepresentinthe D approachhasadvisedagainst unde~a~in analysisbymeans of the D tree. The issue of thestickyor flo the smile i s however fundamental and will therefore be dealt with in detail in the conce~tual~y identical, but n~merically superior, approach presented in the next chapter.
~e saw in the previous chapter that the DE;tree, despite o f f e ~ a n very ~ elegant and intellectually appeal in^ solution to the problem of pricing options in the presence of restricted-stochastic-volatilitysmiles, and of obtaining the local volatility surface, is prone to numerical difficulties and gives rise to a very noisy estimate of the localvolatility surface, showedthat part of the problem stems from the very fact that we do not a priori both the local ~ o ~ eof~the t s variable whose process is to be mapped onto the tree. This makes it diffic~lt to choice of the comp~tational check het her the (a priori a r b i t r a ~ ) the end, turn out to be app~opriate to handle the problem at hand. In indicated that, even for mildly ti~e-dependent volatilities, or for shallow smiles, rather the algorithm can easily fail to find a feasible solution,,or tends to prod~ce implausible shapes for the local volatility surface. do recommend a wealth. of n ~ m e ~ c tricks al to circumvent some of the oblems (see, for instance, the references i nce, however, these remedies are either c i s introduced in the present chapter, very general and eas~-to-imple~ent met~od by ~ e a l of ~~ s hich a s~ooth local volatility function can be easily obtained from a set of quoted plain-vanilla option prices, This t e c h ~ i ~ provides ~e a simple and powerful alternative to the construction of a tree such as be argued that, for the s ~ ~ ~ c u r~~ o co s fee ~ t ~ u c t the i n ~ ~ t uv re o ~ ~ t o fi the ~ i ~ ~n~erZy ~rocess? i n ~ it is nu~erically faster, much more stable and ~etter-be~aved than the trinomial tree methodology. Furth.ermore, if properly imple~ented,, this technique is virtually guaranteed to fail only if true arbitrage o~portunities exist a ~ o n g sthe t input prices. In this context, it is essential to note the qualifier true, which indicates that the arbitrage violations are not an artefact of the n ~ m e ~ c a
procedure9 but really indicate the presence of a pre-~~nsaction-costs arbitrage possibility in the market prices-see the discussions in Chapter 5, ection 6.6 below on this point. essential to point out, however, that the procedure described below does not provide a pricing engine, and the information about the local volatility must therefore be coupled with a suitable valuation al~orithm.Direct availabi the local volatility makes Monte Carlo an obvious choice for the latter. techniques could, however, be employed, such as the Explicit o r ~mplicit ifferences schemes (i~plemented, this time, in the 'traditional9 way, i.e. starting from owle edge of, rather than obtaining a v~iances). In pa~icular,the Explicit Finite larly well suited to the task of ca~culatin~ options by ~ackward ind~ction9 given its ability to cope in a robust manner with the case when the coefficients are either time- o r state-dep~ndent (see ~ i l m o t (1998)). t rrespective of the method chosen, the importa~tpoint is that the approach presented below conceptually de-couples the two distinct aspects of option pric in the p r e s ~ n c ~ smiles, i.e. (i) the extr~ction of from the market prices of cess, and (ii) the evaluation of the n e c e s s a ~ expectations from tot have cle enting themethod it is i ~ ~ o r t a n underwhich the approach can be i ~ p l e m e n t e ~ .
necessary s in order to impl~ment the metho he set of ~ s s u ~ p t i o n ely coincides with the assu~ptions underlying the reported in whatfollows in a concise form, not onlyfor clarity and ease of reference, but also from a sli htly different angle, i.e. from the point of view of links between local and implied volatilities. The reader is referred and 5 for a discussio~ of the issues related to the assumed market co~pleteness of the rest~cted-stochastic-volatility models. These assumptions are as follows.
ote that the volatility term cr(St, t ) at time t has been allowed to depend at most on time, and on the realisation of the un~erlying process at that particular point in time. As mentioned in the previous chapter, the function U ( * ) is a ~ e t e ~ ~ i ~ i function of a ~ t o c h ~ squantity t~c (S,) and time. A few c o m ~ e n t s are in order. bviously, the process (6.1) for the underlying cannot depend in any way on the strike at which we might have chosen to transact our option,
Another way to look at the matter is to recognise that the same process must be used to price options with different strikes, and cannot therefore contain any s~i~e-specific information. An i ~ ~ Zvolatility, i e ~ on the other hand, can depend on strike, because it is just a short-hand way to quote a price. It might be useful to recall again the de~nition of implied volatility given in Chapter 4 as the wro number to put in the wrongformula to obtain the right price,. With the procedu presented in this section, we are therefore trying to find the s t r i k e - i ~ ~ e ~ e n ~ e local volatilities that can reproduce the m ~ k e t - i ~ p l i e sd ~ r i k e - ~ e ~ eimplied n~e~t volatilities. There is nothing magic in the process: as we have already pointed out, strike-dependent implied volatilities are not true volatilities at all (neither a~erage nor i ~ s t a ~ t a ~ e o u and s ) , therefore are not linked to a time int instantaneous (local) volatility by any obvious relatio~ship.
. In ~ q ~ a t i o (6.1) n above there is one single source of risk, h

in the underlying stock, S. As discussed in Chapter S, Section implies that the markets are complete, and that, therefore, risk-neutra~valuation will yield a unique price for an option on S. There is a perfect functional relationship between the f ~ t u r e level of the stock price and the value of the future local volatility: at a given point time, if the value of S is known, then the volatility is perfectly known. as well. anis terminology, this is called the Effective Vola It is a special (li~iting)case discussion at the beginnin
. In realityweonlyknow prices at discrete maturities and strikes. For co~putational purposes we must therefore create a smooth and diffe the implied volatility as a function of strike and m a reasons that will become apparent later on, this implied volatility surface is re to be s ~ o o t h enough to allow for continuous derivatives with respect to t up to second order, and with respect to maturity up to first order, incl that, despite the fact that we advocated the use of a smooth price function in order the implied volatility surface to carry out the DK tree const~ction, required to display the c o n t i ~ u i of t ~ derivatives required in this c prove below that if the i ~ ~ l i volatility ed surface is not sufficiently urin the local volatility. Since the approac disconti~uities this chapter and the ode1 are conceptually totally equivalent, of these approach neither requires nor knows about the conti~uity derivatives may shed some interesting light on the origin of some of the numerically unsatisfactory results we observed in the previous chapter,
~~~
repeatedly mentioned above, implied volatilities are simply t way of quoting a price. Since, for the process described by quation (6,l), the lack formula in general nger provides a sol~tion of the f~ndamental Partial ~ifferential E~uation ( (6.2) shownbelowwith the call/pL~t boundary conditions, we can no longer impose
. As
e r words, if the process is of the form (&l), implied lack volatilities are t volatilities at all (neither instantaneous, nor average).
It is essential to point out again that ti~e-homogeneitycan be a or floatiness property of implied volatilities and of local volatilities, but stic~ness of the smile can only be properties of the i ~ ~ l~1~~~ i e ~ vul~tilit~e,s, not of the instantaneous (local) volatilities. It is a priori impossible to tell, from inspection of the local volatility function, whether the resulting future smile will be, for insta~ce,oati in^ or sticky. As ent ti one^ in the previous chapter, the only sureproof way to determine the nature of the future smile is to
l. obtain, by whatever eth hod, the local volatility surface consistent with the
chosen model and today9s observed market prices; . calculate ~ ~ t plain-vanilla ~ r e option prices using t ~ 'S local ~ ~ volatility y surface; 3. obtain the implied lack volatility from these prices; and 4. observe whether the resulting ~ t ~ implied r e volatility s~rface displays or lacks any of the features we might deem desirable (time homogeneity, stickiness, etc.). lation of future option prices can be accomplished either by a tree, and by using a node in the future as the new 'root' to compute prices, or, perhaps, by using a ado simulation started at a future point in time, z, for an arbitrary value of the underlyi~g,St, and with a volatility function given by a(S,, t>, t 2 z. Note that, from each node in the a fully tree, or from each possible future value of the stock, S,, there o~ginates implied volatility s t ~ c t u r eIt . is these implied volati~itysurfaces that can then be compared across times to maturity and across strike levels to ete ermine whether they are time-homogeneous andlor sticky o r floating. resent the method mentioned aving clarified these important points, w e to extract the local volatility surface. call the plain-vanilla option (call or put) C ~ , r (S); t , this indicates the price of a call or put evaluated at time t , when the stock price at time t has value S,
when the strike is K , and when the maturity of the option is T . Since ris~-neutral valuation is still possible, we h o w that, even if the Scholes f o ~ u l no a longer applies, the call price still satisfies the fu lack and Scholes PDE:
[ a p t + (r d)sa/as + ;G&S2a2/aS2 r]cK,T(t, S) =0
(6.2)
where d is the dividend yield, and all the other symbols have the usual mean Equation (6.2) is known as the backward (~olmogorov) equation, since it tell ~ ~ In order r to d use~this equation how the option price propagates ~ ~ c in time. one needs to know the distribution of option prices at a given future time. Typically the user will know this price distribution at expiry (the initial condition). This, however, is just a matter of convenience. If one knew the pricesof the option C K , r ( t ,S) at any generic time z T , one could work ones way backwards to today. Equation (6.2) constitutes the usual starting point for the discretisation of PDE approaches, either in the finite differences or in the trinomial-tree impleme~tation. Ifweknew the values of the derivativesinEquation (6.2), thenwecould combine this knowledge with the observation of todays option prices,C K , r ( t ,S), and solve for theonly u ~ o w n i.e. , the local volatility. The problemwith quat ti on (6.2), however, is that we do notknow a priori an analytic solution of the PDE for the option price C K , T ( t , S) which we can differentiate (we only lack and Scholes equation does not hold in this case, and that gammaandtheta derivatives are therefore inappro could be tempted to circumvent the p r o b l e ~ by making use of our of the implied volatility function to obtain prices, and differentiate the latter; unfortunately this would require the evaluation of the first and second derivatives of the call function with respect to the ~ ~ o c ~ and ~ r of ~ the c efirst , derivative with respect to time. As we discussed in Chapter 4, Section 4.17, in the context of the evaluation of the delta of a plain-vanilla option, these derivatives can be carried out only if we assume to h o w how the implied ~olatilityvaries as a functio~ of S and t , i.e. if a ~ i ~ ~ land /aS a ~ i ~ ~were ~ / known. a t This, in turn, is r ~oating (or other) nature tanta~ount to a s s u ~ i n g ~ o w ~ e about d g e the sticky o of the smile, and about the degree of time homogeneity (if any) of the implied volatility function. All these problems can be by-passed thanks to the existence of a dual equation, known as the forward (Fok~er-Planck) equation:
-=
Note how, in moving from the ~olmogorov to the Fo~er-Planckequation, the time to maturity, T , has replaced the calendar time, t , and the strike, K , has replaced the stock level, S. The Fo~er-Planck e~uation describes how the option price
propagates in time. Typically one knows the distribution density of the price today (a Dirac a), and one wants to discover how this density spreads out as time progresses, given the drift and volatility of the process. In complete analogy to the case o olmogorov equation, there is, however, nothing special about t a ~ ~ distr y'~ .in general, if one knew the distribution at anygiven future time, one could determine how this distribution evolves over time from this arbitrary future time origin. A derivation of Equation (6.3) is given our immediate purposes the essential feature to note is that the unk~ow call ~ function are now to be evaluated with respect to strike and time to Further~ore,even if we do not h o w the 'true' function (solution of s (6.2) and (6.3) and dependent, amongs other variables, on the 'true' ty a($,t)), we do know a 'black-box' (the lack function itself) function, ~ volatility, ~ which Z can ~ produce e the ~ (by de~nition) correct time, of the ~ es as a function of strike and time to maturity. As we shall see immediately ow, this is allwe needin order to be able to extract the local volatility function. ime we should note that, in the light of this disc~ssion~ the descripmechanism presented in Chapter 4, ~ection 4.7, nolonger
~~~~~~
he two equations (6.2) and (6.3) above hold for any t , 0 5 z 5 T. In particular, therefore, theyapply to t = 0 (today). In this case the F o ~ e r - ~ l a n c equation becomes
m the discussion above, we have assumed to h o w the prices of options of strikes and maturities via the implied volatility surface. In other words, we ow the quantity C K , ~ ( O S), as a function of K and T, and therefore we are in a osition to evaluate (numerically, but very accurately) the following derivative^:
ote that in evaluating the derivatives above one cannot simply use the closedck derivatives because the derivatives of the pricing function by the co~espondinglack-de~ved f o r ~ u l a e with the i ~ p l i e ~ volatility as an input. (Once again it is useful to recall that an implied volatility is simply 'the wrong number to put in the wrong f o r ~ u l a to get the right price'; and that any use of the i ~ p l i e d volatilities other than for obtai~in~g a price is
unwa~anted and unjusti~able. See again the discussion in Chapter 4, Therefore, in evaluating the derivatives (6.5) to (6.7) one ~ u s use t numerical de~vatives. F u ~ h e ~ owhen r ~ , taking the numerical derivatives with respect to the strike one must change in the Black formula not only the strike itself, but also the i ~ ~ l i e ~ ~ ~o l aw t is ia lf i~u ~~ cit i o o f ~the s ~ t~i~e). Once these derivatives have been obtained, one can return to Equation (6.4) and solve for the only unknown, Q&, to obtain:
This is the local vo~atilitythat will prevail at time T ( t = T ) if the future stock price is equal to K(St = K ) , given todays prices of the plain-vanilla options. ~q~atio (6.8) n we have therefore solved the problem of determining the lo volatilit without any tree constructi shall show explicitly below that, conceptually equivalent to t tirnation of much smoother for the increased efficiency in the numerical estimation of the local vol with the approach described by Equation (6.8) i s that one i s not imposi recombination of a computational tree, which is always a burdens om^ numerical constraint (see also Chapter 9, Section 9.5). Fu~hermore, the a wa ~ time-con sum in^ ancillary constraint of constructing a contrav~iate cussion in Chapter 5 , Section 5.4) is no lo S ( ~ . ~ ) - ( 6 . 8to ) determi~e the volatility s ly much simpler to obtain, but also faster to compute and more accurate. are shown in the following section.
st of allcheck that, if the implied vol~tilities are obtained from a time-constant, no-smile implied volatility function, then the resulting local volatilities are constant in time and space. This entry-level test, it will be recalled, was successfully passed also by the P> approach, butonly after purging the numerical errors by fitting to the proxy market prices (see Figure 5.7 in Chapter 5, Section 5.6). For this simple test, instead, the approac~ proposed in this chapter i ~ ~ e d i a t e gives l y virtually perfect results witho~t any need for an ancillary const~ction. oving away from the trivial (no-smile, constant-volatility) case, the techni~ue was shown to begin to produce considerable numerical noise already
Partofthederivativescouldbeevaluatedusingtheanalyticexpressionsforthe to evaluate nume~cally~ ~ i and~~ ~ i l ~ andtheBlackkappa,butonewouldstillhave See Chapter 4, Section 4.7. Since the procedure described above is extremely fast anyhow, one can afford to be rather careless in weighing which route provides the nu~ericaily superior approach.
/~
for the case of a no-smile, time-dependent implied volatility surface. In this case the correct resulting local volatilities must be constant across the space (i.e. stock price) direction, and mustshow (see Chapter 1) the time dependence implied by
(Note that, in quat ti on (6.9), the dependence of 5 on S has been dropped because we are now in a no-smile situation.) ~e have seen that if one attempts to recover these relationships with acceptable precision within the framewor~ of the tradirocedure, one already encounters rather awkward numerical problems e simple maturity-dependent-implied-volatility case, which is shown nvenience in Figure 6.1. sirnilar numerical problems are encountered when adopting the procedure this section, even whenthe matu~tydependence of the implied volatilities is much stronger than what could be dealt with by the As one can appreciate from Figure 6.2, which displays the results for the purely tim~~dependent case, the nume~cal quality of the solution is much hi the fact that the same switch from calls to puts is used in moving across the atthe-~oney level. In comparing the tw res one should also note that, in order to avoid numerical problems, in the case the maturity-dependence of the inputs had to be much more limited (see the captions of ~igures 5.8 and 5.1 1 in hapter 5). ~ r i ~ a ~ the ~ ctwo i esurfaces shown in Fi ures 6.1 and 6.2 look so ~ualitatively different that one might wellwonder whether, apart from nume~cal noise, the DK urface would indeed converge to the same solutio^ given by E ~ u a t i o (6.5)-(6.~). ~s n other words, apart from numerical noise, can one count on the uniqueness of the local volatility surface once the prices of the pl~in-vanillaoptions and the drifts (risk-less rate and dividend yield) are given? To answer this question, let us begin (see Appendix 6.1 for a proof) that the second derivative of a call price with respect to the strike is equal to the price probability density:
(6.10)
ecall that we have a ~ s u ~ to ed know the prices for all the p~ain-v~nilla o~tion (i.e. for options of any maturity and any strike). As a consequence, we know a Z E the densities, Once the drift has been specified (by no arbitrage): and if one accepts that the process for the underlying is described by E~uation (&l), then the solution to this problem is unique, i.e. there is only one process and one
If the drift is not specified, then the uniqueness can for a nice example. no longer be guaranteed. See Dupire (I 993)
m S
$$?
0 0
3
g
. W
0 9
I_.
CT
C
0 0 0 0
0 0 0 1 0 0 0
$$
$88
a ,
state-dependent volatility function that will produce all the densities. Therefore, again under the working assumptions, if we find a solution, we can rest assured that it is the s01ution.~It can also be shown (DK(1998) Reiner (1998)) that p ( t , S, t , S), the problem is equivalent to obtaining the transition proba~ili~ies, between different states at contiguous pointsin time, Therefore the DK tree and this approach are indeed conce~tually exactly equivalent. Giventhe theoretical equivalence of the two approaches, in so far as the distillation of the local volatility surface is concerned, and the clear n ~ ~ e r i c a l superiority (also shown at length in what follows) of the method presented here approach, we shall not compare in a sequential manner the local surfaces imp lie^ volati~itiesobtained as described above with the co~esponding by the DK construction for the cases of smile only, combined smile and matu- , rity dependence, etc. We shall instead attempt to develop in the next sections a er stand in^ of how several input implied volatilities relate to the ocal volatility surfaces, This task is ~ n d e ~ a k below en for several c o m ~ o n albeit , somewhat stylised, shapes of the srnile.
some intuitio~ about the link between the implied and the local es, this section will begin by presenting several possible shapes for the two and highlighting their most salient features. A.~ualitative e~planation of the reasons behind some of these features will be given in Section 6.5. Z n the ~eantime, several pointers are dispersed with the results presented below as a reminder that a particular feature or behaviour will be referred to in the followin discussion section. The shapes of the implied volatility surfaces examined below are rather are not supposed precisely to reflect any particul~r market on any ather, they attempt to display the salient features of specific mar~et e equity smile,, the FFXsmile o r the interest-rate smile). Given this choice, some of the smile features are introduced in a piecemeal fashion (e.g. the ~ a t u r i t y and strike dependences are sometimes intro~ucedseparately) in order better to illustrate the influence of both variables on the resulting local volatility function. It should be clearly stated that, as a result of this choice, the input implied volatility surfaces are s o ~ e t i ~ not e s very realistic (or even asymptotically ~nanciallycompatible with the absence of arbitrage). This i ~ p o ~ aissue n t is addressed in Section 6.8.
The class of financial markets that this type of smile is supposed loosely to mimic consists, for instance, of the FX options where both underlying FX rates are currencies of mature economies (the shape of the ~:emerging-market-currency smile is much more similar to the equity case). The salient feature of the FX market is not so much the symmetry, but the fact that the volatilities of out-of-themoney calls do not systematically trade at a p r e ~ u m or a discount with respect to the same-delta, out-of-the-money puts. The relative volatility of puts and calls (often quoted in the market in terms of the 25-delta calVput risk reversal4) can in reality be very variable, but, as a first approximation and after avera~ing over sufficiently long periods, one can say that the smile is roughly symmetric for calls and puts. The first case we want to examine is therefore that of an implied volatility surface displaying a s y m m e ~ c constant quadratic smile5 and with no timedependence, as shownin Figure 6.3. The result in^ local volatility surface is shown in Figure 6.4. The first important feature to notice is that, despite the fact that we started from a symmet~c implied volatility surface, the resulting local volatility surface is strongly tilted in the direction of the low stock prices. In other words, according to the model embodied by Equation (&l),future high realisations of the stock price will be associated with a decreasing volatility, and vice versa. At this point is tempt in^ to quote Blacks own words:
I have believed for a long time that stock returns are related to Volatility changes. When stocks go up, volatilities seem to go down; and when stocks go down, volatilities seem to go u p 6
shall return to this point at greater length later on. For the moment one can also notice that the local volatility for short calendar times (i.e. close to today) and very high o r low values of the underlying tends to be much higher than any of the implied volatilities. The qualitative reason for this effect is also explained in what follows. In the meantime it is instructive to compare the implied and local volatility functions for a given time (see Figures 6.5-6.7). Note that special care must be exercised in making the comparison: the x-axis represents the stock price for the local volatility function, and the strike for the implied volatility; on the yaxis one can read calendar time for the local volatility curve, and maturity for the implied volatility. Despite the fact that we seem to be comparing the proverbial
See Chapter 4 for a brief discussion of risk reversals and strangles in the FX markets, and of how they can be used to infer info~ation about expected future currency movements, and Cooper and Talbot (1999) for some empirical data relating to the $/Yen rate. We shall argue in Section 6.8 that this implied volatility function is not realistic if extended to longer and longer maturities. We are using it here as a didactic tool. Black (1976), quoted in Dumas et al. (1998).
E ~ i c iExt~actio~ ~~t of t
~t~~~ Local Vo~atility
e F ~ t ~ Local r e ~olati~ity
95.OOYo 85.00% 75.00% 65.00~~
55.00%
45.00%
35.00% 25.00% 15.00% 75

80
85 95 90
100
105
110
115
120
125
6.5 The implied and local volatility functions for time = 0.1 year. See the text about the meaning of time
45.OO0/o
40.00%
35.00%
30.00%
25.00%
20.00% 15.00% 8075

85
90
95
100
105
110
115
120
Same as Figure 6.5 for time == 0.4 year
apples and pears, we shall show later on that looking at the two lines together can be helpful in order to develop some i~tuition about the link between the two surfaces. We are now in a position to introduce ti~e-de~endence on top of the smile effect, as shown by the implied volatility surface depicted in Figure 6.8, ~e retain symmetry in the s ~ l e - d e ~ e n d e n cThe e . resulting local volatility is shown in Figure 6.9. Note that the same feature of very high local volatilities for low and high values of the underlying is still present at short times, but that it dies
10.00% 4
75
80
85
90
95
100
105
110
115
l20
125
.7
Same as Figure 6.5 for time = 1 year
off much more quickly with time. The convexity of the local volatility, always positive in the case of a time-independent symmetric implied volatility surface, appears to change sign for distant future times (t >2.5 years) andvery low stock levels; otherwise the surface appears to become asymptotically flat. See the discussion in Section 6.5 about this point.
The second shape for the implied volatility surface we are going to examine is the half-pa~abolicsmile shown in Figure 6.10. This stylised shape i s supposed to capture the most salient qualitative feature of equity smiles, or of smiles found in the FX option market for the dollar against emerging market currencies (in this case the high volatility is on the side of the ~cal~emerging-mar~et-cu~ency put). The co~esponding local volatility surface is displayed in Figure 6.1 1, together with the time cross-sections for 0.4 and 1 year (Figures 6.12-6.14). The most notewo~hyfeature of these curves is the fact that, despite the monotonically increasing implied volatility smile for strikes towards the in-themoney puts direction, the resulting local volatility function displays a very clear ~nimum in moving from high stock prices to low stock prices. This ~ i n i m u m , it will be argued later on, is not a numerical artefact, and becomes relatively more and more pronounced as time goes on (see Figures 6.11-6.14, where the usual caveat about the meaning of time applies). ~ntroducing a time-dependence in the implied volatility as shown inFigure 6.1 produces the local volatility surface displayed in Figure 6.16,
e F ~ t Local ~ ~ Volati~ity e
5
v)
0 ._. m
W
t
7
~ Local r ~ Vo~atility
I _
. -
Q,
v)
V
0 l U3
I ; '
~ i c i ~Extr~ction nt of t
35.00% 33.00% 31 .OO% 29.00%
25.00% 23.00% 2 l 00%

*
19.00%
ure 6 . 1 2 The implied and local volatility functions for time text for the meaning of time
= 0.1 year.See the
40.00%
35.00~0
~0.00~0
25.00%
20.00%
15.00% 80
85
90
95
100
105
110
115
120
3 Same as Figure 6.12 for time = 0.4 year
35.00%
~ 0 . 0 0 ~ ~
25.00%
20.00~0
15.00%
10.00%
75
80
85
90
95
100
105
110
115
120
125
Same as Figure 6.12 for 'time' = 1 year
The most salient features to note are
1. the change in convexity for long times and low stock levels; 2, the overall asy~ptotic flatten in^ of the local volatility surface; and 3. the fact that the ~ n i m u m inthe local volatility function for a given time of n the stock price is less pronounced, but still present (see as a f ~ ~ c t i o Figure 6.17).
A ~ualitative explanation of the reason why an implied volatility surface which
for o~t-of-the-mone~ is flat for in-the- one^ calls and monotonic all^ increa~ing calls can give rise to a local volatility with a ~ i n i ~ u is m given in the follow in^ subse~tio~.
Finally, we examine below the case of an onot tonically decreasing implied volatility surface, This surface is meant to describe a p p r o ~ i ~ a t e the l y shape of the smile encountered in the interest-rate world. As is clearly noticeable from ures 6.18-6.20 there are now no significant structural features in the local volatility curve (the very small 'blip' in the top ri~ht-hand corner of the threedimensional graph i s due to ~ u ~ e ~ c a l ~nt~oducing ti~e-dependence in
This 'noise' is almost certainly due to the fact that in order to evaluate the Black formula one has to makeuse of numericalapproximationsforthecumulative normal distribution.Whilst,in normal regimes, this is usually the last worry for option pricing, the power of the approximation becomes stretched when derivatives of such extremely out-of-the-money options are evaluated. The approxi~ationto the cumulative normal distribution used was found in ~ b r a i ~ o w i and t z Stegun (1972), whose upper bound for the absolute value of the error is l . 5x lw".
~ i c i e Extractio~ ~t of the Future Local Vo~ati~ity

29.00% 27.OO%
53
17.00%
1 5.00% 80
85 110 105 90100 95
1 1 5
120
the implied volatility surface does not bring about any unexpected features. The only noteworthy feature is that the steepness of the local volatility surface (against stock price) is more pronounced than the slope of the implied volatility surface (against maturity); see Figure 6.20. We shall try to understand this behaviour too in the following section.
in this section we want to gain an intuitive understanding of the links between the implied and local volatility surfaces presented in the previous section. in the reasoning presented in what follows to account for the various features reported n of thought. Namely: above we shall follow a c o ~ o line we shall imagine that we are performing a Monte Carlo simulation8 of the evolution of the underlying using the local volatilities obtained from the various implied volatility surfaces analysed above; we shall look in turn at plain-va~illa options with particular strikes and maturities; we shall try to determine which paths are mostlikely to be signi~cantly sampled (i.e. to provide a large pay-off) in the Monte Carlo simulation.
The practicalities of how to carry outa Monte Carlo simulation using local volatilities are touched upon in a latersection.Forthemoment we shall limitourselvesto a simplethoughtexperiment during the course of which Equation (6.1) is discretised as S,,, = Sold(1 p , ( $ & , t))ht Solda(&id, t)dhtE, where E E: N(0, l), and all the other symbols have the obvious meaning.
$ 8 8
0 0 0
5
$?$?$?$?$?$?&e$?
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
T-
22.00%
21 * oo /o
20.00~0
19.00~0
18.00%
The implied and local volatility functions for time = 0.4 year. See the n~ time text for the ~ e a n i of
Note that, ~ 1a ~0 art~ i c u l~ ~ r~ tthe h local , volatility is no longer a stochastic quantity, and can be thought of as a purely time-dependent volatility, with the time-dependence implied by the values of the stock along the chosen path. In other words, a ~ each o ~~ o io ~ f ~ this t ~ ~ we t can h write
a(t)
= a(&
t)
(6.1 1)
If, for a p~ticular plain-vanilla option, it were indeed the case that a few sirnilar paths contribute the largest weight in the Monte Carlo average, then it would be tempting to impose that the implied volatility for this plain-vanilla option should still be linked to the insta~taneous time-dependent volatility by the usual relationship 2 (6.12) ~ ( udu ) = ~ aBlackt
. I
where the integrand in the equation above is the time-dependent volatility along the impo~ant path. In general this will of course not be true: to begin with, for any finalrealisation of the stock price, and hence of the pay-off, the Monte Carlo simulation of our thought expe~mentwill produce a series of paths that will sample the local volatilities encountered from the starting to the final point, Furthermore, if one truly knew the path, it is not clear what meaning (if any) one can associate with the concept of volatility, If, however, for some special plain-vanilla options one can make the rather heroic approximation that it is sensible to concentrate on one or a few important paths that end in the money andprovide a significant pay-off,
xt~actio of ~ the ~
~ Local t Volatility ~ ~
then, along these paths, Equations (6.11) and (6.12) can provide an appro~imate tool tounderstand the shape of the resulting local volatility, Needless to say, the approach will be, in general, a very crude approximation. Nonetheless, for carefully chosen strikes and maturities, it can provide, as shown below, some useful insight and intuition. We start by examining the case of a symmetric, quadratic smile. In order to understand the qualitative features displayed in Figures 6.3-6.9, we shall focus attention on one pa~icular,and relatively short, maturity, say half a year, and consider at-the-money (ATM) options (calls to begin with). For the sake of simplicity we shall also ignore interest rates; therefore the at-the-money-spot and at-the-money- orw ward levels coincide, and have the value &l00 ( K ( The implied volatility for this call option is, say, 19.55%. First, we want to try to understand the shape of the local volatilities ~ o u n d the origin. If we have a parabolic smile as a function of strike, i.e. if
we have pointed out in the previous section that the local volatility for a s ~ ~ Z ~ ~ t ~ t (i.e. ~ iclose t eto today) will approximately di will be lower for St = K( strike K = S(0) (see for S will be me qualitative shape (see Figure 6.5); than the implied volatility for matu~ity t and than the implied volatility for matu~tyt
) will diverge as t approaches zero (see
er
Figure 6.5). order to make sense of this behaviour let us imagine that we are r ~ n ~ i n call option onte Carlo simulatio~ to value a short-~aturity at-the-~oney the local volatility obtained from the procedure described above. A few of the values of the local volatility sampled at time intervals of At = 0.1 are shown in Table 6.1.
(2100) at time intervals equal to 0.1 year
Implied 96 19.55% 1 00 1 04 19.99% 18.20% 16.81% 15.70% 1 08 25.79% 22.22% 19.82O/O 1 8.06% The local volatilities around today's price of the stock
Local volatilities 92 27.09% 24.37% 22.33% 20.73% 20.40% 18.86'/0 17.62% 16.60%
I a,34vO
17.01Q /' 1 5.94% 15.05%
Let us first try to explain the observation that, for relatively short t , the local volatility between 0 and t is lower for S, = AT^) than the implied volatility for matu~ty t and strike K = S(0) (see Figure 6.5). To begin with, all the paths that end above the strike will contribute a strictly positive pay-off, and the completely straight path contri~utes no pay-off at all, Given the short mat~rity, those paths that begin to develop below the strike are unlikely to give a very large final pay-off (they are actually quite likely to end up out of the money and to provide no pay-off at all). Therefore the espectation over the Monte Carlo pay-offs will mainly be affected by those paths that start from g100 and see the price level increase more or less steadily. Remember also that, for the p~ticular implied volatility surface considered, the local volatility increases on both sides of the at-the-money level, and that the local volatility is a unique function of the spot level. If the almost-straight-upwardpaths are indeed the important ones, then the average of the local volatilities (as given by qua ti on (6.1 2)) along these paths must approxi~ately equal the implied volatility. The implied in turn, is, by ~ e ~ n i t i othe n , price that would obtain if one experienced a purely time-dependent average volatility equal to 19.55%. In an imprecise but int~itively helpful way, and as long as time is relatively short, one can therefore think of the volatility for these s h o ~ - m a t u ~ options ty as equal to the average of the olatilities encountered along the paths themselves. If, for a ~ o m e n t we , take as a fact that the local volatility increases on ~ o t sides of the at-the-money level above the corresponding levels of the implied conclude that the local volatility volatilities (see Figure 6.5), we can i ~ e d i a t e l y paths must be greater than the local e~~erience along d any of these u p w ~ d volatility at S(0). This must be the case for the average of the local volatilities to produce the correct implied volatility, given that we have taken for that the local volatility rises on both sides of the at-the-~oneylevel above the i~plied volatility. In other words, for the average of the local volatilities to give rise to the correct implied volatility, either the local volatilities would have to be flat, or some of the local volatility values will have to be below and some above the average itself. Since we are temporarily assuming as a fact the monotonically rising shape of the local volatility, the first alternative is not acceptable, i.e. the local volatility cannot be constant. It therefore follows that, in order to pro the correct average, as one ~ a v e away s f r a the ~ origin alang any o f t ~ e i~~artant ~ ~ y - a ~ - c ~ n t r i~ bu in a tt the ~~slocalvolatili~ ~ u s be t ~ i ~t h ~~ ~ n e ~ r ~~ ~( 0 l )( ) .0 can at this pointturn to the explanation of the fact needed for the a r g u ~ e n t above, i.e. the sharp andmonotonic increase in the local volatilities for short times as one moves away in either direction from the at-the money level. Let us identify first the important paths. The shorter the maturity, and the the strike, the more the underlying stock has to move towards and beyo strike witho~t desc~bing too tortuous a path in order to provide a positive pay-off to the sim~lation.(More precisely, the l i ~ e l i ~ o o of d very t paths is very small, and these tortuous paths will very rarely occur in our Carlo thought
simplify the reasoning, we are therefore going to assume that, aturities, the impo~ant paths are virtually straight lines. Let us then consider two out-of-the-~oneystrikes, K1 and K2 (with K2 = and K2 more out of the money). Given the assumed shape and slope o the implied volatility corresponding to K2 i s greater than the implied volatility co~espondingto K1. F u r t h e ~ o r e ,let us recall that both implied can be approximately thought of as averages of the local volatilities straight paths. Let us suppose that we have determined the successi volatilities that produces the correct average for the less out-of-the-~oney option (struck at .K1) .The next option (with strike K:! = .K1 AK)has a higher implied volatility, and therefore the co~esponding average must be higher. terms in the average of the local volatilities up to K1 must be the same. Therefore the extra term (i.e. the local volatility for the stock level from K1 to K Z )must be sufficiently high to produce, by itseEf, the correct KZ-average; in particular, it will certainly have to be higher than any of the values encountered along the path to K I , Clearly, the more out of the l~oney K1 is, the higher the extra local volatility element will have to be to produce, by itself, the correct lity. This is indeed borne out, for instance, by Figures 6.5, 6.6 and 6.7. reasoning along the same lines, it i s also easy to see that, as the maturity of the option tends to zero, a constant parabolic smile implies a divergi volatility: the shorter the time to reach a given fixed strike, the higher volatility necessary to give rise to the correct average. This dive give rise to anypathological behaviour, since as time tends to zer of the stock price reaching a finite distance from the origin a This implies, however, that the local volatility surface must carefully as t tends to zero. In general, the log-normal character of the distribution of the underlying does create an a s y ~ eeven t ~ if the starting implied volatility surface were pe~ectly relatively close to the sym~etric. For small times, and values of the u~derlying at-the-money level, however, this asymmetry i s small, and what we said about a call option also applies to put options (recall that we have assumed zero i~terest
atters change when we move from a symmetric smile to a one-sided parabolic (see Figures 6.10-6.17). Clearly, if we are in the high-stoc~-priceregion every thin^ has to beflat: once we have moved deeply enough into th implied-volatility region, the local (percentage) volatili the level of the underlying and we are back in familiar Close to the at-the-money strike though, there is now an asymmetry between calls and puts. In the case of slightly out-of-the-money puts, for short maturities the same reason in^ as before still applies: the important, paths are the same, and it does not matter (much) what happe~s to the paths that begin to wander from th ds high stock levels, since they are unlikely to provide much pay-off t, once we have made a plausible argume~t, based on the pay-offs
of slightly-out-of-the-money puts, to link the local volatility and their implied volatility we must also account for the i d e ~ t i c implied ~l volatility for slightlyi~-the-money calls. In other words, same-strike calls and puts must trade at the same implied volatility (otherwise calVput parity wouldbe violated). For the short maturities we are considering, the paths that contribute significantly to the slightlyin-the~moneycall pay-offs are now those that start off in the high-stock-values ut if the local volatility surface were to keep the shape of the implied rve (as it approximately does in the case of symmetric parabolic smiles) and therefore to flatten off, then these impo~ant paths (for the calls) could not build up the correct average throughout their life: they wouldhave to start low, as argued above, and find a similarly low volatility all the way to their expiry. Therefore the local volatility must rise both towards the out-of-the-money-calls and towards the out-of-the-money-putsregion. This is indeed borne out by Figures 6.10-6.17, and explains the existence of the ~ n i m u m . This observation is important not just for the insight it gives about the origin of this local minimum, but also, and actually more, because it casts light on the possible future shape of the implied volatility function. As was mentioned in the previous sections, the future implied volatility surface cannot, in general, be gauged by inspection from the shape of the local volatility: the only correct way to estimate it is to carry out, by whichever numerical method one chooses, a discounted e~pectation starting from different future til~e/price nodes. In the particular case we have just e~amined, however, one can irnmediately see that, given the local volatility surface obtained using Equations ( 6 . ~ ) - ( 6 . ~not ) , even the qualitative shape of todays i ~ ~ l i vola~ility ed surface will be obtai~ed in the future unless the stock price happened to be close to todays at-the y level. We have, in fact, seen that the local m i n i ~ u m in the local volatility is necessary to account, at the same time, for the implied volatilities of both puts and calls, But this local ~ n i m u m is only encountered (see Figures 6.11 and 6.16) close to today, S at-the-~oney level. Therefore, we can indeed fit the prices today of plain-vanilla calls and puts implied by a volatility surface with a shape such as that depicted in Figure 6.10 using a model for the stock dynamics described by Equation (6.1); but, by so doing, we are i ~ ~ lthat, y in i general, ~ ~ thef ~ t u rie~ ~ l i e d v o l sa u~ ~i alciwill e ~ look ~ u a l i ~ a t i ~ e l y d i ~ e r e ~ t f what r o ~ it looks like today.This has profound effects on the future rehedging costs for, and therefore on the prices of, non-plain-vanilla options. This to ic will be discussed at great length in what follows. can finally analyse in a ~ualitative way the last case, i.e. the onot tonically decreasing implied-volatility surface shown in Figure 6.18. This volatility shape is often associated with the smiles found in the interest-rate option markets. The important thing to note is that the local volatility curve now looks virtually like a replica of itself at any future point in time and for any level of the underlying. In particular, ~ualitatively it looks, from any future tirnelprice node, almost exactly like the local volatility seen from the origin today. The latter, however, prices todays market options correctly (by const~ction). Therefore, by and large, one can
i c i ~~ ~xtt r ~ c to i fo t~
expect that this will also be the case in the future, and that the srnile surface will be reasonably time-homogeneous. Note that, if one looks carefully at Figure 6-29, one can see that the iso-localvolatility lines do not run exactly parallel to the time axis (as the iso-impliedvolatility lines do in Figure 6.18). This is probably due to the log-normal nature of the underlying process; the effect, however, seems to be rather small. ~ n l i k e the case of the one-sided implied volatility surface, it therefore appears that using the model described by Equation (6.l) for the monotonic-s~le case can give rise to a roughly self-replicating future implied volatility surface. In other words, in the case of interest-rate options, we do not have to believe that the w o r l ~ t o ~ o r r o w will be q u a l i t a t i ~ ~ l y d i ~ e r e todays nt~ro ~o r l ~ w (at least as long as plain-v~nilla options are concerne~) in order to use quat ti on (6.1) in a logically se~-consistent ~ a n n ~This r . observation also ties in well with the Smile Tale l presented in ~ection 4.3 of Chapter 4. In practical terms, the actual future re-hedging costs that we shall encounter during the course of the option will be similar to what our modelhasassumed in coming up with the cost of the option itself. As mentioned above, this topic will be analysed in greater detail in a and 6.8. Before underta~ing this task, however, it is important to highlight some absolute (as opposed to ~ ~ ~ d e l - d e p e n d e no-arbitrage nt~) conditions that must be satisfied by any implied volatility surface. Finally, it is importa~t to point outagain that some of the local volatilities used for il~ustrativepurposes in the discussion so far were neither very realistic, nor s . the wo~ld they have been arbitrage-free if extrapolated to longer ~ a t u ~ t i eIn follow in^ section we hig~light those conditions on the implied volatility surface that, i ~ e s ~ e c t i v of e the model, must be satis~edin order to avoid arbitrage. These conditions will also be of great relevance in the context of the discussion in Sections 6.7 and 6.8, where the case of the pricing of a continuous double barrier option is treated in detail.
In this section we explore some simple constraints on the implied volatility surface by looking at it as a function of strike and maturity. The purpose of the analysis is to show the implications of these constraints for the res~lting local volatility functions. It is shown in whatfollows that the importance of these observations is very far-reach in^ in so far as the ability of the restricte~-stochasticvolatility models to reproduce in the future realistic smile surfaces is conce~ed.
See the discussion at the end of Section 6.8 for a clear definition between absolute and modeldependent arbitrage.
miles can raise the relative value of out-of-the-money options with respect t-the-money options. This increase, however, can never produce a price for a given option greater than the price for a more in-the-money option with the same expiry on the same underlying. If the condition fails, then there clearly exists the possibility for an obvious arbitrage. With this first caveat in mind, let us examine ~ a appear plausible; an implied volatility smile surface which might, ~ r i facie, see Table 6.2 Note that there is a time-dependence of the at-the-money volatilities, but the smile's increase (~uadratic) is the same across ~aturities,i.e. the implied volatility is of the form
O(T>==: uATM(T)
+a
(~
(6.13)
where K is the strike, CFATM(T) is a function of time to expiry only, a is a constant, and SO is the value of spot today. Let us compute the prices for calls and puts that this surface implies (for simplicity we shall assume zero interest rates; this will not affect our conclusions because we shall compare, in the first instance, options with ~ i ~ e rstrikes e ~ t and the , ~ a expiry); ~ e see Table 6.3. In the right-hand columns of the call price matrix and in the left-hand columns of the put price matrix we find clear violations of the price condition, i.e. prices for out-of-the-money options rea at er than prices of more in-the-money options with the same expiry. One might be tempted to conclude that the reason for this arbitrage possibilit is simply that we have allowed the smile togrow too sharply. If one wan d to retain the quadratic behaviour of the smile with strike, then one could try to reduce the constant a. This is shown in Table 6.4.. One can then check the call and put prices obtainable e new model implied volatility surface (see Table 6.5). spite the fact that the smile is now much less pronounced, the ~iolation areas hted in bold in Table 6.5) have receded to the long-~aturity part of the but have not disappeared. From this simple example, one can therefore conclude that decreasing the constant a would not provide a lon by increasing the maturity the (relative) effect of a fixed difference in implied volatility becomes progressively greater. Therefore, given a ~~e~ difference in i~plied volatilities, one can always find a maturit~lon e n o ~ g h for the out-of-money option to be worth more than a more in-the-money option, i.e. for the price violation to appear. We have therefore reached the conclusion that, in order to avoid a model~independent arbitra e,lO t ~ e s ~ i ~ e ~~~~t s t e~ee~ c~ re e~ s ,s ore precisely, it must decrease with ~ a t u ~quickly ty nough to (more than) compensate for the increase in relative value due to the increase in maturity. eeping this firstresult in mind, one can try to model a better implie surface by i ~ p o s i n g that the quadratic proportionality coefficient should decay
lo
See Section 6.8 for the distinction between ~odel-dependent and m~del-in~epend~nt arbitrage.
ojC6ojn'
M M M M
19\9u3iz>
M M M M
Table 6.3 Call and put prices obtained for the implied volatilities and maturities in Table 6.2. The areas for which arbitrage violations are encountered have been highlighted in bold
Call prices 80
20.1 5359 20.72 602 21.41 095 22.09785 22.75789 23.3839 23.97536 24.53393 25.061 9 25.561 6 26.03523 26.48481 26.91 21 1 27.31 874 27.7061 2
Put prices
84 16.1 625 16.6987 1 7.3 1 071 17.91 11 18.4801 6 19.01 449 19.51 527 19.98488 2 0.42 59 1 20.8408 21.231 74 2 1.60067 2 1.94928 22.2 7909 22.591 4 84 0.1 625 0.6987 1.3 1071 1.91 1099 2.4801 56 3.01 4493 3.51 527 3.984881 4.42591 2 4.840801 5.231 74 5.600666 5.949283 6.279087 6.591 396
88
12.2251 1 12.79582 13.38818 13.94505 14.45964 14.93433 15.3731 1 15.77986 16.15798 1 6.5 1038 16.83951 17.14743 17.43594 17.70655 17.96058 88 0.2251 1 0.795824 1.38818 1.945053 2.45964 2.934328 3.373106 3.779858 4.1 57985 4.51 0384 4.839506 5.147432 5.43593 7 5.706546 5.960581
92 8.4281 99 9.1 22796 9.7503 5 7 10.30585 10.80146 11.2478 1 1.65292 12.0229 12.36242 12.67514 12.96402 13.23145 13.4794 13.70955 13.9233 92 0.4281 99 1.122796 1.750357 2.305846 2.801 46 3 247802 3.652921 4.0229 4.36241 7 4.6751 44 4.964022 5.231 446 5.479401 5.70955 5.923305
96 5.0271 36 5.91 1334 6.607665 7.1 893 1 9 7.691 275 8.1 33259 8.527893 8.883007 9.205249 9.498937 9.767556 10.01 387 10.2401 1 10.444812 1 0.63 944 96 1.0271 36 1.91 1334 2.607665 3.189319 3.691 275 4.1 33259 4.527693 4.883007 5.205249 5.498937 5.767556 6.01 3869 6.2401 1 1 6.44481 22 6.63944 1
100 2.497504 3.495777 4.2371 08 4.841 454 5.355787 5.804449 6.202045 6.5581 88 6.87961 8 7.1 71 291 7.436984 7.679657 7.901 688 8.105022 8.291 275 100 2.497504 3.495777 4.2371 08 4.841 454 5.355787 5.804449 6.202045 6.5581 88 6.87961 8 7.1 71 291 7.436984 7.679657 7.901 688 8.1 05022 8.291 275
104 1 .I 13025 2.043285 2.772 675 3.380865 3.9053 1 7 4.366643 4.77822 5 5.1 48932 5.48503 5.791 29 6.071 369 6.3281 59 6.564001 5.780821 6.98022 9 104 5.1 13025 6.0432 85 6.772 675 7.380865 7.90531 7 8.366643 8.778225 9.1 48932 9.48503 9.791 29 10.07137 10.32816 10.564 10.78082 10.98023
108 0.553261 1.359627 2.0681 09 2.688082 3 2 3 7781 3.730883 4.1 77233 4.5 84064 4.956838 5.299795 5.61 6298 5.90907 6.1 80351 6.43201 3 6.665642 108 8.553261 9.359627 10.0681 1 10.68808 1 1 23778 11.73088 12.17723 12.58406 12.95684 13.2998 13.61 63 13.90907 14.18035 14.43201 14.66564
112 0.372045 1 .I 351 1.87499 2.551 039 3.1 66305 3.727883 4.243675 4.71 9461 5.1 601 03 5.569551 5.95 1 044 6.307272 6.640486 6.952601 7.245253 112 12.37205 13.1351 13.87499 14.55104 15.1 6608 15.72788 16.24368 16.71946 17.1601 17.56955 17.95104 18.30727 18.64049 18.9526 19.24525
116 0.34821 9 1.18145 2.0321 82 2.82831 8 3.563603 4.24281 1 4.8722 1 7 5.457608 6.003931 6.51 5338 9.995306 7.446766 7.872205 8.2 73 749 8.653235 116 16.34822 17.18145 18.0321 8 18.82832 19.5636 20.24281 2 0.87222 21.45761 22.00393 22.51 534 22.99531 23.44677 2 3.8722 24.27375 24.65324
120 0.415336 1.433003 2.482485 3.470462 4.387404 5.2381 48 6.02986 6.76931 5 7.4 623 7 1 8.1 33971 8.728276 9.308797 9.85852 10.38001 10.87546 120 20.41 534 21.433 12.48243 23.47046 24.3874 25.2381 4 26.02986 26.76931 2 7.4623 7 28.1 1397 28.72828 29.3088 29.85852 30.38001 30.87546
80 0.1 53588 0.72601 9 1.410945 2.097855 2.757894 3.38389 6 3.975357 4.533933 5.061 9 5.561 6 6.035235 6.484807 6.91 21 1 7.31 8741 7.7061 21
Table 6.5 Call and put prices obtained for the implied volatilities and maturities in Table 6.4. The areas for which arbitrage violations are encountered have been highlighted in bold
Call prices 80
84 16.07463 16.41456 16.85222 17.30368 17.74298 18.161 88 18.55821 18.93207 19.2844 19.61642 19.92943 20.22466 20.50324 20.76623 21.01 457 84 0.074629 0.41 4564 0.85221 6 1.303677 1.742978 2.1 61 879 2.5 583 1 2.932068 3.284395 3.61 6424 3.929435 4.22 3659 4.503252 4.76623 5.01 4573
88
12.1471 4 12.59201 13.08196 13.5536 1 3.9946 1 14.40404 14.78381 15.13643 1 5.46436 15.7698 16.05469 16.32072 16.56935 16.80187 17.10939 88 0.1 471 39 0.592 007 1.081964 1.553601 1.994605 2.40403 7 2.783808 3.13643 1 3.464357 3.769796 4.054688 4.32071 8 4.569353 4.801 869 5.01 9386
92 8.366428 8.999966 9.581 903 1 0.1 0009 10.56359 10.98141 11.36067 11.70685 12.02422 12.31 61 6 12.58531 12.83319 13.06436 13.27747 13.47486 92 0.3 66428 0.999966 1.581 903 2.1 00087 2.563591 2.981414 3.36067 3.706847 4.0242 1 7 4.3 1 61 63 4.58541 4.8341 9 5.064357 5.2 77472 5.474862
96 5.001 237 5.870952 6.556597 7.1 2941 2 7.623 673 8.058757 8.446883 8.796356 9.1 13132 9.401 666 9.665397 9.90705 1 10.1 2883 10.33256 10.51976 96 1.001227 1.870952 2.556597 3.129412 3.623673 4.058757 3.446883 4.796356 5.1 13132 5.401 666 5.665397 5.907051 6.1 28834 6.332561 6.51 9755
100 2 A97503 3.495777 4.2371 08 4.841 454 5.355787 5.804449 6.2 02 045 6.5581 88 6.87961 8 7.1 71291 7.436984 7.679657 7.901 688 8.1 05022 8.291 275 100 2.497504 3.495777 4.2371 08 4.841 454 5.355787 5.804449 6.202035 6.558188 6.87661 8 7.1 71291 7.436984 7.679657 7.901 688 8.1 05022 8.291275
104 1.085642 2.000924 2.71 9237 3.31 8258 3.83462 1 4.288879 4.693938 5.058543 5.388958 5.68986 5.964856 6.2 1 6801 6.448005 6.660369 6.855485 104 5.085642 6.000924 6.71 9237 7.31 8258 7.833621 8.288879 8.693938 9.058543 9.388958 9.68986 9.964856 10.2168 10.448 10.66037 10.85548
108 0.479076 1.21 938 1.878981 2.459003 2.974242 3.336778 3.855342 4.23 65 74 4.58551 4.906082 5.201 424 5.474085 5.7261 68 5.959437 6.1 75386 108 8.479076 9.21 938 9.878981 10.459 10.97428 11.43678 11.43678 12.23657 12.58551 12.90608 13.20142 13.47408 13.7261 7 13.95944 14.17539
112 0.2571 09 0.871 653 1.496166 2.077269 2.61 0603 3.099965 3.55021 8 3.965825 4.350592 4.70771 5 5.039871 5.34931 5 5.63 7964 5.90746 6.1 592 18 112 12.2571 1 12.87165 13.49617 14.07727 14.6106 15.09996 15.55022 15.96582 16.35059 16.70772 17.03987 17.34932 17.63796 17.90746 18.15922
116 0.1 8436 0.759082 1.400492 2.022682 2.607937 3.1 541 74 3.663405 4.1 38607 4.58278 4.998669 5.388703 5.755004 6.099424 6.423582 6.728898 116 16.1 8436 16.75908 17.40049 18.02268 18.60794 19.1 541 7 19.66341 20.1 3861 20.58278 20.99867 21.3887 21.755 22.099442 14.42358 22.7289
120 0.1 74246 0.787743 1.505028 2.21 6094 2.894044 3.533095 4.1 33673 4.698079 5.229053 5.729281 6.201 231 6.6471 14 7.068892 7.468298 7.846871 120 20.1 7425 20.78774 2 1.50503 22.31 609 22.89404 23.53309 21.13367 24.69808 2 5.22 905 25.72928 16.20123 26.6471 1 27.06889 27.4683 2 7.84687
20.04976 20.34583 20.77089 21 2 3 1 71 21.69335 22.14235 22.57353 2.9851 9 23.3771 5 23.74992 24.1 0432 24.441 26 24.761 67 25.06643 25.35461
Put prices
80 0.049 764 0.34583 0.770893 1.23171 3 1.693351 2.142348 2.573528 2.9851 85 3.3771 47 3.749924 4.1 04325 4.441 263 4.761 666 5.066432 5.356409
with time s~litablyfast. A possible description of the functional form for the implied volatility function embodying these desiderata could therefore be
o(1~7) =~ATM(IJ~)~ ( I J ~ ) ( ~
so)
(6.14)
a(T) = ao exp(-Piir)
The surface thus obtained is shown inFigure 6.2 1, For this surfxe, one can immediately check that the call and put prices now behave correctly (see Table 6.6). The functional forms presented above for the smile and for the decay constant are not supposed to show a particularly realistic fit to observed market data. Given their very clean form, however, it is possible to analyse the ~ualitative behaviour they implyfor future option prices. In pa~iculal. is it very instructive to e x a ~ n the e local volatility surfaces obtainable using the method described in the previous sections using good and bad implied volatility inputs. in pa~icular,shows the local volatility surface obtaine~ with the first im~lied volatility rnatrix (Table 6.2)-which we know to admit arbitrage. The first observatio~is that the resulting (unacceptable) local volatility surface bas lost the symmetry of the parent implied volatility function (recall that we started from a quadratic smile); this i s due to the log-normal distribution of the underlying price process. The surface appears to be slightly decreasing over time, and, by inspection, could produce a srniley implied volatility surface in the future (whether floating, sticky, or mixed is impossible to saywithout carrying out the calculations described in the previous section: recallthat the sticky or oati in^ nature of a smile is always a feature of the lie^, not of the local, volatility surface). A priori, i.e. once again, simply judging from the shape of the surface and short of calculating future option prices, it appears that whatever future smile might exist it will be less pronounced in the future. We can thenrepeat the calculations described in the first sections of this chapter in order to obtain the local volatility surface using acceptable implied volatility surfaces such as the ones obtainable using ~ ~ u a t i o n (6.14) s and (6.15). Two such surfaces are shown in Figures 6.23 and 6.24, The results displayed in Figure 6.23, in particular, were obtained with the acceptable implied volatility surface of Figure 6.2 l. It is immediately obvious that, at the short-time end of the curve, the qualitative behaviour is quite similar to the one obtained for the unacceptable input implied volatility surfaces. omp paring Figures 6.23 and 6.24 one can also notice that, for long times, there remains a very m o ~ e s but t ~oticeab~e depende~ce on time (illcreasing in Figure 6.23 and decreasing in Figure 6.24). F u ~ ~ e ~ o the re surface is more symmetric across the strike direction. The fundamental feature displayed by these two figures, however, is the fact that the dependence of the local volatility on the stock price has all but disappeared for maturities beyond a few years. Therefore the local volatility surfaces obtained from ~ a r a ~ o l i c i ~ p l i volatilities that have been rendered acceptable by introducing a decaying smile can ~ r o ~no ~ s& ~ ie l e at s all in the ~ u t ~ r e !
$ ?
0
+ 1
2 3 g
$?$?
"1l
0 0 O 01
I
0
I
N m
II: . . .
x
W a , 1J
nt
e ~ ~ t Local ~ r V e o l a t i l i t y
Y-"""
"
" " "
. -
8
CJ
h
Q
This observation is very impo~ant, and the study with the very clean implied volatility parabolic surfaces used as inputs for the above discussion points to the following conclusions.
. In order to have acceptable prices, i.e. in order to avoid higher prices for more out-of-the-money options and, therefore, arbitrage, the implied volatility smile surface must display a decreasing degree of smiliness as a function of maturity.
. If one introduces a fattening smile surface, as required by Conclusion 1, then the future local volatilities that can be extracted from it tend to lose their price dependence: they only (and very mildly) depend on time.
. As a consequence, these local volatility surfaces seem unable to produce future implied smile surfaces with any structure (or, rather, they seem to be only able to produce flat ones).
Therefore this state of affairs implies that todays smile surfxe is not only not time homogeneous, but is destined to disappear! Admittedly, the results reported above have been obtained with a particul~ behaviour of smile flattening (decaying exponential versus maturity, as shown in (6.15)). Also, the smile for the shortest matu~ty was arbitra~ly (albeit ~~uatio n ~lausibly) chosen. Therefore, strictly speaking, the conclusions reached above are not general, even for purely parabolic implied volatility surfaces. It is nonetheless dif~cult to see how different patterns consistent with the required decay of the smile surface could produce results that are qualitatively truly different. These observations, and onc cl us ion 4 in particular, have profound im~lications of the volatility for the users of mar~et-calibrated models based on a speci~cation surface purely as a deterministic function of time and of the underlying. Note that no assumption was made in obtaining the local volatility functions from the implied volatility surfkce other than assuming that
(6.16) If the conclusions reached above are indeed as general as theyappear, then users of this class of models with implied volatilities of the FX type are therefore implicitly end or sin^ a view of the world where future re-hedging costs (as predicted by the model) are materially different from those that will be encountered during the life of the option. This has obvious9and potentially very dangerous co~se~uences for the risk manage~ent and the very pricing of the options concerned. This topic is explored in detail in the next two sections by focusing on a pa~icular option i n s t ~ m e n(the t contin~ous double barrier), and by drawing from the analysis of the results some general conclusions about option pricing in the presence of smiles. efore e m b a r ~ n g on this discussion, however, it is interesting to report the results of empirical work carried out by Dumas et al. (1997) on the hedging
IT
= IT(&, t )
performance of the DK, Dupire (1994) an ubinstein (1994) approaches, These researchers examined the predictive and h ing ~erformance of these restrictedstochastic-volatility odels in the case of equity (S& period from June 1 8 to December 1993. Theyused ~choles ( ~ ~ MO ~ 3 as ) a bench~arkstrategy against of the more complex restricted-stochastic-volatility app The latter was implemented with an implied volatility input which was smoothed across strikes and maturities. Despite the fact that Dumas et al. come from a very different angle, they reach very similar conclusions to the ones ex equity options are concerned, regarding the hedgin approach. In particular, the study mentioned abo conclusions that the more parsimonious lack and Scholes modelworksbest in the sample, at least on the basis of an information criterion m~ntioned in their paper; the prediction errors obtainable from th estricted-stochastic-volatilitymodel are larger, compared with the Black and holes case, when the extracted local volatilit function is used to value options one week later; ratios ete ermined by the Black and Scholes model appear more an the corres~onding quantities obtained using the more complex restricted-stochastic-volati~itymodels. n the basis of their study, Dumas et al. conclude that simpler is better. The observations presented in earlier parts of this section, although arrived at by following a very different line of reasoni~g,concur with this conclusion, with the proviso, however, that simpler might well be better, but even this better might not be all that good. I am not aware of empirical studies similar to the one conduct~d by Dumas et al. for FX and interest rates. On the basis of the results of the tests reported and discussed in Sections 6.3-6.5 of this chapter, one might surmise that similar 1 .S were conducted conclusions could be reached if a study such as Dumas et a interest-rate area the restricted-stochastic-volatility for the FX case; but that in the might provide a more realistic, and useful, pricing tool. analysing a pricing approach potentially suitable for interest-rate option (Chapter 7), and before indicatin~which modellin provide a better description of the FX and equity option marke case analysis is presented in the closing sections of this chapter of an importa~t study. Apart from its intrinsic interest, the accompanying isc cuss ion can provide uidance in practical a~plications.
r , so~etimes, boxes) are instru~entsthat pay &l at maturity unless a barriers o
ici
reference rate (often a spot rate) touches or breaches two pre-set barriers at time throughout the life of the option. If either barrier is touched or n no payment takes place. Clearly, for the value of the continuous double barrier tobe non-zero at any point in time, the current value of the ~eference rate must be greater than the lower barrier and smaller than the higher barrier. It is not ~ifficult to write the PDE that must be satisfied by the price of the option; whzt is more comp d is finding a solution that matches the initial and boundary conditions. is greatly assisted in the search for a solution by the L ~ n i q ~ ~ ntheorem, ess ,inverynon-technical terms, states that, if we and initial conditions, find a solution to the PDE satisfying both the b o u n d a ~ we can rest assured that it will be the solution. ~nalytic se~i-closed- for^ equations do exist in the literature for the doublebarrier problem, which enjoy different de rees of general it^ no^-constant interest onstant volatilities, etc.; for a c o ~ p e n d i u of ~ useful formulae see 1). In this section I describe a general semi-analytic method to o b t a i ~ an a ~ p r o x i ~ a t i n series g of portfolios satisfying with arbitra~y precision both the initial con~itionsand an increasin~ discrete s~b-set of the continuous boundary conditions. The technique, known as static portfolio however, t, what 1believe are useful extenreplication is not new. I shall ~ ~ e s e n sions: the solution obtained below will hold, in fact, not only for the non-flat term s t ~ c t u r e of inter~st rates or ti~e-depende~t volatilities, but also in the case of sticky o r floating smiles. ~espite the fact that the a l ~ o r i t h r~ ~se~te below d will be cast in terms of a ategy, it should only be ed, in my opinion, as a means towards a very ~ e ~ eand ~ aintui l transparent sol~tion, rather than as an actually i ~ p l e ~ e n t a b trading le strategy. Its mainmerit, I shall its ability to account for the smile effect, andtomake the prici between the sticky and oati in^ smile regimes particularly clear. Apart from the int~nsic interest, the discussion will clarify s future re-he~ging costs of option strategies i and should therefore be o modelling approaches, a e apparent sim~licityof the approach, the static portfolio replicaactually involves some very subtle pointsin option pricing. The discussion of these issues can quickly bring the reader into rather deep stochasticcalculus waters. I have chosen, instead, to follow a less direct but mathe~atically simpler and intuitively more transparent route in order to explain both the reasons for the need for a careful treatment, and the solution of a simple paradox
A couple of techni~al de~nitions: the ~~~~~u~ ~onditions must be s ~ t i s by ~ the e ~ solution for all values of the und~rlying variable at the final time (optional expiry); the ~~~~~~~yconditions must be satisfied by the solution at any point in time for values of the underlying equal to either of the barriers.
introd~ce below, ~ efore isc cussing the static replication a~proach, it is therefore essential to clarify some i m ~ o ~ a points n t linked to the finite cost (even in the absence of bid-offer spreads and other frictions) in u~winding the r e ~ u i r e ~ hedging ~ositions. This is done by means of the example below.
he po~folio built using the procedure described above will ce~ainlyproduce the required pay-off if the underlying stock price ends up anywhere between t ever vt o ~ u ~ e~ ~ i te ~~ b ~ e~ o u ~ t ~ ~a ~ o &l00 and El 10, w ~ t ~~o ~ a ut what happens if the stock price breaches either b The trader who put in place the static strategy made up of a very of calls and puts will have to unwind all these positions as soon a t Figure 6.26 one can i~mediatelysee, in fact, that, either barrier, the value of the piain vanill the value of the double barrier rapidly diverge, In this context th the finite perforation cost becomes extremely useful: once again, not be able to catch9 the stock price exactly at E100 (or &l10) a cost owing to the finite differenc between the actual transa explored e theoretical unwinding level, As in the mo~ey-for-nothi ~aradox above, by m a ~ n g the tra~ing interval smaller and smaller e erf for at ion cost (wl~ich, as we argued, depends on a J A t ) can be made arbitr~ily small. Unlike the previous example, however, i~creasing the trader to more and more potential crossi of the barrier leve to more and more small losses): with the tinuous-double-barrier case, once level is touched, the option is dead, the finite perforation cost is once and can therefore be truly reduced (in our friction-~essworld) without ~ a v i n g to pay any price. There stili r e ~ a i n s , however, o of calls andputs (displayed in ad to c a l c ~ ~ a the t e expected conditions at any of the i n t e ~ e isc counted value of the later pay-offs contin~ent upon the stock price bein at barrier. This e x ~ e c t ~ t could i o ~ be p e r f o ~ e d ana~yticallysimply by u ack formula. The volatility that entered this calculation was t~erefor ~ t ~volatility r e for the stock price at thechosen intermediate time, If the volatility perfect~yd ~ t e r ~ i n i s t i then c 9 the market in plain-van ill^ op ive us information about the i ~ p l i e d future volatility (see but would also provide the trader with the i~struments to lock in this future volatility. In a v et er minis tic-volatility world, therefore, givin option prices for different maturities is equivale~t to providin at all the intermediate expiries. ces s t a ~ i n g resence of smiles, however, the ~roblem is much more ~omplicated. In order to tackle the roble^ in the presence of s ~ i l e s the , trades could take either of two differe~t routes: the first one (model-~ased) would be to
l. start from an assumptio~ for the dynamics for the underlyin instance9the one described by ~ q u a t i o (6.1) ~ in Section 6.2;
2. determine, using the prices of traded plain-vanilla options, the local volatility ( $ , ,t ) ; surface a 3, evaluate the future discounted values of the pay-offs (i.e. the future option prices) by using the chosen numerical method (e.g. a trinomial tree if the DK construction had been used to extract the local volatilities, or, for instance, a Monte Carlo procedure) and by ~lucing h i ~ s ut e eithe~ ~ b a ~ at n t ~ ~e ~ppropr~ate f ~ t u r point e in time; 4, solve the 2 x 2 system on the basis of these future option values in order to determine the required amounts of calls and puts.
Note that, given the assumption in point 1, recovering the future option values obtained from a given chosen model is as essential in order to prevent arbitrage as the recovery of t ~ ~ ~ prices a y of ~ plain-vanilla options. In other words, if the ~ o ~were e l correct, assuming any other value for the future values of the plain-vanilla options would be an arbitrage violation no less severe than using a non-market price for a spot-starting plain-vanilla option. The only difference between the present and future prices, in this respect, is that different models will predict different (model-dependent) future values for future option prices, but all (correctly calibrated) modelsmust accept the same spot option prices on immediate penalty of arbitrage. Therefore the arbitrage violation incurred by assuming future option prices not identical to the ones predicted by the model is model-contingent, whilst failure to recover spot option prices is a modelindependent arbitrage violation. The second (theoretically more dubious, but practically pro~ably safer in practice) strategy would be the following: 1. The trader could make an assumption about the nature of the srnile (sticky or floating) and about its degree of time-ho~ogeneity. Let us assume for a ~ o ~ e that n t the trader believes that the srnile surface (sticky or floating) should be time homogeneous. 2. Given this assumption, the trader could then calculate the future value of the e~pectation of the discounted pay-offs by enforcing directly the sticky or float in^ assumption about the smile and using the Black formula with the appropriate f ~ ~ t u implied re volatility as an input. In particular, if the trader assumed the snde to be floating, she would translate the smile surface to either barrier, and the options would therefore have either todays at-the-money implied volatility, or the volatility corresponding to todays option with a strike out of the ~ o n e by y an a ~ o u nof t pounds equal to the barrier; if the trader had assumed a sticky smile, the relevant vo~atilitieswould simply correspond to todays volatilities for the upper and lower barrier levels. Note that assuming knowledge of the future smile surface greatly simpli~es the task of producing the future option prices, since a smiley volatility is the wrong numberto put in the wrong formula to get the right price, and the lack formula can therefore be employed by virtue of the very definitionof a smile.
3, From the values thus calculated, the trader wouldthen have to solve the 2 x 2 system as above and determine (different) amounts of calls and puts
to match the boundary condition. It must be stressed again that if one believed in a particular model desc~ption of the d y n a ~ i c s of the underlying, then the second procedure would, in general, be theoretically incorrect: the assumed future prices would not be consistent with the und~rlyingfinancial model and todays market prices. In the light of the discussion about the likely realism of a description such as the one provided by Equatio~ (6.1) for equity or FX products (see Sections 6.4 and 6.5), however, feeling constrained by the model output should probably not be considered an imperative (or perhaps even desirable). It is clearly up to the trader how confident she feels about her own ~nancial intu~tion,or about the realism of a given n any case this example makes perfectly clear the link about the future re-hedging costs, the option value and the underlying assumptions about the smile type. Table 6.8 shows the portfolio replication option prices obtained with a constant (no smile) volatility (Const), a sticky smile (Fixed), and a floating smile (Float), for a constant or a time-dependent volatility. Despite the fact that the discussion of the specific pricing of continuous barrier options is not the main thrust of the argument, it is worthwhile pointing out that a floating smile can produce a signi~cantlyhigher option price than a sticky smile, which, in turn, produces higher option values than the constant (no-smile) volatility assumption. ~eturni~ to g the general pricing philosophy behind this example, one could object that, if the trader were to follow the second approach, she would, in a sense, assume to h o w the answer beforehand, since she would be imposing the future implied volatility surface, rather than obtaining it from the no-arbitrage dynamics of a model. Whilst there is certainly some truth in this criticism, it must also be kept in mind that, after all, the acceptance of a model ultimately depends on its ability
The double-barrier option prices calculated using the portfolio replication technique with a constant volatility (Const), a sticky smile (Fixed), a floating smile (Float) and flat or time-dependent volatility Flat volatility ATMVOL (%) 10 9 8 7 6 5 4 Const 0.0061 0.01 67 0.0411 0.091 4 0.1 834 0.331 2 0.5374 Fixed 0,0068 0.01 73 0.041 4 0.0911 0.1 820 0.3283 0.5329 Float Float 0.0074 0.01 78 0.041 6 0.0908 0.1 81 2 0.3269 0.53 15 Time-~ependentvolatility Fixed 0.0065 0.01 73 0.0420 0.0926 0.1 849 0.3327 0.5388 0.0069 0.0178 0.0424 0.0929 0.1 848 0.3323 0.5382
xtractio~ of the Future LocalVolatility
185
to reproduce qualitative features that the trader feels confident with. The at-leastapproximate reproduction of todays smile surface, it has been argued at length in this and previous chapters, is one such qualitative feature, and therefore the two approaches are not so different after all. Probably a more const~ctiveapproach is to recognise that different plainvanilla-option strategies will be needed to hedge different exotic products: if the hedging strategy were truly static in nature,12 as might be the case, for instance, for a single-look binary option, then todays option prices would be all-important. If substantial future re-hedging is implied by the strategy, as in the case of an option with a forward-setting strike discussed in Chapter 1, Section 1.7,then recovering future option prices becomes essential, and therefore the user might prefer not to place excessive reliance on the predictions of a model, andor to correct and supplement its prediction with exogenous (and, possibly, theoretically incompatible) information. It is also useful at this point to revisit the example ofan option with a forward-setting strike, analysed inthe case of purely time-dependent volatilities in Chapter 1, Section 1.7. We argued in that context that, in order to hedge the volatility exposure arising from the option, the trader had to lock in the future portion of the volatility between the time of the strike reset and the option ~aturity. We also pointed out how a plausible hedge would have to display, before the reset of the strike, an appreciable kappa, but no delta and no gamma. We proposed a long and a short position in wide, symmetric strangles to fulfil, at least a~proximately,these combined requirements. Finally, we noted that the chosen strangles ceased to produce a flat profile, needed to give no delta and no gamma, for smaller and smaller movements of the underlying from the original at-the-money position as the strike-resetting time approached. During the life of the option the strikes would therefore have to be adjusted; the closer to the time of the strike reset, the more frequently these ~eadjustments are likely to take place. In the presence of smiles the problem becomes much more complex, To begin with, one cannot assume that the trader will manage to effect these strike changes at no cost (i.e. the strategy will prove anything but self-~nancing); fu~hermore, these readjustments will become progressively more frequent as the strike reset approaches, and as the maturity of the options which must be bought and sold in order to readjust the strangles shortens. In other words, the shorter the maturities the higher the likelihood of re-hedging and the steeper the smile curve is likely to be. The trader will therefore have to take into account in her price-making the future smiley implied volatility of options with different degrees of in- and outof-the-monieness and different maturities. A s a consequence,fu~ure option prices (i.e. the fut~~e smiley implied volatility), might well turn out to have as great an
l 2 Note that, despite the name, the static portfolio replication strategy is not static at all, requiring as it does a very dynamic re-hedging if the barrier is breached.
impact as the spot option prices that determine the cost of putting the strategy in place today. In the light of this analysis, and of the discussion presented above regarding the pricing of continuous double barriers, would the reader prefer to rely on her financial knowledge and intuition, or on the output of a restricted-stochastic volatility model in order to estimate the nature (stickiness, time homogeneity, etc.) of the future prevailing smile? The situation is clearly very different in the case of interest rates on the one hand, and equities on the other. The type of smile present in the former e, its likely financial origin (see Chapter 4, Section 4.3, and Sections 6.4 and ) and the overall smaller magnitude of interest-rate smiles make the output of a model based on a price d y n a ~ c s of the type described in Equation (6.1) ~lausible and convincing, and create little incentive to override the model output. For equities and FX, on the other hand, explaining the smile on the basis of a restricted-stochastic-volatility modelling approach seems to imply a much less plausible future evolution for the sInile surface itself, and might therefore prompt the user to supplement or replace the model outputs with potentially inconsistent but financially more appealing information. A s a final caveat, one should r e ~ e m b e that, r much as it might be desirable, in some cases, to override the predictions of the model, one must still be careful not to intro i ~ ~ e ~ possibilities e ~ ~ e of ~ arbitrage, t such as the ones highlighted in model might ince we have pointed out that a restricted-stochastic-volati~ity account in a satisfactory way for smiles in the interest-rate area, we pursue this approach in the next chapter by proposing a model that, thanks to the closedform solutions that it affords, allows easy calibration to smiley option prices, and easy extensions to the Monte Carlo application for exotic options. For those asset classes where the underlying ~ n a n c i a l m e c ~ ~ n responsible ism for the smiles makes the rest~cted-stochastic-volatility frame~ork inappropriate, a completely different model~ingapproach is then introduced in Chapter 8.
FT
Let
Call($,, K , T ,t ) be the time-t price of a call on stock S with strike K and maturity T .
@(ST) be the probability density for the stock price t time t being S,. value a
ST at time T , contingent on its
Ignoring discounting, and given any terminal price dist~bution, we can write for the price time-t price of a call:
7
where (a b)+ = max(a b, 0). In order to evaluate the first derivative of the functio~ call with respect to the strike we can first re-write Equation (A.6.1) as
(ST K)@(ST)~ST
and then make use of the result
: I'
f (t, x)dt
= f ( t ,t ) +
iK
F
This gives
~ i f f e ~ e n t i aagain t i ~ ~with respect to K finally gives
where use has been made of the f u ~ ~ a m e n ttheorem al of calculus:
7.
The treatment presented thus far in Part Two has attempted to explain the dependence of the implied lack volatility not only on the option maturity -a feature that can be accounted for simply by requiring the instantaneous volatility to be time dependent-but also on the strike, by means of the mo~elling approach referred toin Chapter 4, Section 4.6, as the restricted-stochasticvolatilitymodel. In this framework the rate or equity price dist~butionis assumed to be generated by a process of the form
dX/X = pdt
+ g(X, T) dz(t)
(7. 1)
where X denotes the rate or price, as appropriate, and the other symbols have their usual meanings. If this choice is made, then ris~-neutral valuation (although lacks f o ~ u l a for ) calls and puts still applies. of this modelChapters S and 6 have illustrated two different imple~entations ling approach, the first based on the const~ction of a t ~ n o ~tree; a l the second centred around the preliminary evaluation of a local volatility surface to be used as an input to a distinct pricing engine. As we pointed out, despite appearances9 the di~erencesbetween the two approaches were rather s u p ~ r ~ c i aonce l : the model assu~ptionembodied in Equation (7.1) has been made, there exists, in equivalence between fact, a fundamental underlyin~
Part of the material in this chapter is based on a paper by Me~endereskiand Rebonato (1999) which has been sub~itted for publication. I would like to thank Dr Merfendereski for allowing the material to appear before it is in the public domain.
l. the specification of the local volatility surface a(S, t ) , for any S and t ; 2. the speci~cation of the plain-vanilla option prices CK,T(S, t ) , for all K and 1 " ; 3. the specification of the probability densities, @(ST), for the price, for all T
(at least as long as the drift is specified by a no-arbitrage requirement). ote that the qualifiers 'for all S and t ' , 'for all K and T" and 'for all 7" are essential to ensure the uni~uenessof the mappin om one description to unique local volatility the other, i.e. to ensure, for instance that there exis (ST);or to ensure that function o(S, t ) capable of generatin the integration of the product anof ST) times pay-off the to), of of a call or a put of strike K should produce the price today, CK,T(SO, the co~esponding plain-vanilla opti o r , finally, to ensure that, ifwevalued r a putby means of, say, a nte Carlo simulation for the underlying a call o the appropriate no-arbitrage drift and o(S, t ) as volatility, then all today's vanilla option prices would be recovered. In this light, the DK approach can be seen as one of the computational ap~roaches consistent with Equation (7.1), and which takes specification 2 above a start in^ point. The extraction of the local volatility surface S apter 6 would naturally lead to option valuation along the lines The approach resented in this chapter starts from the same ~nancial'picture' but follows yet another route, i.e. specification 3: option prices are arrived at by modelling directly the probability density generated by the diffusion (7.1). ~onceptually, the equivalence of, for instance, the tree-based and the densitybased approaches can be seen by noting that, since the underlying distribution is accessible from traded prices via the relatio~ship (see ~ppendix 6.1 of ~ h a p t e 6 r for a d e ~ v a t i o n ~
suitable trinomial trees locally one can, at least in principle, attempt to constr~ct nsistent with any set of (feasible2) ob e, Dupire (1993), Ru~instein (1994) o r ivalence, however, can be of little h tical implementations: it is well known, for instance, that a small noise in the market prices can give rise to wildly ~uctuati densities, and, as a consequence9 oothing and judicious fitting of to a highly unstable implied process for X. the underlying p~ce/time/strike surface is often advocated, but the fact remains that the resulting distributions can be very strongly dependent on the details of ne might wonder whether one should worry at all if the underlying distri~ u t i o nis not 'nice looking', as long as the prices of plain-vanilla calls and puts are correctly recovered. ~btaininga well-behaved and 'plausible' terminal
The observed option prices cannot be totally arbitrary, but must (i) be consistent with no-ar~itrage and (ii) give rise to a strictly positive density; see Dupire (1993).
rices
distribution, however, can be important as soon as one moves away from the evaluation of plain-vanilla options: all numerical methods, in fact, perform, in a direct or indirect way, an important sam~ling of one or more (joint) price di~trioptions that have been used in the ~onstruction, butions. For those plain-~a~illa via ~ ~ u a t i o n l), of the terminal distribution function, any procedure that goes (7, ack from the distribution itself to the prices is virtually guaranteed to produce ood prices; this is, however, no longer ~uaranteedto be the case even other plain-~anilla options with the same expiry but different strikes, and wou in general for exotic options. the densities obtained es '7.1-7.3: Figure 7.1 dis~lays S obtainab~e from the implied ~olatilities in Figure 7.3. The latter, in turn, were assumed to be given by the supe smooth ~uadratic smile and a smallrandom noise term, of than one-hundredth of the volatility itself. From this nois prices of plain-vanilla calls and puts were then obtaine the proba~ility-wei~hted pay-offs. These prices were th volatilit , and the latter compared with the correspond in^ input (noiseless) quanone can observe, despite the fact that the noise is extremely srnall (and, ,much smaller than a typical bid-offer spread), the errors in prices for different strikes and the same ~ a t ~ r i t i e obtained s by numerical inte see Figure 7.2(a) and the densities in Figure '7.1 can still be considera~le; Clearly, if one were to evaluate more complex options the errors pro~ably W be even more severe. circu~~vent this ~roblema new approach ) ) is proposed in this section, by means
0.6
0.7
0.8
0.9
1.1
1.2
1.3
1.4
ThedensitiesobtainablefromEquation implied volatilities in Figure7.3
(7.1) using the prices from the
1.4
1.2
1
I l
U)
0.8
0.6
0.4
0.2 0
-0.2
-0.4
(a) Pricing errors in units of kappas (PV change for a 1 change in volatility) from the densities and volatilities shown in Figures7.1 and 7.3. See the text for a precise description of the procedure. (b) As (a) for a similar noise in the volatility input
dist~b~tion of s the quantity X , which can be interpreted either as a spot or a forward process, is assumed to be generated by the by-now-fa~iliarprocess descri~ed by ~ ~ u a t i o (7.1) n above, It is these terminal distributio~s,however, rather than the process itself, that are directly ~odel1ed,and this is done by a s s ~ ~ that i n they ~ have a particular four-parameter functional form (described the lognormal distrib~tionas a special case. The later on) which enco~passes idea behind the approach is to produce t e r ~ i n a ldensities which are aswell
rices
19.00%
0.6
-1
0.7 1.1 0.81
0.9
1.2
1.3
1.4
The smooth and noisy implied volatility curves used to obtain the densities in Figure 7.1
behaved as the chosen functional form will allow, and, at the same time, sufficiently flexible to account convincin~ly for observed prices. The approach will be shown to be of simple practical il~plementation because, for each expiry, the best combination of parameters that gives rise to an optimal (in a sense to be described) match tomarket call and putprices can be found using a very efficient and rapid mixed nume~ca~analytical procedure: this is because closed-form solutions are presented below not onlyfor call and p consistent with this distribution, but also for the cumulative distribution aris from the chosen density. Thanks to these analytic solutions, the needed to calibrate the model to ~ a r ~prices e t can be rendered ofnonThe tec~nique can also easily lend itself to onte Carlo si~ulations plain-vanilla options, From fitted distributions co~espondingto matu~ties it is in fact possibi obtain the equivalent diffusions (see b the proces~ ('7.1). It would also be possible to cast the whole approach in terms of forward (rather tha cesses. This would be p~ticularly appropriate in order to simulate via lo approach long-dated, path-dependent options. This for war^-rate-based impiementation i s relatively easy and practical because the co~elation st~cture of the joint st~butionof the forward rates obtai a 'traditional, log-normal nte Carlo evolution i s little p e ~ u r b e by replacing the log-nor~al evolution with the more general process o~tainable from the fitted densities; fu~hermore, it is also possible to determine a quantitative estimate of the deviation from the desired and the obtained co~elation
surfaces. The well-established techniq~es available to simulate jointly lo variables can therefore be easily adapted to the case at hand. ce these additional results are in themselves quite complex, and connected with the approach presented here, they are not p further. It is, however, wo~hwhilepoint out at this stage that form-solution approac~ presented here for a coherent pricing system encompassing p he approach presented in this chapter is o f thistype. ~ o n e t h e ~ e it s s enjoys several thatmake its detailed analysis wo~hwhile: tion function is directly mode~led,the result behaved and plausible shape. ied volatilities (input prices) ca to radically different dist it conversely follows that an approach starting directly from the distribution can ariety of market prices with little loss of pr~cision.sing what follows one can often argue that, when the fit i rtunities might exist, From this point of view the approach not only efficient fitting to existing prices, but can also provide in~ications
A further positive feature is thatwe express the closed-form solutions for eir de~vatives in terms of the integrals of a f a ~ i of l~ functions -normal dist~butions belong as a special case; in addition, the functional form of these closed- for^ solutions is such that they have an easily for k-like appearance, making their use easy ai~htforward stomed to pricing using the arke et-stand ck formula. This ~ it ~appear, ~ ~ since, ~ after c i e c o m ~ e nis t actually more relevant than it ~ ~~ h the existence of a smiley v tility quote simply stems from a f a ~ l i of ~ t ~ a r ~ e t ~ ~ ~ iwith cip the a n t ack s conceptual and computationa~ framewor~s. the disc~ssion in Chapter 4, Section 4.1.) i n a ~ l yas ~ far as E~ropeanoptions are concerned, the stoch~stic nature of st rates can be automatically inco~orated in the treat~ent by assigning the o ~ (see ~ the ~ d ~i s ~ ~ s s i o in n apte er I). This volatility volatility of t h e ~ process can either be obtained directly from the market, or obtained using the X spot volatility and the volatility of the discounting factor (which, in turn, can be derived from the volatilities of traded instruments such as caplets). The same is not true for approaches based on the evolution of the spot process, where a whole new computational di~ension must be added in order to account for nondeterministic rates. Clearly, the longer the e ~ ~ i of r yan option, the a1 impact of the stochastic discounting. a~proach proposed in what follows is, of course, not without its shortcornut these are shared by all the spot-based mode analysed in the previous S, In particul~,when imputin~ an underlyin distribution from traded (7.2) or using the procedure desc~ibed below) prices (either directly via Equ~tion
one is i~plicitly makin rownian ~ltration assumption, i.e. one i s implicitly a s s u ~ i n gthat all the ncertaintyaboutthe future price evolution is pro ccessive realisations of the price itself, and of no other considerations about market completeness~andabout the validity of usi t diff~sion when markets are c o ~ p l e t e but the mixed jum~-diffusion process, have ection 4.6, and they are therefore not owever, find it useful to revisit that section at this point in order limitations of the approach w These s h o r t c o ~ n are ~s o the rnethodology ~roposed in and are shared by all the rest~cted-stochasti~-volatility a~proache no~ethelessbe kept in mind when actually hedgi~gand risk-rna positions.
The set of market call and put option prices available on a ~ s k - ~ e u t rprobability al density that can be estimated using a approac~es, In the presence of smiles, this mar~et-imp lie^ distributio~ i s different from the l o ~ - n o r ~ a l .a continuum of call (put) prices were available, one could I f make use of the followin relations (see A p p e n ~ 6.1, i ~ ~ h a p t e 6) r to e s t i ~ ~ athe te risk-neutral density function ~ ~from (the call ~ price ) function C( (l993)):
Since, in practice, call (put) prices areonly available for discrete strikes, one has to a ~ p r o ~ i ~the a t second e derivative in Equation (7.4) by a finite difference ~ppro~imatio~. The a~proach that is, in princi~le, least data-polluti is to make use only of the actually quoted prices in order to estimate this sec0 derivative. ~nfortunately, res 7.1-7.3 and explained in the a c c o ~ p a n ~ discussion, in~ the for the theoretically desirable totally n o n - ~ a r a ~ e t r i c f e a t ~ of the procedure is that the result in^ density function is often hi sometimes multi-~odaland, in general, not even guaranteed to be everywhere positive. A si~ple and ~ ~ l l - k n improve~ent o~n (see, for example, to model the price/ ike elations ship by means of a suitably smooth ~L~nction. i ~ Whilst the results far brief outline of the ocedure is reported in ~ p p e n d 7.1. the density so obtained can be better behaved, this i s generally only the case when
the fitted stri~e/maturity/volatilitysurface is described by a very smooth function, uch a smooth function ~ i g not ~ t account for plain-vanilla option prices in a very satisfactory way. The obvious remedy is to increase the ice-fitting function, typically by allowing it to have more de the strike/~aturity/volati~ity surface is assumed to be given, c spline through the observed liquid market prices, as more knots in the pline were retained, re and more prices (in the knit, all the prices) can be atisfactoril~ recovered. increasing the number of knots, however, wildly ensities can easily be re-introduced. It is therefore impo~ant to note e approach is still non-parame~ic, information has nonetheless been om the actual prices to the model surface. It is therefore not a priori obvious that the infor~ational content lost in going from the actual prices to the m o ~ e surface l is ade~uately co~pensate for ~ by not having to make assu~ptions about the underlying distribution. In order to obviate the above s h o r t c o ~ ~roposea new implementation of the p from the market prices of calls and puts, and atte function that best reprod~cesthe market pri we can now rest assured that the resulting density function choice of functional form for the density functi unknown) density function in terms of m ~ a r a ~ e t e rwe s , therefore cast oblem in terms of an dimensional optimisation: we seek to ~ i n i m i s e x2 defined as:
where the weights wci and wpi may be chosen to specify li~uidity or other quality
Jo
a ;
s ~ ~ n e u t rdensity al function. The notatio~s emphasise the model call and put price depe~dence on the parameters of the distribution, 8 .
functional dependence on the parameters, are specified. Twodesiderata however, already follow from looking at the problem in this manner: on the one hand, the density function should be s ~ o o t h and well behaved; on the other hand, a sufficient richness and fle~ibilityto account for a widevariety of observe t prices should be retained. Analytic tractability is an added bonus. h these criteria in mind we have chosen as the fitting density functio~ a ararneter family of probability distributions the Generalised 2 inwhat follows) (~umminset al. (1990)). The choice is motivated by the fact that
2 density, besides being flexible, can easily accom~odatefat tails (which result in the cornmonly observed implied volatility smiles observed in the market); when using certain c o ~ ~ i n a t i o n s it de~enerates into a log-normal dist~bution of the four parameters, thus allowing a convenient, consistent, and smooth transition to flat volatility models; and it affords analytic solutions that make the approach both practical and cient.
where a , b, p, and q are the four free parameters, and B(., defined in terms of the g a m a function as
a )
is the beta function
For future use, the moments of the G 2 distribution can be derived and ex~ressed
as follows:
~olatilitys ~ i l e s can be i n t e ~ r e t e as ~ deviations from the lo~-normal(L what follows) density for the underlying ~sk-neutral density, a qualitative feel for the shape of the distribution, andto ed compa~sonwith the log-normal distribution, a simple ical tool for observing this deviation and which enables c o m ~ a ~ s o across n different maturities and indeed across different markets is now described. Using the following transformation from variable x to variable y , y=lnx a
, i.e. x = epY+a
(7,10)
,p~fx),is transfor~ed into a standardnormal density, p y(yj=

n j a ~ and i ~Cornel1 (1970)):
(7.13) can be a p ~ ~ i to e dany fitted risk-neutr~1 de~sity, and the This same transfor~ation deviation of the r e s u l t i ~ plot ~ from ~ o r ~ a ~ can i t yive a direct visual indication of such features as skewness o r kurtosis, In ~articular, for the G one o~tains
IPe(Br+a)lSoc;sz(eP'i",
In thischapter,we
focus on tb plottedandcomparedwith a st
a, b, p, 41,
2 distri~~tion and th d normalplot,as sh
-3
-1
Standardised L N and matching GB2 fits for difTerent maturityoptions
the set of parameters obtained by fitting the density function to a series of caplet prices of different maturities ranging from 6 months to 10 years (see Section 7, for nu~erical details). The cluster of nor~alised fitted GB2 curves displayed in Figure 7.4 shows that, despite the difference in maturities, the different smiley distributions, once 'nornalised' ,display a strong ~ualitative si~ila~ty.
In this section we derive closed- or^ expressions for call and put prices, both in terms of the most general G 2 ~istribution,andin the case where certain are imposed. restrictions on the values of the para~eters As mentioned above, one of the desirable features the approach presented here is that the pricing formulae can be expressed in lack-like form. In order ht these similarities, we begin by recalling that the first two ~ o m e n t s of a LN distri~ution,
defined in terns of the parameters a and 6, are given by
If, for si~plicity, we set interest rates to zero, the call and put option prices are
(7.17)
The call and put price integrals can then be solved to yield the price in terms of ,and the two LN para~eters
( * ) , as usual, denotes the cumulative normal dist~bution, E~uations (7.18) where N not have the usual lack appearance. However, one can i ~ ~ e d i a t e verify ly t a = ln(F) o"/2
p=04
where F is theforward price, t is thetimeto expiry, and CT is the volatility. ~ubstitutinginto (7.18), the familiar form of the lack call and put option pricing formulae is obtained:
(7.20)
emarkably similar results can be obtained for the call andputusing the CB2 ~ i s t ~ b u t i oAgain, n. the call and put option prices can be defined by the following integrals:
00
CallGB2(K) = PUtGB2(~) =
~*
00
( xK)+$'%B2 ( x )dX
x)+$?GBZ(x)
(7.21)
whichcan be solved in absolutegeneralitytoyieldthe call andput prices in terms of the strike, K , and the four GI32 distribution ~arameters:
where now
1 F2(a, b;c;x)
is the hyperg~o~etric functio~, defined as
and the other sym~o1s have their usual meaning. The results obtained so far can be rendered more transparent and for zero interest rates) by not in^ that ( a ~ a i n
easy to use
(7.24)
This being the case, we can then substitute b = F A where

(7.25) N p I /a,4 I/4 to obtain the following call and put price equations in terns of the strike K, forward F , and thee of the four GB2 distribution parameters, a, p, and 4:
A.=
, 4)
-F
[ ( K / A F ) I ~ ( ~ + ~ ~I/a, ~)p ~F~ 4;[1~ p 4 W P W ,4 W )
(7.27) Note that one of the parameters has been eliminated by imposing the no-arbitrage forward ~ondition.~ Although these expressions appear somewhat more unwieldy than (7.22) and (7.23), they are already beginning to be in the familiar Black form, involving as they do a weighted sum of F and K. ~nfortunately, unlike the LN case, we cannot directly e l i ~ n a t a e second distribution parameter by introducing the volatility, Q. One can, however, deternine an 'equivalent' volatility of the new distribution fitby imposing that not only the first, but also the second moment of the LN and GB distri~utions should be identical, i.e. that ,a+P2/2 = bB(P l/', 4 'la)
+ + + l/a; (
e2a+2P2
Solving for a and , 8then gives
= bzB(p + 2/a7 4' l a ) N P ,4 )

H p
4)
(7.28)
+ 2 / a ,4 -2 / 4
5 ( P ?4)
It is interesting to note that, despite appearances, this manipulation is exactly analogous to the imposition of the second (forward-rate) condition in the DIC tree construction.
e that in the expressions above the second m o m e ~ of t the L sen to be derived from the at-the-money volatility, It should be stressed that there is no a priori reason why the moments of the two distributions should be matched, and that, by so doing, one is restricting the ~exibilityof the CB2 distribution. Whether the procedure thusmodified still retains sufficient richness after this reduction in the n ~ ~ b of er degrees of dom is an empirical issue that will be addressed later on, introd~ctionof the e~uivalent ~olatility is sim~ly motivate ed with if the user w a ~ t ~ d flexibility of the G ne the e~uivalent volatility, o q u i v , in terms a, p and q. Note that since this e~uivalent (e.g. the at-the-~oney imp lie^ vo~atility), 2 distribution ~arameters, i.e. p and q. It do indeed allow sufficient ~ e ~ i ~ i lto ity model most realistic forms of the observed market smiles, skews, frowns, etc. in an efficient ~ a n n e r . Armed with this result, we are now in a position to simplify further the a~alytic expressio~ for the G 2 call and put price formulae ( 7 . ~ and ~ ) (7 27) by r the expressions in terns of the incomplete beta function, I z ( p ,q) ( ~ b r a m o ~ un (1972)). The latter is defined as follows:
and, as before, 1F2(a, b;c;x) is the hypergeometric function. function ( A b r a ~ o w i and t~ ~ t e ~ the following relation for the hyper~eomet~c (1964, Eq, 15.3.4)) (7.33)
i
and ~ e r f o r ~ i n some g algebraic acrobatics, we finally obtain the following simplified expressions for the GB2 call and put prices:
where z is defined as
2=
+(hF/K)a
(wK)a
(7.35)
~ o l ~ t i for o~ Smiley s
This simplified form of the C 2 call and put price analytic e~pressions(7.34) is nowreallyvery s i ~ l a to r the familiar Black(and Scholes) optionpricing formulae re-derived earlier (see ~ ~ u a t i o (7.20)), n and, like the Black and Scholes expression, allows simple, robust, and efficient evaluation of vanilla options. In place of the curn~lativenormal, AT(.), we now have the incom~letebeta ion, .lx(p,q), which, like N(.), also takes on values between 0 and 1. AnalS to N ( * )tabulati~ns, standardised plots of Z J p , q ) exist for different (p, q) xample, Figure 7.5). Numerical appro~imation of Zz(p,q ) is also and efficient numerical routines are available in the l i t ~ r a t ~ r e .(1992, Section 6.4)). of the call and put price formulae are shown in Figure 7.6(a) and put __ call parity. 'degenerates9 into a LN distribution f the four parameters a, 6, p, and q. For those all and put price f o r ~ u l a e (7.34) become i~entical to the Black rrnulae, i.e. the Black pricing formulae are a S trikes' case of the more general GB2 formulae by enforcing the relations ('7.24) and (7.30), one can first ensu second rnoments for the GB2 and co~espondingLN dist~but two free param~ters.As p and q become 'large'4 the resultant CB2 distri~ution
0.6
0.2
0.4
0.8
Figure 7.5
/ , ( p , 9) versus z for different ( p , 9) pairs
A valuefor p and q ofapproximately 80 issufficiently'large'tomaketheresultsvirtually indistin~uishable from those obtained using a log-nor~al distribution. Much larger values can give rise to n u ~ e r i cproblems. ~l
4000 2000
6000 100008000
(a)
2000
4000
6000
8000
10000
(b)
parity
(a) GB2 call;
(b) put prices;
(c) the sum of (a)and
(b), showing call-put
unifor~y converges to a LN,thereby ensuring that our option pricing formulae approach the Black formulae. As far as the price of calls or a put in the presence of smiles is concerned, empirical rules of thumb commonly used in the market can provide an adequate, if approximate, solution to the problem of providing a quote for maturities or strikes intermediate between the liquid ones. Where these rules of thumb break down, however, is in the specification of the risk statistics. In particular, for cases like the normal distribution it is easy to show that the gamma or kappa (Vega)
4000 1
3000 2000 :
1000 :
-1 000
-2000 : -3000 -4000
r
'
=L , p
ure 7.6
continue^)
statistics obtainable from inputting in the Black formula the smiley out-of-the money volatilities significantly differ from the corresponding true normal value; see Chapter4, Sections 4.2 and4.7. It is therefore veryuseful that the G dist~butionallows the closed-form evaluation of both the delta and am ma for calls and puts. Explicit di~erentiation in fact gives (7.36)
APUt
(+
" 9
4 -"Iz(4-;,P+-p
(7.37)
(7.38) Note that (7.38) is of a similar functional form to the GB2 density itself. Comparing the above G 2 'greeks' with the Black deltas and g a m a s , (7.39)
(7.40) (7.41)
10000 4000 8000
6000
(a)
0.0005
O"0004
0.0003
0.0002 0.0001
w e 7.7
(a) GB2 call delta Aca//;(b) GB2 call and put gamma l?ca,/ = rPul
one can readily notice f u n d a ~ e n t a l s i ~ l ~ i t which i e s , can allow traders to preserve their Black-like intuition (see Figure 7.7). Again, note that the expression for amm ma, (7.4 l), is of a similar funct~onal form as the base density.
In.this section the CB2 function is fitted so as to reproduce the plain-vanilla option prices from a variety of markets. It is important to point out that the dis~ussion presented at the end of Chapter 6 indicates that the smile surface produced by a restricted-stochastic-vo~atility diffusion such as the one in Equation (7.1) tends to
give rise to reasonably time-homogeneous~ ~ smile t surfaces ~ ~ only if e t o ~ 'S~ y smile surface is of the interest-rate type. Also, the discussion about the financial ~ e c h a n i s ~ possibly s responsible for the smiles (Smile Tale l in Chapter 4) gave an indication of the reason why a dependence on the underlying in the volatility term could give a plausible description of the origin of the smile effect in the case of interest-rate options, Therefore we believe that this is the area where most pro~table use can be made of the model described in this chapter. Unlike other researchers who start from identical assu~ptions(i.e. Equation (7. l))I do not r e c o ~ e n d the approach described in this chapter to price options in the FX or equity cases. Even for these asset classes, however, the ~ e t h o d o l o ~ y described above can provide considerable help if one had to ark-to-market plainvanilla option positions, perhaps put on the books of an investment house a lon time ago, when the strikes are very far away from any quote currently obtainable through brokers in the markets. In this more restricted light (i.e. as a provider of a 'snapshot' picture, rather than as an explanation of the true price dynamics) we show below that the GB2 approach would give rise to a useful and flexible tool. Needless to say, the most interesting application remains in the option pricing arena. The empirical analysis was carried out by looking at all the caplet prices implied from quoted broker cap prices for different strikes and maturities on l ~ (8 July 1998) for DE several days. The results obtained for a p ~ t i c u date are reported in what follows as representative of all the days investigated, The equation used for the forward price5 of a caplet is as derived in Section 7.3:
Call = FI,(q l / a ,p
+ l/a) -~ I ~ (p) q ,
(7.42)
where Iz is the incomplete eta function, z is as in Equation (7,35), h is given by
and where B ( p , q), as above, is the beta function. After discounting, the above pricing equation was used to value caplets with different strikes, and to find the optimal values of a, p, and q, that minimise the sum of the square errors between the cap prices obtainable using the above equation and the prices observed in the market for different ~aturities. As many sets of distinct distribution parameters were used in the search as the number of cap maturities, and the assumption was made that all caplets between two successive option maturities would be priced by the same set of parameters. The results are reported inTables 7.1 and 7.2 whichshow the market and fitted caplet prices, respectively. The excellent quality of the fit is apparent. In
Since we have made the equivalent diffusion assumption, andwe work in the forward measure, the present value is simply obtained by ~ultiplying the forward value by the discount bond ~aturing at caplet-pay~ent time.
C C
C
2
C
C
c
C C C C
c c
C
c c
C
c
C
c e c
C
c
c
particular, it is worthwhile pointing out that all caplets, except those whose corresponding cells in the table are highlighted in bold, were fitted to within one basis point. The parameters that lead to the best fitare shown in Table 7.3 and in Figure 7.8. As can be noted, there is a reasonably smoothvariation of these ~arameters (especially a and p) across ~aturities.This would be of great help if one had to quote a price in a caplet of expiry intermediate between two optimisation dates. As explained above, by equating the first and second moments of the lognormal and GB2 dist~butions we can derive an e~uivalent volatility which can be compared with the caplfloor implied volatility curve. The results are shown in Table 7.4, A s a further test the approach described in this chapter was used to fit the prices of calls on the FTSE in the s u m ~ e of r 1998.6The market data comprised the Forward Index value and the implied volatility for different strike prices and maturities which range from three months to two years. The price data for the FTSEl 00 are shown below in Table 7.5, and Tab1 7.6 shows the market option prices for these data.
Table 7.3
1.o 1.5 2.o 2.5 3.O 3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0 7.5 8.0 8.5 9.0 9.5 Best-fitCB2 para~eters
a
Maturity (y)
b
0.04960 0.06232 0.07355 0.08371 0.091 66 0.09959 0.09763 0.1 01 55 0.1 021 6 0.10186 0.1 0358 0.1 02 65 0.1 0293 0.1 01 01 0.0991 0.1 0224 0.101 17 0.1 031 6
P
2.923 3.666 3.21 6 2A92 2.61 4 2.479 2.345 2.338 2.33 2.1 32 2 .Q78 1.907 1.854 1.706 1.698 1.692 1.687 1.683
4
5.756 10.01 10.85 11.22 11.49 11.51 10.64 10.63 10.63 9,351 9.423 8.178 8.029 6.901 6.893 6.887 6.883 6,879
5.618 3.436 2.889 2.651 2.489 2.438 2.474 2.457 2,459 2 .S40 2.503 2.587 2.580 2.684 2.658 2.636 2.620 2.608
As mentioned at the beginning of this section, the ~ ~ u ~ c suitability iuZ of the approach to the of equity case is not addressed in carrying out this test, which is only meant to show the ~exibility the approach. ~l~ernatively, the approach employed could beof use as an i n t e ~ o l a ~ i o ~ e x ~ a p o l a exercise for ~ a r k i n g - t o - ~ purposes. ~~et
" " " " " " " " " " " " " " "
" " " " " " " "
" " " " " " " " " " " " " " " " " " " " " * " " " " " " "
9 10 11 12 13 14 15 16
17 18
ure 7.8
Plot of best-fitGB2 parameters Implied and equivalent volatilities
Maturity (y)
Log volatility Equivalent Equivalent volatility (%) volatility (%)

13.48 14.99 16.06 16.26 16.46 1 5.89 15.15 14.39 13.64 13.30 1 3.05 12.79 12.54 12.29 1 2.05 11.79 11.55 11.30 13.52 14.94 1 5 *93 16.07 16.21 1 5.63 14.89 14.1 4 13.41 13.06 12.80 12.55 12.29 12.05 11.80 11.55 11.31 11.06
log volatility (%)

0.04 -0.06 -0.1 3 -0.1 9 -0.25 -0.27 -0.26 -0.25 -0.24 -0.24 -0.25 -0.25 -0.25 -0.24 -0.24 -0.24 -0.24 -0.23
1 '0 1.5 2.0 2.5 3-0 3*S 4.0 4.5 5.0 5.5 6.0 6.5 7.0 7.5 8.0 8.5 9.0 9.5
The volatilities given for the different contracts were converted to premiums lack model, and these prices were then used to determine the G distrib~tio~ that best fits these prices. The model prices thus obtained are given in Table 7.7. The agree~ent can. again be observed to be excellent. The p~ameters for each maturity were found to be as given in Table 7.8.
le 7.5
The FTSE implied volatilities used as input data (29 May 1998)
Spot = 5843.32 (29 May 1998) Strike Maturity Forward Disc. Sep-98 Dec-98 Mar-99 Jun-99 Sep-99 Dec-99 Mar-00 Jun-00 5915.50 6000.1 l 6079.46 6128.55 6195.65 6269.07 6341.98 6383.58 0.98 0.96 0.95 0.93 0.91 0.90 0.88 0.87 5700 0.2530 0.2522 0.2528 0.2535 0.2539 0.2543 0.2546 0.2550 6050 6000 5950 5900 5850 5800 5750 0.2493 0.2495 0.2505 0.25 15 0.2519 0,2524 0.2528 0.2533 0.2456 0.2469 0.2482 0.2494 0.2499 0.2505 0.2510 0.25 l S 0.2420 0.2442 0.2458 0.2472 0.2479 0.2485 0.2491 0.2498 0.2385 0.2415 0.2434 0.2450 0.2458 0.2465 0.2473 0.2480 0.235 1 0.23 18 0.2387 0,2360 0.2409 0.2385 0.2427 0.2405 0.2436 0.2415 0.2445 0.2425 0.2454 0.2435 0.2463 0.2445 0.2286 0.2334 0.2361 0.2384 0.2395 0.2406 0.2417 0.24'28
The market option prices for the data in Table 7.5
Strike Maturity Sep-98 Dec-98 Mar-99 Jun-99 Sep-99 Dec-99 Mar-00 Jun-00 5700 418.81 586.56 727.12 837.32 950.32 1061.2 1167.8 1250.2 5750 5850 5800 385.79 553.83 694.3I 804.79 917.58 1028.2 1 134.7 1217.3 354.00 323.51 52 l.93 490.86 662.1 5 630.57 772.79 741.21 885.27 853.33 963.43 995.70 1101.9 1069.4 1184.7 1152.3 5900 5950 6000 6050
. . . . . " _ I _ _ " .
294.41 266.76 240.62 43 1.28 402.89 460.62 539.56 569.19 599.56 649.34 679.35 710.02 790.54 759.96 821.71 868.85 93 3.46 89990 1005.4 974.12 1037.2 1057.4 1088.6 1 120.3
216.06 375.54 5 10.77 620.15 730.09 838.46 943.36 3 026.7
able 7.7
The model option pricesfor the set of data in Table 7.6
050 6000 5950 5900 5850 5800 Maturity 5750 5700 294.62 323.69 385.71 354.07 Sep-98 418.5 460.51 490.73 Dec-98 586.74 553.87 521.86 599.37 630.34 Mar-99 727.42 694.38 662.01 709.84 740.92 837.66 804.85 772.601 Jun-99 Sep-99 950.59 917.61 885.1l 853.10 821 .S8 93 l .39 995.55 963.25 Dec-99 1061.4 1028.2 1037.2 1069.3 Mar-00 1167.9 1134.6 1101.8 Ju~-00 1250.2 1217.2 l 184.5 I 152.2 1120.3 266.91 43 1.24 569.13 679.35 790.58 899.98 1005.6 1088.8 240.62 215.77 402.94 375.64 539.65 5 10.94 649.49 620.26 760.10 730.14 869.02 838.51 974.33 943.46 1057.6 1026.8
Sq. error
0.258 0.080 0.240 0.3 l3 0.196 0.130 0.111 0.101
The best parameters for theto fit the FTSE data in Tables 7.5-7.7 Maturity Sep-98 Dec-98 Mar-99 Jun-99 Sep-99 Dec-99 Mar-00 j un-00
a
P
0.51 1387 0.248835 0.09594 0.076328 0.0663 73 0.058059 0.051 503 0.048309
4
4.1 441 75 0.81 6654 0.255468 0.231 991 0.262421 0.305228 0.361 9 0.37601
16.6881 9 22.781 68 46.38553 48.61 783 48.28537 49.1 467 50.33071 49.34221
In this chapter a parametric approach to modelling p~ain-vanillaoption prices in the presence of smiles has been presented. A GB2 distribution, defined as
was fitted by ~ n i ~ i s i n g
Analytic call and put price formulae were derived for the case of zero-interest rates which canbe expressed in such a way as to display a Black-like appearance:
The delta and gamma statistics were also expressed in terms of simple closedalso turned out to bear a strong forma^ resemblance to the form e ~ ~ a t i o nwhich s, co~espondi~ Black-like g formulae. Putlcall parity is preserved, as well as the identity for the gamma of a call and a put. These features would therefore allow the user not only to price but also to hedge vanilla options. By virtue of these closed-form solutions the minimisation needed to determine the para~etersof the GB2 distribution can be accomplished inan extremely efficient manner. Also, the optimised para~eters display a smooth behaviour as a function of option maturity, thereby allowing pricing by means of parameters
obtained from a simple interpolation of the co~espondingvalues for nearby ~ a t u ~ t i eAs s . the parameters p and q become large the dist~bution unifo~llly converges to the log-normal distribution, which can therefore be considered a special case of the more general approach presented here. The limits of applicability of the approach are to be found in the eequivalentdiffusion, assu~ptions.These limitations are, however, shared by all those approaches (such as the Dupire or the DK) which, i~plicitly or explicitly, start from E~uation (7.1) . Finally, it is impo~ant to point out that, once the distribution p~ameters have s , procedure presented in Chapter 6 been obtained for a series of m a t ~ ~ t i ethe can be easily employed to obtain explicitly the dependent ~olatility coefficient in E~uation (7.1). This task is greatly facilitated by the smooth ch~nge in the p~ameters across maturities and strikes. Once this volatility coefficie~t has been extracted, consistent Monte Carlo simulations or finite difference schemes can be carried out in order to price exotic options.
An i~plementation of Shirnkos (1992) approach can be effectedalong the following lines: carry out first a cubic-spline fit through the n liquid market option prices for a range of
strikes with a given expiry, T.For the chosen set of basis functions, the second derivative of the cubic-splin~ fit is a linear function of strikes, K i , thus producin {Ki,( o T ( k l j ) } density function defined by n 2 pairs of strikes and density ordinates, for i = 0, l, ..., n, n 1. More precisely, if one denotes by P(A) the pro~ability of occurrence of event A, the various densities, and the values of the strikes for which the densities can be taken to be zero, can be obtained by means of the following relationships:
pi
=~
( ~ i < - 1
K <K i ) = i [ ( o T ( ~ i - I )
(oT(Ki)](Ki
Kj-1)
for i = 1, 2,
I)
+.
12
(A.7.1)
where
Pn+l = P(K >K E ) = l
(A.7.4)
i=l
A s mentioned in the previous chapters, a variety of financial mechanisms have been proposed in order to account for the observed market smiles. More often than not, analytic tractability, rather than empirical and financial reasonableness, has informed the choice of modelling approach. There are important practical and theoretical reasons why many of the mathematically appealing models that have been proposed fail to provide a satisfactory theory of smiles in the case of equities; to refresh our memory, they are presented below.
The qualitative shape of the equity smile has been relatively constant over many years (certainly after the 1987 crash); approaches such the ones examined in Chapters 5-7 can fit (by construction) todays shape of the smile surface, but, in the equity and FX cases, tend to produce a very different ~ t smile ~ surface. ~ Therefore, e to be consistent and to price options reasonably, the user must believe that the shape of the smile we observe today, and that has been with us for so long, is destined to disappear in the future. The equity smile is pronounced for short maturities and shallower and shallower for longer maturities; fully stochastic volatility models tend to produce the reverse effect, or have to be parameterised using strongly time-dependent coefficients which, in turn, imply a time-inhomogeneous future behaviour. (See, for example, Dasand S u n d a r a ~ (1997) for an empirical study.) In other words,whenweuse fully-stochastic-volatility models thus parameterised we are again implicitly subscribing to the view that the future will be qualitative^^ different from the present and the recent past.
Useful discussions with Ernanuele Arnerio are gratefully acknowledged,
3. In the equity case the smile is by and large floating (see the discussion in Chapter 4, Section 4.4); it is therefore reasonable to impose the condition that, whatever model is used, it should generate a future smile surface which approximately maintains its shape as a function of the future level of at-themonieness (see also the discussion in Chapter 6, Section 6.8). It has been pointed out in Chapters 4-7 that those modelling approaches that, within the restricted-stochastic-volatilityframewor~, successfully recover todays prices for plain-vanilla equities or FX options seem to fail to reproduce a selfsimilar future implied volatility surface. In this chapter a completely different pricing approach is therefore introduced which attempts to account qualitatively for all the empirical observations highlighted above, incl~ding the persistence of the floating nature of the smile. In pa~icular,the mixed jump-diffusion model presented below naturally produces floating smiles of an approxi~ately time-homogeneous shape which are more pronounced for short ~aturitiesand shallower for longer expiries. Despite the appeal of the methodology, the analytical, numerical, and, not least, financial difficulties of the approach are not to be underestimated. The results shown below therefore should be seen as a pointer towards what I believe are the most exciting current developments, rather than constituting the last word on the matter.
The financial model presented in this section formalises and makes more precise theSmileTale 2 presentedin Chapter 4, Section 4.4. We assume that there exists a first set of econornic agents (pension funds, institutio~al investors, m ~ t u a l funds, etc., referred to in what follows as the fund players) who are natural holders of equities. One of the main sources of risk to which they are exposed is the possibility of large downward jumps in the equity indices in which they are (or, have to be) invested. For this reason they are natural buyers of insurance, typically in the form of out-of-the-money puts, and, by so doing, they bid up the prices of these options. There is then another set of economic agents (proprietary traders, investment banks, etc., collectively referred to in what follows as the arbitrageurs) who have no preferential habitat in terms of option strikes. If the dynamics of the underlying were exactly diffusive, and the trading universe exactly of the type,then these players wouldnaturally exert a pulltowards re-equilibration of the option prices by selling the expensive out-of-the-money options, and Vega-hedging themselves by buying cheap at-the-money options.
Note that, ina Black world, the ratio amount of at-the-money option neededto neutralise the trader against the Vega exposure automatically provides the correct gamma hedge as well. The relevance of this point is dwelt upon later on.
217
If the process of the underlying index were truly a log-normal Brownian diffusion, then the arbitrageurs would reap a risk-less profit at the expense of the fund players, and therefore the smiley shape of the strike/implied volatility function could not persist in equilib~um.In the real world, however, there exist rare but large jumps. Over these rare events the continuous-tirne re-hedging that is necessary for the arbitrageurs to capture the 'extra value' between the at-the-money and the out-of-the-money volatilities can no longer take place. By itself, this fact will create a variance of returns for the arbitrageurs, andif the latter are risk averse, and perceive the variance of their portfolio returns as their 'risk', they will demand an extra compensation for a s s u ~ n g the jump risk: in other words, until the smile is steep enough, the arbitrageurs will not move into action. It is also important to note that, given his strategy, over a jump the arbitrageur will always make a loss. To see why this must be the case let us consider his strategy in greater detail, The arbitrageur observes a smile curve which is flat for values of the strike above the at-the-money level, and which rises monotonically for strikes below the at-the-rnoney level. He therefore enters the following transactions (see Table 8. l and Figure 8.1) : he sells an out-of-the-money put; he constructs a zero-cost delta-neutral replicating portfolio to cancel the delta exposure of the out-of-the money put; he neutralises the Vega exposure of the out-of-the-mo~eyput by buying an amount of at-the-money put with the same Vega; he e l i ~ n a t the ~ s residual delta exposure introduced by the last transaction by dealing in an a p p r o ~ ~ a (delta) te amount of stock and borrowing or lending cash.
Table 8.1 Two sub-portfolios, before (the two columns on the left) and afterthe jump (the two columns onthe right). For each of thetwo columns, the first column refers to the sub-portfoliowith the out-of-the-money option,and the second to the sub-portfolio with the at-the-money option. The stock price before the jump is at 100 and the residual maturity one year; thetwo strikes are at 100 at 80. The amount of cash before the jump makeseach sub-portfolio worth zero. A 'ratio' amount of portfolio 2 makes the Total portfolio kappa and gamma neutral. When the jump occurs the stock pricemoves from 100 to 80
Strike Price Delta Gamma Kappa Cash Port Ratio Total portfolio
80.0000 6.3725 -0.4602 0.0248 31.7562 -43.1 862
100.0000 21 .l 859 -0.845 1 0.01 49 19.0534 -88.7954
0.0000
1.6667
-0.0000
-0.0000
80.0000 6.3725 -0.1 120 0.0248 19.0534 -12.3906 2 .g456 0.4800 -1.01 75
100.0000 21 .l 859 -0.4602 0.01 49 39.6953 -53.9828 4.01 69
-0.0004
-0.0006
99.5
100
100.5
-0.000~
-0.001 I
f
tock Price
The P&L of the Total portfolio described in Table 8.1 as a function of the post-jump value of the stock. Note how the Total portfolio is amm ma-neutral at the origin
Therefore, if it were the case that own war^ j u ~ are ~ ~ s o r e ~ ~ e ~ u e ~ t j u ~ ~then, s , on average, the expectation3 of the value of the overall portfolio will be negative. Note that, for the moment, the delta and Vega amounts put in place by the arbitrageur are those suggested by the usual Black formula used with the at-themoney volatility. Also, we have, rather arbitrarily, assumed that the arbitrageur is trying to obtain the biggest bang for the buck, i.e. that he has chosen to buy the option which has, at the same time, the ~ a ~ i m u difference m in implied volatility, but whichis closest in strike to the chosen out-of-th~-~oney put. Given the equity smile curve he observes, the option chosen as a buy is therefore the at-the-money option. This assumption is not necessary, and the choice of option strikes will be discussed later on, If the trader has correctly guessed the volatility of the purely time-dependent diffusive component of the process (see Section 8.3 for a more precise description of the mathematical set-up), and this volatility is exactly the at-the-money volatility, then the overall portfolio will make exactly no gains and no losses until the first jump. (~ontinuous, friction-less trading is, as usual, allowed.) M e n the first jump occurs the portfolio will no longer have an overall zero value, and, as shown above, over a downward jump the overall portfolio will always makea loss. The portfolio made up of the out-of-the-~oney put, which the
Note that, for the moment, we are taking this expectation with respect to the real-world (econometric) measure.
arbitrageur has shorted, and its delta-neutralising portfolio, will in fact increase proportionally in absolute value by more than the corresponding at-the-money put and delta hedges. The effect of a particular downward jump on a portfolio that has been kept balanced at all times before the rare event, as explained above, is shown inTable 8.1, The portfolio P&L over jumps of di~erentamplitude (upward and downward) is shown in Figure 8.1. A s mentioned above, even if we assume that, in the real world, the likelihood of occurrence of upward jumps is the same as the probability of downward jumps, the presence of jumps, of whichever sign, will still introduce a finite variance to the terminal value of the portfolio. If the arbitra~eur is risk-averse, and perceives risk in terms of portfolio variance, then he will therefore demand some compensation for this form of risk even if the expected jump amplitude were zero (or, up to a point, even if it were positive). Conversely, if downward jumps are more likely thanupward jumps, even risk-neutral arbitrageurs will require a compensation above the at-the-money volatility before selling the outof-the-money put. Clearly, the maximum jump compensation will be required by risk-averse arbitra~eurs who espect d o w n ~ a r d biased jumps. It is essential to note that in the model we have outlined the process of the underlyin~equity index is of a mixed diffusive-jump nature. The jumps, in particular, are of random amplitude, and the market is therefore incomplete with respect to their occurrence (unless one introduces other ~ n s t r u ~ e nthat t s price isk-neutral valuation will therefore fail to provide a ~~~~~e~ option price for the problem described above, and the result will depend on the appetite for risk of the arbitrageurs. This feature is obviously computationally unpleasant, but, in a way, it is one of the central features employed in this chapter in order unt for the stylised facts that characterise equity smiles. sections a particular modelling approach will be presented which can be used in order to account for the presence of smiles in the e uit market by means of a mixed jump-diffusion process. Several existin starting from Merton (1990), have dealt with the case of mixed jumphe approaches that can be found in the literature differ in eemed possible. It is therefore important to state clearly fundamental securities a s s u ~ e d to exist at the traders disposal for hedging purposes in the different approaches. In one important category, to which the model presented below belongs, the hedging of an option with anything but the underlying stock (or index) is disallowed. The second pricing framework allows hed ing of an option position (e,
It is possible, as we shall show below in a special case, to hedge with a ~ nu~ ~b e r of ~hedging ~ e options if the jumps are random but their possible ma~nitudes finite in number. Only an ~~~~~~e number of hedging options could hedge random jumps with a continuum of possible amplitudes. More precisely, we can no longer rest assured that there exists a unique ~ a ~ i n g ameasure le (see, of example, Pliska (1997) or Bjork (1998)).
out-of-the-money puts) with one or more other option positions (e.g. the at-themoney puts); the last modelling approach assumes that the trader can include in his strategy an infinite number of options thereby creating, as shown below, a perfect hedge against the infinite possible realisations of the jump amplitude. In other words, the second approach recognises that ~ o degree ~ eof option hedging is possible, and likely to be entered into by the market p ~ i c i p a n t sbut , that the resulting portfolio does not exactly replicate the t e r ~ i n a pay-off l of the plainvanilla option. This observation will become crucial when the discussion moves from calls and puts to the case of exotic options. Finally, it is impo~ant to stress again that, given the relative market positions (long or short) in the model sketched above of the fund players (who seek insurance) and of the ar~itrageurs, the latter will either refrain from trading or take a precise side of the trade; in other words, if they trade at all, the arbitrageurs in the model sketched above will be seEEers of the expensive out-of~the-~oney options and buyers of the cheap at-the-money options. In this description, one is therefore not placing oneself in the framework of a market maker who is asked to make a price in a security (trading in a possibly incomplete market) without knowing whether she will be bidding for or offering the security itself. Therefore, despite the fact that, in some circumstances, options can be priced, even in incomplete markets, using a risk-neutral framework (on this point see Bjork (1998)), this is not the case for the financial model we are desc~bing here. Before embarking on these tasks it is however necessary to introduce some of the fundamental mathematical and financial concepts connected with mixed jump-diffusion processes. This is undertaken in the next section.
Since jump processes are less widely used in financial applications than ~rownian diffusions, I shall review, in as simple a manner as possible, the accompanying t e r ~ n o l o g yand formalism. (See also Merton (1990) for one of the earliest descriptions of jump processes in the context of option pricing. Many of the applications of mixed jump-diffusion processes are in the defaultable-bond arena. Of these, one of the most interesting papers of which I am aware that deals with the necessary change in measure when jumps are possible is ~ c h o e n b u c ~ ~ r ( l 9 9 ~ We shall assume that the stock behaviour is described in the real(econometric) world by a mixed jump-diffusion process of the form:
In Equation (8.1) S ( t ) denotes the stock price, as(t)its percentage volatility, dz(t) is the increment of a standard Brownian motion, A ( S ) is the (random) amplitude of the jump, and (JJV(t) is the increment of a counting process. A s for the latter,
a counting process N ( t ) represents the number of times a jump has occurred between now ( t = 0) and time t. We therefore put N(0) = 0. If we impose that the jump process is Poisson in nature, we h o w from the elementa~ properties of Poisson processes that, if h is the frequency (intensity) of the jumps, the probability of k jumps having occurred out to time t is
PEN($) = k ] = esp[--ht](ht)'/lk!
(8.2)
In order to obtain the expected value of the increment, d N ( t ) , of the counti~g process, let us recall that, for Poisson processes, the occurrence of a jump is independe~t of the occu~ence of previous jumps, and that the probability of two simultaneous jumps is zero. Let usthen assume that, out to time c , j jumps have occurred. Then, over the nest small time interval, dt, one extra jump will occur with probability h dt (see Equation (8.3)), and no jumps will occur with probabi~ity1 h dt. The change in the counting process will therefore be 1 with h dt. Therefore probability h dt and 0 with probability 1 E[dN(t)] = I*h dt
+ O"(1 -h dt) = h dt
(8.3)
If the jump frequency is constant, then the expected number of jumps from time 0 to time t is therefore given by
We can then define the process M(t) as
zero: By the definitions given above, the expectation of dM(t) is, by const~ction, hdt] = E[dJV(t)] hdt = hdt hdt E[dlM(t)] = E[dN(t) = = :
(8.6)
or, in other words, the process M ( t ) is a m ~ i n g a l e The . process ~ ( tis )often a t e ~ process, since it is constructed so as to referred to as the c o ~ ~ e n ~counting ensure that on average it exactly compensates for the increase in the umber of jumps accumulated in N ( t ) ; see Figure 8.2. use of these definitions, Equation (8.1) can be re-written as
Note that we are still not saying anything specific about the drift for the stock, and, in particular, we are not invoking a risk-neutral evolution for the stock cs price. With Equation (8.7) we have simply re-written the original ~ y n a ~ for
3
1
I
-1
-3
i ~ ~8.2 r e A realisation of a counting process N(T) and its associated c o ~ ~ e n s a t e ~ process M(T). Note that the expectation of M(T) is 0, .whilst the expectation of N( T) grows with time(at the rateh)
the stock price as a process with a new drift and a compensated counting process which multiplies a jump amplitude A(S). We have notsaid anything specific about the latter either. Let us now consider an arbitrary function, f ( S ) , of the stock price. If its underlying dynamics is of a mixed jump-diffusion nature, its stochastic differential equation (SDE) will contain a familiar Ito-like component, and a new c o n t ~ b ~ t i o n arising from the jump events. More precisely:
af +z o S s d z ( t ) + [AA(f>lm(t)
(8.8)
The first two terms are familiar. The last term ( [ A A ( ~ dN(t)) )] has the following mean in^: the symbol A is a differencing operator, whose action on the argument is to produce a jump of magnitude A in the variable S: AA(f) f(sr +A(s>, $1 f(St, $1 (8.9) Therefore Equation (8.8) means:
= = I
(8.10)
where the compensated process has been re-introduced. Let us now make some assumptions about the amplitude A(S). After the jump the stock price moves from S to, say, S A($), Let us impose that the quantity A ( S ) should be of the form
A ( S ) = S(t)(a l)
(8.12)
If that is the case, then the change in stock level before and after the jump is given by
Safter
Sbefore = Sbefore(a
1 )
(8.13)
i.e. the difference in the stock price before, Sbefore, and after the event, Safter, is propo~ional the value of the stock itself before the jump. This is particularly appro~riate for the case of equity index markets, where jumps in the index are meaningful in terms of their ~ e Z ~ t i v rather e, than absolute, magnitude.6 Note that, if a is greater than one, the jump will be upwards, and downwards if it smaller thanunity. Also, a cannot be negative, so it is a ~ ~ e a l i n to g describe it as a ~ o g - n o ~dist~buted ~ ~ l y variable, with expectation pa and variance : ; 0 E[ln(a)] = P a
0 var[ln(a)] = :
(8.14)
If these assumptions are made, then the dynamics for the underlying stock is given by dS(t) = [p($) A(S)h]dt
(8.16) with p = p,/S, From ~ q u a t i o (8.1 ~ l)-the extended Itos lemma-one can easily obtain the process followed for the particular but importa~t case when the function of S is given by f = ln(S). In fact, recalling that A = S , ( al) and S, + A = a s , ,one can write
+ + ~ s ( t ) S ( tdz(t) ) +A(S) M(t) =[ ~ ( + t )S(t)(al)h] dt + ~ ~ s ( t ) Sdz(t) ( t ) + S(t)(a 1 )dM(t) =+ dS(t)/S(t) = [ p , ( $ )+ (a1)hl dt + ~ s ( tdz(t) ) + (a-l)
M(t)
If the Dow-Jones or the J?TSE were re-based tomorrow by dividing the current value of the index is emphatically not true for interest rates, where the absolute level does have an intrinsic, and not purely conventional, meaning.
by a factor of 10, then nothing financially meaningful has happened; the same
where use has been made of the fact that, for f = h($), ~ f / = ~ l/$ S and ~ 2 f / ~ = S--l/S2, 2 and the last term follows from the de~nition of [&(.f)] ~ ( t = )[ f ( S +A) f(S)]W ( t ) applied to f = ln(S). Given the properties of the logarithm, one can write the term [In($ +A) ln(S)] as [ln(S + A ) In(S)] = [ln(Sa) In(S)] = and therefore Equation (8.17) can be re-written as
p d In(S) = [
o
(8.18)
$U-:] dt + dz(t) + ln(a) dJV(t)
(8.19)
or, equivalent~y,
p io: dln(S) = [
+ h ln(a)] dt + CT dz(t) + In(a)
in terms of the compensated process d ~ ( t ) Either . espression can then be integrated from time 0 to a generic time t , to give (focusing on Equation (8.19) for concrete~ess):
~
dln(S) = ln(S(t)) ln(S(0))
Equation (8.21) shows that it is possible to simulate a realisation of the stock price from time 0 to an arbitrary large time t using a s i ~ g Zas~long ~ as ~ one ~ ~ , can draw three random quantities: the usual realisation of the ~rownian process: Z ( t ) = :+/t, with E E N(0, 1); the realisatio~ of a, which we posited (in Equations (8.14) and (8.15)) to be log-normally distri~uted with E[ln(a)] = pa, var[ln(a)] = : : 0 and the realisation of N ( t ) , about which we h o w that E [ N ( t ) ]= At, and var[N(t)] = At and that it is distributed according to P [ N ( t )= k ] = exp[-ht](~t)~/k!.
In this respect, the Monte Carlo simu~ation of a jump process is only marginally more cumbersome than the simulation of a digusion process. This result might suggest that one could follow the procedure consisting of
si~ulating, by means of Equation (8.21), many realisations of the stock price; evaluating the option pay-off;
discounting the option pay-off; repeating many times and averaging the discounted pay-offs; compa~ng the obtained values with the market prices for plain-vanilla options; varying the p a r ~ e t e r of s the simulation until the best overall agreement can be found across maturities and strikes. Despite its intuitional appeal, the approach outlined above is not as straightfor war^ as it might appear, since, given the financialstory presented in the previous section, one should expect that the risk-neutral valuation would not be capable of providing a unique option price, since risk aversion might enter in the formation of the latter. It is therefore essential to explore in detail up to what point (and if at all) a claim on a stock that can experience jumps can be hedged. For this purpose, let us try to hedge a plain-vanilla option on an asset which undergoes a nixed jump-diffusion process by using the underlying stock itself as a hedging instrument. Let us call C(S) the contingent claim, and let us make use again of the extended Itos lemma introduced before. Our strategy will be to attempt to eliminate the Poisson component first, and to deal with the usual diffusive term afterwards. In analogy with the approach followed for diffusive processes, we shall try to solve for the 6 amount of stock that will neutralise the Poisson component, (For a similar approach, see eumee et d . (1999), keeping in mind, however, that their setting, and consequently their results, are somewhat different. Their Figure 1 is of particular interest.) Let us define I1: = C ( S ) &S, and write
+ 6[p(t) +A(S)h]dt + a;(;(t)S(t)dz(t) +A ( S )~

Let us concentrate on the terms in dM(t). If we want portfolio against the Poisson jump, we must set [ C ( &+ A ( S ) , t ) C(&, t ) ] = 6A(S) = 6S(a 1)
( ~ (8.22) ) to try to neutralise the
If one recalls (Equation (8.13)) that Sbefore(a 1) = Safter Sbefore, then Equation (8.23) is very suggestive, because it is r e ~ n i s c e n of t a Black-delta amount of stock. Indeed, if it were the case that C(aS) = aC(S), then one could write the term [C($, A(S), t ) C($,)] as C(S)[al], and the expression for the a ~ o u n t 6 would simply be given by 6 = C/S, with the amplitude a not appearing at all. However, this is a very special case, i.e. the case when the contingent claim is a linear function of S. In general, the 6 amount will depend on the amplitude a. I f this latter quantity were a known number (i.e. if the distribution we have posited
h o w exactly how much of the underlying stock we have to use in order to neutralise the effect of the Poisson jump. We would then simplyhave to take care of the Brownian diffusion tern, which we could handle using the classic Black and Scholes reasonin introducing a A amount of another contingent claim on S. We can now under in a more precise manner the arguments previously made (see Chapter 2) that the introduction of a Poisson jump of known amplitude requires the use of an extra hedging instrument. The discussion about the dangers and limitations of ing on the basis of the equivalent diffusion obtained from a distribution ated by a mixed jump-diffusion process (see Chapter 4, Section 4.6) can e re-visited at this point. In the general case, however, the amplitude a is a random variable. A s a consequence, we cannot h o w the amount S with certainty, and our hed i~pe~ec In t .the context of the strategy described above of the arbitrageu sells the o~t-of-the-mone~ putand buys the at-the-money option, there is no a priori knowable amount of stock that will insulate both components of his po~folio from the jump. Given the is cuss ion above, and qua ti on (8.21), which provides an efficient numerical solution to the evolution of a mixed jump-diffusion process by means of a single long jump, at this point one can re-visit the possible simulation' approach to the problem of the pricing of smiley equity options outlined above. It is well known that the investors risk aversion cannot be recovered from the prices of traded options if the underl~ing process is a pure diffusion: Blacks for~ula is independent of (i.e. is valid for) any pattern of risk aversion. In the presence of the kind of market incompleten~ss generated by the presence of jumps it would seem, however, that the risk aversion might be recovered from the option prices. ~nfortunately, given the vast predominance, in the option literature if not in reality, of diffusions, as opposed to jump processes, pra~titio~ers have become very lazy in distinguishing between risk neutrali~y,absence of arbitrage, absence of risk aversion, etc. In what follows we shall therefore attempt to clarify some of these important concepts, in so far as they directly affect the problem at hand. It s h o u l ~ be recalled, to begin with, that if the process for the un~erlyi diff~sion and all the usual perfect-m~ket assumptions are met, then we need ~ s ~ - n e u t rinvestors al for the Black and Scholes formula to apply. Whatever their risk aversion, it will not enter their appraisal of the value of the option because the variance of returns of the hedged portfolio can be made exactly zero, In other words, in the well-known expression p = r ha (where p is the drift, a the volatility and h the market price of risk), it is not h that goes to zero, f the ~ o r ~ o Z i o j . (i.e. the volatility o enever one is dealing with a diffusion and one constructs a tree, or p e ~ o r ~ s te Carlo simulation, to price a simple option, one should therefore have, in the back of ones mind, a parallel tree o r simulation in which the accompanying replicating hedge is evolved (backwards or forward as appropriate).
for a were to collapse to a Dirac S), then we would
Strictly speaking, it is only the c o ~ ~ portfolio i ~ e that ~ will have zero variance, but, since life is short, we often take a short-cut and dispense with the construction of the parallel tree: we recognise that the variance of return of the option by itself is anything but zero, but obtain the correct result by pretending that the risk aversion of the investor (rather than the variance of the overall portfolio) were zero. In other words, in the expression p = r ha we cheat, and put h rather than a equal to zero. The advantage is that we can now dispense altogether with the parallel tree built in the back of our mind containing the perfect hedges. When, however, a perfect hedge cannot be put in place, perhaps because of the existence of a jump component in the process, this convenient sleight of hand is no longer possible. The variance of the portfolio now becomes crucial, together with the risk aversion of the investor. Valuing the option by itself (without the partial hedging strategy we have chosen) no longer makes sense, since the perfect hedging portfolio somewhere in the back of our mind no longer exists. At the same time, since there is no such thing as the perfect hedging strategy, there is ng portfolio. We have therefore reached the first result: the trader overall variance of return for every partial hedge she might dream up. The ~nancial complexities, however, are not over yet. Since we have been spoiled by the familiar ~rownian processes with perfect hedges, we have become accustomed to identifying absence of arbitrage with the existence of a unique price for the option. This is, however, a special case that applies only when a perfect replicating strategy can be put together, and the overall portfolio (i.e. the one made upof the optionand of the parallel perfect hedging strategy in the back of our mind) is purely deterministic. In eneral, given a set of possible prices for different securities in different states of the world, three different cases can arise (see Pliska (1997) for a very clear discussion):
1. a pricing measure exists and is unique; 2. there are in~nitely many pricing probability measures; 3. there are no pricing measures.
It so turns out that the case 1 applies to the case when markets are complete, claim is replicable (i.e. when markets are but 2, is true if not every conti~gent not complete). ~rbitrage, however, should not be permitted to exist not only in the restrictive case l but also in case 2. Looking at the problem in this manner, obtaining a unique price from a no- bitr rage argu~ent, as is possible in case 1, is the exception rather than the rule, and the lack of uniqueness of option prices in the case of incomplete markets is of a very fundamental nature. an incomplete market a single quoted price for an ecause, out of the countless possible pricin that could co-exist in an incomplete market, the market chooses a single one: since the quoted price is the level where the market clears, i.e. where for every willing buyer there was a willing seller, the pricing measure is chosen among the
in~nitely many possible by the process of market clearing. (This, incidentally, is orks (1998) often repeated quote: Who chooses the measure? The market ! . ) Whether, to what extent, and under what conditions the investors, risk aversion can be recovered, by itself or in conjunction with other statistically accessible quantities, from observations of option prices in incomplete markets is a complex digression. Were topic, which would require more than a short and se~f-contained it willsuffice to say that, given the context of Smile Tale 2, and given, in p~ticular,the fund managers need for insurance, the market price for jump risk might be easier (or less difficult?) to estimate than inother, more symmetric, situ~tions. Let us therefore assume for a moment that we have found an implied methodology by means of which, from the prices of plain-vanilla options, we manage to gain access to the risk aversion of the arbitrageurs (see, however, the discussion at the end of this section). The technical difficulties would be far from trivial, nce we are not dealing with a market incompleteness arising from a non-traded rownian motion (see Bjork (l 998)), but with a market price of risk arising from jumps (~choenbucker (1996) deals very clearly with the mathematical complexities connected with the appropriate change in measure). This market price of jump risk would therefore not enter as an additive term to the risk-less rate of growth, but as a term that would alter the objective amplitude and frequency of the jumps. Let us ignore also these dif~culties for a moment. Even if we could gain access to this jump risk aversion, what we could obtain from the market prices of options would only be related to the average utility function of the market agents. For all its intrinsic interest, this quantity would not tell the individual trader how an equity option should be priced given her risk aversion (which will, in general, be different from the markets). To use Bjorks ain, in incomplete markets, the market itself, not the trader, chooses the pricing measure. Let us nonetheless make, for a moment, another heroic assumption: let us suppose that the trader, for reasons better h o w n to herself than to anyone else, has decided that, after all the labour she has put into obtaining the markets jump risk aversion, she will adopt it as her own risk aversion. Even inthis case she would still have to evaluate the (non-zero!) variance of return from the option and her chosen hedging strategy. It is essential to point out again that the relevant quantity for her is not the variance of returns from the (exotic) option in isolation, but the variance of returns from the option itself plus its proposed hedging strategy. A s for the latter, not only can there be no such thing as a perfectly replicating strategy (by the very definition of market inc~mple~eness~) but there is not even an a priori general agreement about (or knowledge of)
One of the de~nitionsof incomplete market is a market where not every claim See Pliska (1997) or Rebonato (1998a).
is attainable.
the best hedging strategy. As mentioned above, the resulting overall variance of return of the combined (exotic option + hedging strategy) portfolio, would therefore depend on the p ~ i c u l a hedge r chosen by the trader. Looking at the problem in this light, the trader would seem to arrive at a price not only by taking into account her risk aversion, but also by carrying out a portfolio-~~ance-minimising search over all the possible hedging strategies that can be put together using the inst~ments traded in the economy. This is less su~risin than ~ it might at first seem (one can think of the pricing behaviour of an exotic trader who becomes progressively more confident about her hedging strategy, and consequently tightens her price), but it remains, nonetheless, a rather f o r ~ d a b l e task. These conceptual problems are in principle not insoluble, but it is difficult to see how a practical solution to the pricing of an exotic product in the presence of jumps could be arrived at along these lines. A different approach is therefore presented in the next section by introducing some upper and lower bounds for the equity option prices outside which no rational investor should quote, irres~ective
of his risk preferences.
The idea of upper and lower bounds for option prices is not new, and has a pedigree as illustrious as it is ancient, tracing its origin as it does to Mertons (1973) paper on rational option pricing. The difference in the approach described in the next section is that whilst Mertons required nothing more of the investor thanher being ratio~aland preferring more to less, some relatively mildbut powerful assumptions are now added that manage to reduce the width of the no-arbitrage channel in a substantial way.
ritten-Jones and Neuberger (1996) (BJN in what follows) introduced a wayof looking at option pricing which enjoys several useful and desirable features. Their approach is somewhat unconventional in that, relying as it does on the concept of dominating strategy (see, for example, Pliska (1997)), allows the deter~nation not of the price, but of the upper and lower bounds for the price of a contingent claim in a variety of financially impo~ant situations, where a full-blown, risk-neutral valuation is not possibleor appropriate. More specifically, N provide the user with a unified approach to deal with no-arbitrage pricing complete and incomplete markets.At the same time, theysupplyboth the theoretical framework and the numerical tools to treat the case of jump processes, which, as we have seen, could provide a powerful mechanism to explain in financially con~incing manner the presence of smiles in the equity and the F markets, JN also deal with the problem of option pricing whenonly discrete, as opposed to continuous, re-hedging is possible; and, en ~ ~ s s they ~ ~ t
provide yet a different way to look at the hed ing paradox of Crouhy and (jalai (1995) which was first examined in Chapter 2 Section 2.5. Their treatment is, at the same time, very elegant and somewhat unconventional. Its salient features, in so far as they are relevant to the discussion of jumps, volatility and smiles, are presented below. Let S ( 0 ) be the price today of asset S (a stock). We assume that there exists a forward market in the asset S, We must point out from the start t tion is essential for the specific implementation proposed by later on, in fact, they cast a very general, but computationally CO erms of a numerically simpler conventional d y n a ~ i c - ~ r o ~ r a m m i e price paid by BJPJ by doing so is that the class of problems plied is signi~cantly restricted. This important observation will be examined at greater length at the end of ection 8.8, where possible extensions of the method are discussed. The forward price for delivery of the stock at time t will be denoted by S,. A European contin~ent claim, V , is a security whose pay-off at time 1 only depends on the realisation of S at t , i.e. V = V($,). For the sake of simplicity in what follows we shall often omit the adjective European, but this restriction on the type of contin~ent claim is essential, and should always be borne in rnind. Let then n be aprice ~ e ~ ~ i.e. e ~ an c ordered e , collection of prices, Si, with 0 5 i 5 N.It is important to note that the integer N is assume^ to be finite, although S large as one might wish. Therefore a general price sequence is described by
n = {So, S 1 7 $2,
* *
SN}
he price sequence n is said to be ~ e r ~ i s s if ~~ and Ze only if
Condition l , the mildest of the three conditions, is quite obvious and simply requires that, for a price sequence to be per~issible,its first value shouldbe equal to the price of the underlying asset today. Condition 2 prescribes that the jump in the log of the stock price between any two steps should never exceed an arbitrarily large but finite quantity, d. If the underlying true process were a pure diffusion, and since weallow N to be as large (but finite) as we want, this condition would not unduly restrict the properties of a permissible price sequence, since the ma~imumjump size, d, would scale with the square root of the time step. If the underlying true process, however, contains real jumps, perhaps o~iginatingfrom Poisson events, then Condition 2 does pose a s i ~ n i ~ c a restriction, nt and the onus will be on the user to check that the solution to the pricing problem does not significantly depend
on the quantity d. (See the discussion in Section 8.6 regarding this point.) Given the description of the model presented so far, one should note that the treatment does not make any a priori distinction based on the underlying diffusive or jump nature of the stock process. True jumps, finite re-hedging intervals, and, with some caveats, even transaction costs can all be treated on the same footing. A s far as the third conditionXis concerned, it should pointed out that if the stock process were a diffusion, and hedging continuous and friction-less, then it would be equivalent to requiring that the time integral of the square of th instantaneous volatility, o(u), should equal a known quantity, o:vgt :
rt
JO
02(u) du = o&t
(8.24)
Since we h o w that, as far as friction-less option pricing for diffusive price process is concerned: the quantity that matters is just o&t, for this pa~icular case Condition 3 is no more and no less restrictive than requiring a deterministic volatility. The setting presented here, however, is much more general, enco~passing, as it does, both the finite~re-hedging-intervalcase, and the mixed jump-diffusion situation. By imposing Condition 3 we are assuming thatwe h o w a priori the total magnit~de of the moves over the finite re-hedging periods i ~ c l uthe ~ j i ~ u~events. ~ ~ From this point of view, whilst still similar to, Condition 3 is stronger than Equation (8.24) above. We shall return to this point towards the end of this chapter. ~ermissible price sequences are denoted by the symbol P. Since the permissibility depends on the asset, S, via Condition l , on the ma~imum jump size, d, via Condition 2 and on the total variance, v , via Condition 3 they are denoted by d, 4. Having defined a permissible price sequence, one can introduce the concept of residual valati~ities, v i , 0 5 i 5 N , defined for each step i, as
From Conditions 1 and 2 above one can deduce
Therefore the quantity vi conveys information about how much of the total variance we still have available to use at time step i ,
The reader familiar with stochastic calculus will recognise Condition 3 as being akin to requiring that the process should have a finite and known quadratic variation. At least as long as the instantaneous volatility in Equation(8.24) is a purely deterministic function of time.
23
Finally, we can introduce a trading strategy, H , which specifies the number of forward contracts held at time step i , and which we shall allow to depend, in a manner so far unspecified, on the stock price and on the residual volatility at time step i. Therefore, H = H ( S i , vi). As long as we focus our attention on a purely European option, V , we are in a position to consider the strategy consist in^ of
1.. selling the contingent claim at time 0 for an amount of cash equal toV0; 2. paying out to the buyer of V the pay-off (if any) of the co~tingent claim at time T ; 3. accumulating the profits and losses arising from the strategy H (which we shall call the 'hedging profits').
To disencumber the notation in what follows we shall assume zero interest rates. Therefore the profits from the overall strategy are given by
~ge any ~ m o n g s tall the admissible strategies H, we call an ~ r ~ i t r strategy strategy H such that, for all permissible paths (i.e. for all permissible price sequences)
where the inequality is strict for at least one path. if this condition were satisfied, then the seller of the option would ce~ainly receive at least as much, and some, times more, money up-front (V,) than what he has to pay at the end, V ( S T ) plus the costs that the strategy entails, i.e. H ( S i , Vi)(Si+l Si). ~e can, at this point, choose an arbitrary trading strategy, say H I , i.e. we can choose an arbitrary rule that will tell us the position in forward contracts that we propose to hold as a hedge against the European option. This rule will be a ~et~rministic function of all possible values of the residual volatility and stock price.I0 Note that, at this stage, we do not know whether this arbitrary rule will indeed serve as an effective hedge, and we might want to alter it in the future. or the moment it is just a totally arbitra~" rule that takes the stock price and the residual volatility as inputs, and returns the amount of forward contracts that we shall enter.
xj=o,N-l
lo Note that we have not said 'for any time and stock price'. As we shall discuss at greater length in the future, the relevant variable is not time but residual volatility. As usual, we shallassume that all the technical conditions are satisfied. In particular, we obviously require this strategy to be predictable (see, for example, Rebonato (199th)).
33
Since we do not want to allow arbitrage strategies,12 it follows that the price for V we obtain today, Vo, cannot be larger than the maximum cost incurred in paying the final pay-off of the claim, and accumulating the hedging profits or losses for the c ~ o s strategy. e~ The maximum (supremum) is therefore over all the p e ~ i s s i b l e paths: (8.25) Our choice of trading strategy was completely arbitrary. We want, however, 2 1 trading strategies; that is, even if one found Equation (8.25) to hold for a second trading strategy, say 232, such that the right-hand side of Equation (8.2 were always lower than if we had followed strategy H I , the inequality (8.25) would still have to be fulfilled. In general, therefore, the price we receive from the sale of the option today cannot exceed even the lowest (infimum) possible realisation of the quantity { V ( S T ) H ( S i , vi)(Si+l S , ) } over all the possible trading strategies. Therefore
+ xj=o,N-l
Therefore Equations (8.25) and (8.26) above express a minimum upper bound for the price today of the contingent claim in two steps: first, for a given strategy, we determine the price that is certain to cover the total payout (expression (8.25)) for every possible admissible path; then, we vary the strategies H in order to reduce the total pay-out (i.e. the pay-out arising from the final payment and the re-hedging costs) as much as possible, resting assured that the price V0 will have to be lower also than this value. Finally, we ~ e this~minimum ~ eupper bound as
This upper bound has been expressed in terms of a search over paths and strategies. Clearly, we could have reversed the sign in the argument, and we could have obtained a similar expression for the maximum lower bound. Given the complete symmetry of the two approaches, only the case of the minimum upper bound is developed in what follows.
l 2 What is actually prevented is the possibility of a strategy. See Pliska (1997) for a clear distinction between violation of the law of one price, the existence of dominating strategies and of a dominating strategy always entails the possibility the possibilityof arbitrage. Since the existence of arbitrage, for the present purposes no distinction is made between the two cases.
~~~~~~~~~~
Whilst conceptually correct, this approach is computationally quite impractical. In order to circumvent this difficulty, BJN recast the same problem in a different form. To this effect, note that, since the current price of the asset, S, and the total variance determine both the permissible price sequences and the tradin ies, H ,the function V($,v) just defined provides the minimum upper bound as a function of S and v. More precisely, V is a function with a domain given by all the possible values of S and v. At a future point in time, i.e. as the price evolutio~ unravels, the function itself will not change, only its domain will. As a consequence the minimum upper bound in state i can be written as V ( S i ,vi), When the problem is looked at in this manner, there is no special ~ e a n i ~ attache^ g to the initial and final times; in p ~ t i ~ u l aone r , can focus on two con~ecutive7 steps, and write
Note that, in moving from Equatio~ (8.27) to Equation changes have taken place: the supremum is no longer t over realisations of S and v at the next step: for a given (Si, vi) we simply have a finite number of possible terminal desti~ations(Si+l, vi+l) to keep track of, instead of a multitude of connecting paths; also, the sum over time steps has ~isapp~ared, since we are now dealing with a single time step (at a time); finally, V(Si+l) no longer indicates the terminal pay-out, but simply the value of the ~ n i m u m upper bound itself at the next time step, Let us now implicitly define a variable z i by the following relations hi^:
From Condition 2 on p. 231 we can write
Therefore Equation (8.28) can be re-written as
V ( S i ,vi) = inf{sup {V(Si+r,vi+l> h
= V(Si,vi) = i;f
+ h ( S ~vi)(Si+l , Si))> (~ I)}} ~) {sup {v(Siexp(zi), vi -z;) + h S ~ ( e x ~zi
as long as the quantity ! z does not exceed either vi (otherwise the residual volatility vi+l could not be positive), or d2 (so as not to violate the condition on
(8.2~)
the maximum jump size). Since we are now speaking of two consecutive time slices, and the sup re mu^ simply has to be searched over all admissible values of zi, the settin look more like a traditional dynamic programmin~ problem (i.e. a usual tree- or
p- D i ~ ~ s i o Processes n
35
lattice-based backw~d-induction methodology). There still remains, however, the infimum over the trading strategies that makes the problem non-standard. This last hurdle is dealt with by BJN as described in the next section.
. This is obvious from the definition, and can be verified by direct substitution in Equation (8.31)
e
This property follows from the definition (8.31) and the
property (8.30). This property follows from Properties 1 and 2: if the ~ n c t i o nf is equal to zero at the origin, and greater thanor equal tozero for positive ar negative values of z, as long as arbitrarily small in ma~nitude (Property 2), then the derivative of f with respect to z at the origin must be equal to 0.
This derivative
3 is easily calculated:
valuating this ~uantity at z = 0, and equating it to zero because of Property 3, gives
(8.32) Equatio~ (8.32) therefore determines the optimal strategy. Note that if all the Black conditions were met, then the usual expression for the Black-delta holding
of forward contracts is recovered. The optimal strategy, however, holds also in all the much more general cases where the BJN method is applicable. As a final step we can substitute the optimal strategy we have just determined in the expression for the minimum upper bound to obtain
subject to the usual constraint on z (see Equation (8.32)), and with the boundary (initial) condition
V ( & ,0) = V ( & )
(8.34)
Condition (8.34) will be examined in greater detail in the next section. In the meantime Equation (8.33) expresses the solution of the ~ n - m problem a ~ in terms of a search over the admissible values of z. In the expression we have obtained there appears an (unknown) derivative of the minimum upper bound function with respect to the forward price. The set-up is therefore quite similar to the usual Black en~ironment, where initial and boundary conditions allow the solution of a second-order parabolic partial differential equation. It is therefore not surprising that a numerical technique very similar to the binomial (and of which the binomial is a limiting case) can provide a numerical solution to the problem. How this is accomplished in practice is shown in the next section.
In this section we shall analyse the implementation of the approach proposed by Britten-Jones and Neuberger in detail, with a view to highlig~ting both the practical aspects of their construction, and the conceptual implications of their procedure. Let us consider the case of a four-month call option with the underlying, S, at &l00today, and a strike of E100 as well. We shall begin by building in a slightly different way a 'traditional' binomial tree, i.e. we shall discretise a continuoustime diffusion in one of the many and, in the limit, equivalent discrete ways; no jumps will be allowed in the first part of the exercise. Then we shall implement the B~tten-Jones-Neuberger(BJN) procedure in order to deal with the case when jumps are indeed present, leaving all the other numerical features of the problem unchanged. For the sake of simplicity we shall continue to assume zero interest rates. As for the first part of the construction (i.e. setting u p the 'traditional tree'), one normally builds binomial trees by specifying a time step, At, and a, possibly
time-dependent, volatility, C T , If the two states are linked to their parent state by the relationships S , = SO exp(aylAt)
sdown
= SO exp(-oylAt)
then the usual moment-matching condition requires that the stock price can jump to an up9 ora down state with the well-known Cox et al. (1979) probabilities. We shall follow instead a different procedure: in keeping with BJNs approach, we choose to map onto the x-axis, instead of the time variable, a different quantity, i.e. the residual volatility, v = ~(CTyl~t)2, with j a positive integer discussed at greater length in what follows. A mo~ents thought suggests that this is a very natural transformation, since option pricing does not know about volatility and time separately: in the equation the time variable always appears in conjunction with the volatility, in the form crdAt, or as its square. Introducing the new variable v in the context of a diffusive process is a natural and beneficial numerical device that makes time flow faster or more slowly according to whether the (non-constant) volatility is higher or lower, in such a way that the new volatility-adjustedtime flows at an even pace. See Figure 8.3 and its caption for the numerical details, Once we have reached the last step of the tree (which coincides with the option expiry) all the volatility has been used up, and therefore in every state of the world the value of the option, V , is simply equal to the pay-off condition for the co~esponding value of the underlying. At each step in the tree and for each possible state, BJN denote the value of the option with the notation V(Si, v i ) , in order to emphasise that it depends both on the level of the underlying and on the residual volatility. It is wo~hwhileanalysing the ~ e a n i n g of this notation in some further detail: we should remember that one of the crucial assumptions of the model is that we avail ourselves of the exact knowledge of how much total volatility will be used up by option expiry, but we do not assume to know
. 3 The (Si, v i ) grid: the grid was built by requiring that thevarious quantities Si should be given by Si = S, exp(iS), with 4 2 S 2 -4, and that the values of the residual i , should be obtainable as vj = jS2, with j 2 0. At maturityj = 0 volatility, v
23
how this volatility will actually 'occur' during the option life; note that the BJN assumption therefore entails a milder requirement than assuming to know the instantaneous volatility function, a(t),at every time step. For a pureZy ~ i ~ u s i v e process, it is exactly equivalent to requiring that weknow either aavg(T),or a(u)2du, with the two quantities linked by the by-now-fa~liar relationship
1 ;
The same is not true, however, if the process is a mixed jump-diffusion, or if we relinquish the continuous trading assumption, which is the relevant setting for the JN model. Furthermore, if we no longer impose that we can trade continuously, we lose the ability to distinguish between 'true' jumps (i.e. events originating from the discontinuous nature of the process) and 'pseudo' jumps simply induced by our finite trading frequency. Since, however, we assume to h o w the total volatility, ' U , at all points in our price/residual volatility tree we also know how much variance (originating from jumps, finite re-hedging, etc.) we happen to have V = V ( S i ,vi). In particular, the left. Therefore we can truly and correctly w terminal pay-off condition can be writtenin 'S notation as V ( S i , 0) = V(Si), where V($) is the pay-off function at maturity. For a four-step tree const~ction these values can be read, as usual, in the last c o l ~ m n in the grid (which we shall refrain from calling 'the last time slice'), as displayed in Figure 8.4. In general, the BJN discretisation of a pure diffusion on a computational tree can be built by choosing an arbitrary step size, S; placing the grid values along the y-axis (which give the possible levels of the stock price) at Si = SOexp(i8), with i a positive or negative index; j , the values jS2, with j a non-negative allowing for the residual volatility, w integer; and
The pay-off functionV(Si, 0) = V($) for the 100-strikecall is given in the last column. The other values are the discounted option values evaluated as describe in the text
possible states (i
+ 1, j -l ) and (i -l , j -l).
price at a generic (i, j ) point to be linked only totwo
Note that in the last step of the const~ctionthe second index ~ e c ~ e a in~ e moving from one step to the next, because so does the residual volatility. Since, for the moment, we are simply building the BJN lattice in order to discretise a diffusive process, no further conditions are required, and one can easily check by explicit calculation that the same tree would have been obtained by requiring, for instance, the time step, At, to be 0.1 year, and the volatility to be 2 0 . 0 0 ~ , or any other of the countless possible combinations giving rise to a odAt of 0.063246. We are not committing ourselves to any of these choices, however. Wenow step backwards one residual volatility slice, and we therefore place ourselves in any of these possible states which have in common the fact that only an amount 62 of the total variance is still available to be used (they are therefore labelled by the j = 1). The value of V ( S i , l ) is normally obtained as a linear combination of the values V(Si+l,0) and V(Si_. 1,O). The weights p and ( lp) can be easily obtained by imposing that
P=
1exp(--6) exp(6) exp(4)
which, for the example in Figures 8.3 and 8.4 above, happens to give p = 0.484194. Whilst this is correct, in order to buildan algorithm that can be generalised to the case when jumps are present, we choose to follow a different procedure to bring back the option value. First of all we construct the straight line through the two points V(Si+l,0) and V(Si-1,0) using the following two relationshi~s: V(i
+ l , j -l ) = a+bS(i + l , j -l )
V(i-1,j-l)=a+bS(i-l,j-l)
This 2 x 2 linear system uniquely determines the values of a and b. We then make use of the slope and intercept thus obtained to determine V ( i ,j ) as V ( i ,j ) = a
+ bS(i,j )
Although cumbersome, one can check that the values obtained using this algorithm are exactly the same as the values that would be obtained by means of V ( i ,j ) = p V ( i
+ 1, j -l ) + (l -p ) V ( i -l , j -l )
One can now travel all the way back to the root, and obtain the binomial option value of 4.740256 shown in Figure 8.4.
At this point weare in a position to move to the case of a mixed jump-diffusion process. In order to account for the possibility of jumps, the tree const~ction remains the same, but we modify the possible destination nodes from a given parent node. More precisely, we require that from node (i, j ) the reachable points n2), with n a positive or negative (but not zero) integer. should be (i n , j Some further constraints have to be placed on n besides its being strictly positive or negative: looking back at Conditions 2 and 3 (p. 23O), we do not want a move to be so large that the residual volatility from a given node, vi, is exceeded; nor do we want the ~ a x i m u m allowable (log) jump to be greater than d. These two conditions together require that
n 2 5 (j, d2)
d condition, from the node, say, (-1,3) (with S(- 1,3) = 93.8713), the values highlighted in bold inFigure 8.5 are reachable. If either state (2,O)(S = l 13.484) or state (-4,O) (S = 77.6482) hadbeen reached from ( - 1,3), then all the
Therefore, assuming for the moment that we do not have to worry about the
residual volatility available from the parent state would have beenused and, by construction, the stock price would not be allowed to move any more. The flexibility afforded by the choice of the variable v on the x-axis is now apparent: hadwe built a traditional tree with time on the ho~zontalaxis, we would be forced to move to several possible states on the same time slice, thereby making the recombining binomial geometry impossible. Note that in so far as the last two steps are concerned, the possible values of the option are exactly the same as in the simple b i n o ~ a case: l in the case of the last stepI3 simply because, irrespective of which node we might have arrived from, there is no volatility left, and we must therefore simply have the t e r ~ i n a l
that d 2 3
. 5 The states reachab~e from state (-13) are highlighted in bold. It is assumed
l 3 Remember that the word step does not mean time step: volatility-adjusted time, not regular calendar time, ticks by on the lattice.
pay-off condition:
V ( S i ,0) V(Si)
As far as the second-to-last step is concerned, everything must also look exactly the same as in the simple binomial tree, because, on any node ( . ,l) we only have, by construction, one unit of volatility to play with without violating the total volatility condition. We might have arrived at a node (k, 1) directly from the root, if such big jumps had been allowed by the d constraint, but, given that we have only one volatility unit left, we can only move either to node ( k 1,0) or (k l ,0). Once again the aptness of the seemingly awkward set-up to deal with the case of jumps should begin to become apparent. Note also that it becomes mea~ingless to distinguish in this context between truejumps, or the case where the trader went on holiday and stopped re-hedging his position for a couple of weeks: what matters are only the allowable values at re-hedge time. The procedure truly becomes different when we move to the third-to-last step and beyond. At any of the nodes corresponding to these steps there now is a ~ultiplicityof possible destinations, and, even if we wanted to, we could not calculate the option value as a linear combination of its values at the destination nodes. It is at this point that the optimal hedging strategy determined in the previous section comes into play. To see how this can be accomplished, let us look closely at the expression
V ( S i ,vi) = SUP{V(S~ exp(zi), vi $1 h*Si(exp(zi) I>>

Zi
If we are working our way back from the final expiry to today, V(Si,vi) is the new value for the option that we have to determine, and V(Si exp(zi), vi 2;) are option values that have already been calculated, We are holding a hedging , therefore the change amount of forward contract h* (with h* = ~ V / a S i )and in value of our portfolio will be linear in the stock price. Finally, weknow that the stock price in the state where we have to determine the option value is equal to Si. Let us move inside the sup{ operator and consider an arbitrary z?). From Equation (8.33) we h o w that (known) later value V(Si exp(zi), vi if we have chosen the correct Z i , (i.e. the zi for which the expression in curly brackets is maximum), then the new (unknown) value, V(Si, vi), is linked to the later value, V(Si exp(zi), vi z;), by a linear relationship :
S }
v(s~,vi) = V ( S exp(zi), ~ vi z:) h*Si(exp(zi) 1)
What we do not knowin this expression is the slope h*, i.e. the amount of forward contract (or, more simply, of stock, since we are working with zero interest rates)
to hold. We do know, however, that V(Si, vi) must be an upper bound. Therefore, for any possible value of the arrival Si exp(zi), its value must be worth at least as much as any of the known reachable values V(Si exp(zi), vi z?).In addition, it must be a lower upper bound, and therefore, putting both constraints to line that we st ~ e t e r ~ i is n ethe one that has the we deduce that the strai~ht lowest ~ ~ s s i bvalue l e at Si, and s u that ~ all ~ the ~ o i n t s on it lie at or a ~ o v e the v a l ~ e sV(Si exp(zi), vi z?). This line defines what BJN call accessible o~tion the convex hull. Figure 8.6 illustrates the construction graphically. ~ the ~ re~uired ) hedgin~ ratio. After The slope of the straight line ( ~ V /gives this construction has been carried out at each node for a given variance step one can move backwards following the same procedure all the way to the root, where the required minimum upper bound will be obtained. Despite the somewhat unusual const~ction,the algorithm can actually be coded i n a very fast and ef~cient way. The following section presents the computational results obtained following this procedure.
80
90
100
110
120
130
.6 Theconstructionoftheconvex hull straight line: the stock price atthe node (Si, v ; ) where the option value has to be evaluated is 106.21. The diamonds at thereachablenodes labelledArrivalValuesindicatethevaluesoftheoption ;x;). The straight line is the lowest line at S; = 106 such that all the V(S;exp(zi), v r ~ ~ c h a bpoints le lie on or below the line itself. The value of the option (13.00) is then ; = 106.21 given by the value of the straightline corresponding to S
The case study analysed in what follows is that of a plain-vanilla two-month call option, with the underlying at &loo. If no jumps are allowed, and the volatility of the u~derlying is taken to be equal to 2096, then the relative Black price is E3.2564. Figure 8.7 shows the convergence (as a function of the number of steps) of the procedure presented above for the degenerate case when jumps are not allowed (i.e. when the convex hull is trivially constructed using only two points). Similarly, Figure 8.8 shows the convergence results when jumps are indeed allowed. Despite the fact that the convergence is somewhat slower, it can noted that no particularly challenging numerical problems seem to be encountered. The option values for different starting levels of the stock price, S(O), are reported in Figure 8.9 for 124 steps: it is interesting to note that the jump cases give rise to option prices that are always as least as large as, and most of the time larger than, the option prices obtained without jumps; larger ma~imum possible jumps give rise to option prices that are always as least as large as, and most of the time larger than, the option prices obtained with smaller jumps; the increase in option value associated with larger and larger maximum jump sizes seems to become increasingly smaller in moving from the no-jump case to 396, 5% and 6% ma~imum jump sizes. This is quite impo~ant, given the comments made above regarding Condition 2 (p. 230) for a price sequence to be permissible.
3.29 3.27 3.25
a ,
0
3.23
3.1 7
3.154 0
20
40
~ u r n ~ of er Steps
60
80
I t
100
120
140
F i ~ u 8.7 ~ e The convergence of the BIN procedure as a function ofthe number of steps 100, spot 100, interest rates= 0, for the degenerate no-jump case. Trade details: strike maturity = 2 months, volatility = 20% (jumps 3%,5%,6%). Black value 3.2564
3.1
0
20
40
60 80 120 100 ~ u ~ bof e Steps r
140
The convergence of theBIN procedure as a function of thenumber of steps for the jump case. Trade details: strike 100, spot 100, interest rates = 0, maturity = 2 months, volatility = 20%, jumps 3% Uumpl), 5% Uurnp2), 6% Oump3). Black value 3.2564
16.0000 T14.0000
at
12.0000
4.0000
2.0000
88
93
98
103 Stock Price
108
113
Figure 8.9 Option values for different jump sizes as a function of different starting levels ofthe stock price S(0)
From a graph such as the one presented in Figure 8.9 it is possible to obtain the implied volatilities that, plugged into the Black formula, would give the required prices. It is clear from the graph, and indeed by construction, that the impliedvolatilities obtainable from the jump pricesmustbehigherthanthe co~esponding no-jump quantities. What is not a priori obvious from Figure 8.9,
however, is whether this increase involatility varies across strikes, or simply amounts to a rigid upward shift across strikes of the no-jump volatility. In order to explore whether the BJN model gives rise to a smile effect, the prices obtained for different strikes for the case study above have been converted to the co~esponding implied volatilities, and plotted as shown in Figure 8.10 for the ~aximum jump sizes of 3%, 5%, 7% and 9%. Three features are interesting: first, that the increase in implied volatility is not constant across strikes, i.e. the introduction of jumps has given rise to smiles; second, that the increase in this smile is much greater in going from a maxi mu^ jump size of 3% to 5% than from 5% to 7%, and similarly for the move from 7 %to 9%; and third, that the steepness of the smile (i.e. the slope of the implied volatility curve as a function of strike) increases significantly in moving from 3% to 5%, much less from 5% to 7% and almost not at all from 7% to 9%. Also in these respects, therefore, it seems fair to say that, reassuringly, the model seems to depend less and less on the arbitrary value d as the maximum jump size increases. It is also interesting to observe the behaviour of the smile, and of its steepness in particular, as one varies the final maturity of the option. To this effect the calculations were carried out for maturities from 0.25 month to two months at regular intervals of 0.25 month, and the results are shown in Figures 8.1 1 and as the maturity increases, 8.12. Asis apparent, the smile becomes flatter and flatter in good qualitative accord with what is observed in the FX and equity markets. The left part of Figure 8.1 1 could also be profitably compared with the left part of Figure 4.2 in Chapter 4, which shows a market smile surface for the FTSE index: the qualitative simila~ty is quite apparent.
29.00% 1
. 28.00~0
27.OOYo --
22.00% 95 90
85
100
105
trike
110 120 115
.IO The implied volatilities obtained for the case study above and jumps of 3% (ImplVoll), 5% (ImpIVol~), 7% (ImplVol3) and9% (lmplVol4) (128steps)
l'?$8
0
c??
' ? -3
oj
0 0
g g
9
20.00% J
85
90
95
100
105
110
115
trike
Figure 8.12 Implied volatility curves. Comparison of smile curves for different maturities obtained with different maximum jump sizes: maturities of 0.25 and two months of 3% and 6% and maximum jump sizes
Clearly, given the symmetric nature of the jump (at least in log space) the resulting smile is also similarly symmetric; this feature is approximately displayed (on average) by FX smiles, but certainly not by equity smiles. It is possible, however, to introduce in the nume~cal procedure an asymmetric maximum jump size, in the hope that this might give rise to a skewed smile. More precisely, one can replace the condition n2 F ( j , d2) with the two conditions
which, in turn, imply that
With the model thus modified, the same procedure described above was followed, and the co~esponding implied volatilities were obtained for a variety of maturities. Some typical results are shown in Figures 8.13 and 8.14, where, in the up jump, possible values were collected from all reachable states up to five steps ahead, but the d ~ w njump was constrained to reach no further than one, two, three, four or five steps. These different cases are denoted as the (1,5), (2, S), (3, S), (4,5) and ( 5 , 5 ) cases, respectively. Once again, one can note by comparing the two figures that the smile becomes shallower as the maturity increases from 0.25 month to two months; but, more
85
90
95
100
105
110
115
Strike
w e 8.13 The smile surface obtained by several different possible combinations of up and down jumps, from (1,s) (Seriesl) to the symmetric case of (5,5) (Series5). The maturity was two months and all the other inputs wereas per the case study
28.00% 27 .OO% 26.00%

25.00%
24.00% 23.00% 22.00% 21.OO% 20~00~0

19.00% 4
Seriesl Series2 Series3 Series4 Series5
85
90
95
100 trjk~
105
110
115
The smile surface obtained by several different possible c o ~ b i n a t i o n s of up and down jumps, from (1,5) (Seriesl) to the s y ~ m e t r i c case of (5,5) (Series5). The the other inputs wereas per the case study maturity was 0.25 month and all
interestingly, the shape of the smile is now markedly a s y m ~ e t ~and c , whether the slope in moving from at-the-money to in-the-money calls is positive or negative depends onthenumber of allowable down steps. Looked at in this light, the results indicate that the asymmetric BJN jump model could also convincingly describe the situation encountered in the case of the FX smile, the steepness and shape of which change frequently over time, ~ u m m ~ i s i n the g : numerical implementation of the BJN approach has shown that
the method has good convergence properties; the dependence on the arbitrary maximum jump size, d, becomes smaller and smaller as d increases; the approach naturally gives rise to smiley implied volatilities; the slope of the implied volatility curve depends less and less on the maximum jump size as the latter increases, and decreases with increasing maturity, as observed in the FX and equity markets; the resulting smile obtainable from the standard BJN approach is approximately symmetric; this smile can be turned into an asymmetric smile by imposing a different maximum size for the up and down jumps; after carrying out this modi~cation, there exist combinations of maximum up and down jumps that give rise to asymmetric smiles; with asymmetric smiles, the slope of the smile curve moving from at-themoney calls to in-the-money calls can be positive or negative, depending on the choice of d,, and &own.
The results reported above indicate that the approach proposed by BJN (a~though not explicitly created for this purpose) can account in a simple and convincing way for both symmetric and asymmetric smiles. Given the nature of the mechanism responsible for the smile, the BJN approach would produce the type of smile previously described as floating. A very interesting by-product of the discussion so far is that the Crouhy-Galai (1995) paradox mentioned in Chapter 2, Section 2.5, is automatically resolved. The dependence of the optimal hedging strategy on both the instantaneous and the residual average volatility occurs, and becomes increasingly noticeable, when the binomial tree becomes coarser and coarser. But a jump can be distinguished from a Brownian increment in a discrete binomial setting only by making the spacing smaller and smaller and by tracking the different scaling properties of Poisson jumps and Brownian increments with the magnitude of the time step. If the tree is of fixed time step, and, more importantly, if it is coarse, then there is little meaning in distinguishing between the movement induced by the current instantaneous volatility and a true jump. The possible problems with the BJN approach are two-fold: the first, at a conceptual level, is their condition that the quadratic variation should be known at the start. A s mentioned above, this is admittedly no stronger a condition than requiring that the constant (or deterministic) volatility of a Brownian diffusion should be known a priori. This condition, however, is nomally enforced in a purely diffusive setting, and, in this case, as long as sufficiently frequent trading is allowed and the volatility is deterministic and exactly known, there is no such thing as a luckypath (see the discussion in Chapter 2, Section 2.3): over a
finite period of time, if the volatility generating the evolution of the process is known a priori, we can rest assured that the difference between the quantity (In Sj+l In Sj)2 and the average variance, a(u)2du, can be made as small as desired, As for the requirement, again in the purely diffusive case, that the volatility should be exactly known a priori, it is less taxing than it might ~ ~ ~appear, ~ a given the possibility of entering a kappa hedging strategy that can neu~alise the trader to first order against changes in the overall level of the volatility, i.e. in the realised sample quadratic variation (in this context, volatility). If jumps can occur, however, the situation changes radically, because, given a finite trading horizon, the experienced quadratic variation can in reality change dramatically according to whether a jump has, or has not, occurred before expiry. Increasing the trading frequency does not help in this respect, nor does perfect knowledge of the d e t e ~ i n i s ~ volatility. c For the condition on the total variance to hold, one has to assume that the diffusivevolatility out to maturitywill be higher if no jumps occur, and correspondingly lower if one or more do take place, and in such a perfect way that j=O,N-I (ln Sj+1 In = w2 in all cases. Note that now kappa hedging the exposure is not that trivial at all. Furthermore, irrespective of the trading frequency, and with the caveat above, luckyand unlucky paths now do exist. The second problem with the BJN approach is that it is extremely difficult to extend the computational approach to the case of non-plain-vanilla options: we emphasised at the very beginning the assumption of the existence of a market in ~ o ~In the a different ~ ~ states ~ sampled . by the tree, in fact, one finds the value of the same forward contract, not of the spot price. Note that there is no discounting from one step to the other: indeed, not only does a parent node ~ o ~ m u n i c a t e with states with possibly different values of the residual volatility, V i , but, even when two connected states have the same residual volatility, this does not mean that the c o ~ e s ~ o n d i events ng have t&en place at the same time. Therefore, if one used a spot rather than a forward process, one would not know how to discount money from connected nodes. It is therefore far from obvious how to extend the algorithmic approach to cases where several price-sensitive events occur, or where compound options have to be evaluated. Research is currently under way in this direction. Notwit~standingthis criticism, the BJN approach provides what I believe is robably one of the most c o n v i n c i n g ~ n ~ desc~ptions ~ c ~ ~ ~ of the origin of smiles in equities and FX, and allows the incorporation in its formalism of a variety of ~racticallyimportant features, such as finite re-hedging and market incompleteness. If implemented in the asym~etric-jump version it could not only explain the qualitative features of the observed equity smile, but also its essentially floating nature. Monte Carlo and bushy-tree extensions are being explored at the time of this writing.
xj=O,N-l
This chapter explores the role of mean reversion in the context of interest-rate models by drawing a sharp distinction between mean reversion in the real and in the ~sk-adjusted worlds. In this sense the discussion to be found in this chapter is a natural continuation of the arguments presented in Chapter 2, with which it shares several common points, but from which it also differs in fundamentally important ways. In particular, three i~portant points will be made: 1. much as in the equityEX case, the nature of the drift of the state va~able(s) in the real (econometric) world is irrelevant to interest-rate option pricin 2. unlike the equity/FX case, however, in the risk-adjusted dynamics of the (non-traded) spot rate the drift can, to some extent, be specified by the user without necessarily violating the require~entof no arbitrage; in particular, if one so wished, a mean-reve~ing, ~sk-neutral drift could be imparted to the short rate; 3. the effect of this mean-reverting component is to make it possible, at least for suitable choices of the parameters, for a s~o~-rate-based interest-rate model to price market inst~ments such as caplets correctly and, at the same time, to produce an approximately ti~e-homo~eneous evolution of the term structure of volatilities. The discussion will start with the analysis of a somewhat pathological case of mean reversion (i.e. the one encountered in the case of the B l a c ~ - D e ~ a n - T o y (BDT) (1990) model); en route, an often-mentioned paradox encountered in the algorithmic const~ction of the BDT model will be resolved. It must be stressed,
however, that the main purpose of the chapter is not so much to highlight the shortcomings of the BDT approach, but, more importantly, to stress the close link between mean reversion and the evolution of the volatility of the short rate. This observation will then be extended to more general, but still short-rate-based, modelling approaches, and it will be shown how the s h o r t c o ~ n g s of the BDT approach can be avoided. The analysis of interest-rate models based on the evolution of the short rate will not be pursued beyond this chapter, and the re~aining part of the book will deal with the forward-rate-based Brace-Gat~ek-Musiela (BGM) approach. Whilst, in principle, one could still think in terms of mean reversion also in the forward-rate context, it will be shown in Chapters 10 and l1 that, when one is dealing with forward rates, the requirements ensuring ti~e-homogeneityof the volatility structure can be much more readily cast in terms of a time-dependent instantaneous volati~ity. efore tackling these issues in detail, however, itis useful to give a first e~planatio of ~ the reason why mean reversion matters at all in the case of shortrate-based interest-rate models. If the question seems paradoxical, recall that we have shown in Chapters 1 and 2 that the real-world mean reversion, if at all present, was totally irrelevant in so far as the pricing of options was concerned; and that it could not enter, under penalty of arbitrage, the d y n a ~ c s of stock prices or FX rate in the risk-neutral (pricing) measure. Mean reversion in the ~sk-neutral evolution of the short rate can, however, to some extent be specified by the model user, and does have a profound impact on the evolution of such quantities as the term structure of volatilities. The short (and somewhat cryptic) answer to the puzzle is that the short rate is not a traded asset. A more convincing explanation runs along the following lines. Let P be the price of a generic instrument (as long as it is a traded asset). Since we are a s s u ~ n g that the dynamics of the yield curve is driven by the short rate, the price of this asset will be a function of the short rate itself. We can therefore process for the short rate write P = P(r).2 Let us assume that the ris~-adjusted is indeed mean~reverting:
Extensions to short-rate models where more than one state variable drive the dynamics of the yield curve are treated in Rebonato (1997).A whole class of log-normal, two-factor models is introduced of the consol are the underlying factors. For this in that paper, where the short rate and the yield class of models a ~ean-reverting behaviour is assigned to both factors, and the same conclusions about the time ho~ogeneityof the resulti~gterm structure of vo~ati~ities arrived at in this chapter hold true. See also Chapter 16 of Rebonato (1998a).
eversio~ in l~terest-
Intuitively, we might expect that the mean-revering nature of the risk-neutral dynamics for r should enter the drift of P, and therefore possibly affect its variance, but not its volatility. ut we h o w from Chapter 2 that what matters for pricing options on assets is th volatility, rather than the variance, of the underlying. What is not obvious, therefore, is whether andhow the dr$t of the underlying factor (the short rate) can alter the v a l ~ t i l of i ~the traded i n s t ~ m e nP. t In order to see more clearly into this matter, we can begin by using Ito's to obtain the dete~inistic part, pp,of the Stochastic Differential Equation evolution of P(r, t): more precisely, we want to calculate p p and dP(t) = p p dt up ddt) The calculation is straightforw~d, and gives
However, since P denotes the price of a traded asset, we also know that its risk-neutral drift must be equal to the risk-less rate itself pp = rP Therefore, combining the two above equations, one can write:
-+-,u,+"--a
ap
at
ap ar
1 a2P 2 ar2
'
=rP
a(k r) We also know, however, that the term appears in the contribution that Ito's l e ~ m makes a to the stochastic part, up,of the process of P(r, t): dP Gp = ar Or Therefore, combining the equation obtained above for the derivative with the = a(k r), one finally obtains for expression for the drift of the short rate r, the volatili~ of P:
a(k r) The purpose of these algebraic manipulations is to show that the drift of r (and its mean reversion in particul~) affects the v o l ~ t i l of i ~ P, not its risk-neutral drift
(which is beyond the modellers control, given the requirement of no arbitrage).3 The effect of assigning a particular drift to the short rate (in the risk-neutral world) is therefore to modify the volatili~ of the traded asset P , and not at all its fwe were to value an option on P , then this mean-reversion(risk~neutral)drift. I dependent volatility o-p would clearly have a profound impact on the pricing; as a consequence, assigni~g a mean-reverting behaviour to the risk-neutral drift of the driving variable, r, directly and strongly affects the very quantity (i.e. the volatility, as opposed to the variance) that matters for option pricing. The following sections will explore in greater depth this interplay between mean reversion in the risk-neutral world and the evolution of ~uantitiesthat directly affect option pricing.
Of the several one-factor models used for pricing interest rate options, the Black, Derman and Toy (1990) is one of the best known, and is still rather widely used, at least for the evaluation of B e r ~ u d a nswaptions (more widely, in m y opinion, than its properties should wa~ant). Despite its conceptual and practical shortco~ngs, in what follows we shall discuss some features of the BDT model for the insight this analysis can bring to the topic of mean reversion. ~ m o n g s the t most appealing features of the BDT approach are: the capability to price exactly an arbitrary set of received market discount bonds; the log-normal distribution of the short rate, which makes calibration to caplet prices particularly straightforward; and, last but not least, the ease of i m ~ l e ~ e n t ~ t i o n . The first feature (exact pricing of the yield curve) is shared by a variety of r the Hull and White. The second no-arbitrage models, such as the Ho and Lee o (log-normal distribution of rates) is also shared by the Black and Karasinski model (which, unlike the BDT model, displays true mean reversion). Only the BDT approach, however, allows log-normal rates and calibration to caplet prices (in the absence of smile effects) that can be accomplished almost by inspection. From a practical point of view, it is probably only because of this feature that the model is still used by some houses in real-life pricing applications that require ~ a c ~ w ainduction, rd rather than making use of a conceptually superior, but more cumbersome to calibrate, approach such as the Black and Kara~inski.~
The reader might not like the fact that, in the expression for the volatility of P, the reversion speed a enters at the denominator, It would seem that this volatility would he infinite if the reversion speed were exactly zero. In reality, if a = 0, then the term aP/3r would multiply a zero short-rate drift in Equation (9.1), and could not be solved for. Therefore, for the derivation aboveto hold, one .0. must impose that a $ The ?nodern (i.e. B~M-based) approach has not won general acceptance for problems such as Bermudan swaptions because of the difficulties in mapping the non-Markovian process implied by log-noma1 forward or swap rates onto areco~bining tree. Several approximate ways to circumvent the problem have been proposed, but none is without some sllortco~ings.
As is shown in what follows, this ease of calibration to caplet prices is, at the same time, the blessing and the bane of the BDT model, and directly stems from the inflexible specification of the reversion speed, which, as shown below, is completely determined by the future behaviour of the short-rate volatility. In this section I therefore intend to highlight the intimate connection between the ease of calibration to cap prices and the pa~icularlink between the reversion speed and the logarithmic derivative of the short-rate volatility. Let us start by considering a generic Brownian diffusion of the form
d In r(t) = 0(t) dt
ainst(t)
dz(t)
(9.2)
drift, t i c dz the incre~ent of a Brownian process and a i n s t ( t ) with 0(t) a ~ e t e ~ i ~ i s an instantaneous volatility. We have seen in Chapter 2 that, if the drift is deterministic, then the unconditional variance of the variable r out to time T is simply given by
We have also pointed out that if the drift is not purely deterministic, but contains the (stochastic) state variable, then the unconditional variance is no longer simply linked to the time integral of the square of the instantaneous volatility. In particprocess of the form ular, for a mean-reve~ing d In r(t) = [0(t)
+ ~ ( + ( t) In r(t))ldt +
ainst(t)
dz(t)
(9.4)
( g ) , and 0(t) a ~ e t e ~ i n i s tdrift i c compo(with reversion speed k , reversion level + nent) the unconditional variance will, in general, depend on the reversion speed. It is well known from the literature (see, for example Rebonato (1998a)) that the continuous-time equivalent of the BDT model can be written as
d In r(t) = [0(t) f(t)(+(t) In 4 t ) ) l dt with

f(t)
ainst(t)
dz(t)
(9.5)
a In a(t) at
and both 0(t) and a(t)deterministic functions of time. From expression (9.6) one can formally see that it is only in the presence of a time-decaying short-rate volatility -( 0) that the resulting reversion speed (-f)is positive and the model displays mean reversion. (See Section 9.5 for a more intuitive illustration of this observation.) Equation (9.5) can be re-written as a diffusion of the general form: d In r(t) = [a(t)(b(t) -In r(t))]dt
+ a(t)dz(t)
(9.7)
where a(t),b(t) and a($) are ~ e t e r ~ i n i s tfunctions ic of time. The SDE (9.20) can easily be solved, and the variance calculated, to give var [In r(T)] = exp 1--2
LT 1
a(s)ds]
exp
[2
a(s) ds] dt
(9.8)
Note that, by Equation (9.8), the variance of the loga~thm of the short rate depends on the reversion speed. This however seems to create a paradox. We shall show below, in fact, that the following three Facts are true:
. The unconditional variance of the short rate in the not depend on the instantaneous volatility from time 0 to time T At (as one would have been led to expect from Equation (9.3)).
The unconditional variance does not depend on the reversion speed -f (as one might have surrnised from Equation (9.5)).
. One can show (see below) that

where 0 2 ( N A t )is the (square of) the instantaneous short-rate volatility at time T = IVAt, i.e. the total unconditional variance of the short rate from time 0 to time T only depends on the value of the instantaneous volatility at time T.
ow can these apparently contradictory facts be reconciled? How can it be, in other words, that the total variance of r neither depends on the full integral of the (square of the) short rate, nor on the reversion speed? Note that Equation (9.9) is crucially important for calibration purposes: given that the arbitrage-free drift of the forward rates in the Black pricing measure is zero, the market Black implied volatilities ( ~ B ~ ~ provide ~ ~ ( Tdirect ) ) information about the unconditional variance of the relevant forward rates (spot rates at expiry). From the quoted implied Black volatilities of caplets of different expiries the user can obtain almost exactly their exact BDT price by assig~ing a timede~endent short-rate volatility matching the implied Black volatilities. In other words, if one sets a(lVht) = OB~,,,(T), N A t = T
one can rest assured that, for practical purposes, the caplet will be almost exactly pri~ed.~ It is quite well known amongst practitioners that this is the case. What is not generally appreciated is how this can be, since Equations (9.3) and (9.5) would in general suggest that both the instantaneous short-rate volatility from time 0 to time T and the reversion speed - f f should affect the unconditional variance from time 0 to time T. I shall therefore begin to show that Fact 3 is true, i.e. that the empirically known result mentioned above regarding the unconditional variance is indeed correct.
Note that thereis a minor sleight of hand in the reasoning: we are equating variances (of forward and spot rates) rather than volatilities. Since the state variable appears in the drift for the short rate this is not correct (as discussed in Chapter 2). It so happens that, for typical values of the BDT mean reversion, this caveat has a very minor effect.
eve$~ion in lnterest-
A calibrated BDT lattice is fully described by (see the original paper BDT (1990)
for a detailed description of the procedure) a vector r = {ria} (i = 0, k ) whose elements are the lowest values of the short rate at time step i , and a vector CT = {a;} (i = 0, k ) , whose elements are the volatilities of the short rate from time step i to time step i 1.
Every rate rij, in fact, can be obtained as rij = ri0 exp[2ujjJAt]. (At is, as usual, the time step in years.) Let us now define (see Figure 9.1) k random variables Y l , Y2, Y k by
* * * : ,
Yk=
1) At 1 if an up move occurs at time ( k 1) At 0 if a down move occurs at time (k Y3
For instance, for the path highlighted in Figure 9.1, y1 = 0, y2 = 1, and y4 = 0. Let us also assume, as is true in the BDT case, that and 1. the variables yj are inde~endent, 2. the probability P[yk = l ] = P[yk = 01 =
=0
i.
~ i ~9 . 1 ~ rValues e assumed by the random variables down-u~-down-down path hi~hli~hted
y1,
y2, y 3
and y4 for the
The variable x k = xj=l,k yjtherefore gives the level of the short rate at time kAt, and the value of the short rate at time k 4 t in the state labelled by X k is given by rk,X(k) = rk0 exp[2aiXkdAt](9.10) The next task is thento evaluate the expectation E[lnrk,X(kl] and variance var[ln rk,X(k)] of the logarithm of this quantity. To this effect one should note that the dist~bution of x k is simply given by the well-known Bernoulli (binomial) dist~bution P[xk = j ] = C;/Zk (9.11) with
c; =
k! (k j)!j!
(9.12)
Therefore We are now in a position to evaluate E[ln rk,X(k)] (abbreviated as Q):

j=O,k
(9.14)
j=l,k Given, however, the definition of C ; ,
After substituting in (9.14) one obtains
(9.15)
Similarly, for the variance E[ (ln Q)] =
j=O,k
ikC;(ln
rk
+ 2akdAtj)
j=l,k
ut the last term is simply equal to (9.17) and therefore the last s u ~ a t i o n adds up to (9.18)
everion in ~ n t e r ~ t Capletpricesperunitprincipaland ATM strikes for the GBP sterling curve of expiries reportedontheleft-handcolumn,asevaluated using the Black model (column Black), and BDT the model calibrated as described in the text (column B DT) Sterling curve (November 1995) Expiry 01 -NOv-95 31-jan-96 01 -May-96 31-JuI-96 31 -0ct-96 30-Jan-97 01 -May-97 01 -Aug-97 31 -0ct-97 30-jan-98 02-May-98 01-Au~-98 31 -0ct-98 Black 0.000443 0.000773 0.001 148 0.001 559 0.002002 0.002422 0.002746 0.003024 0.003265 0.003471 0.003449 0.003406 BDT 0.000431 0.000757 0.001 133 0.001 548 0.001 994 0.002416 0.002742 0.003020 0.003263 0.003471 0.003452 0.003411
Putting the pieces together the unconditio~al variance is given by
= (hrk)2 = GikAt
+ 2k(sk(h r k ) d h t + G ~ ~ ~+k1)( k ( h r k +k ~ k d h t ) 2
(9.19)
Equation (9.19) therefore shows that Fact 3 is indeed true: despite the fact that the continuous-time limit of the model displays both mean reversion and a nonconstant short-rate volatility, the unconditional variance of the l o ~ a ~ of t ~the m T model only depends on the final instantaneo~s volatilit~ pression (9.19) therefore formally validates the empirical procedure, well known amongst practitioners, calibrate to caplet market prices, Table 9.1 shows the results of calibrating the DT tree using the Black implied vol~tilities as direct input to Equation (9.9). Note, however, that Fact 3 has been proven to be true, but the paradox is still unexplained. With this proviso in mind, let us move to Fact 2.
The above derivation has shown that, in discrete time, the unconditional variance of the short rate is indeed given by expression (9.9). What is still not a ~ p ~ e ~
however, is ~ h the y reversion speed andlor the instantaneous short-rate volatility from time 0 to time T At does not appear in the equation. To see why this is the case it is more profitable to work inthe continuous-time equivalent of the BDT model (Equation (9,5)), and to consider again the general expression for the variance of a mean-reverting dif~usion: d In r(t) = [a(t)(b(t) In r(t))]dt var[ln r(T)] = exp -2 l ' a ( s ) ds!
A s Equation (9.21)
l '
+ a(t) dz(t)
exp
[L
~
(9.20)
a(s) ds! dt
(9.21)
the unconditional variance of the logarithm of the short rate out to time T does indeed in general depend both on the reversion speed and on the values of the instantaneous volatility a(t) from time 0 to time T. This result is completely general, but, if one specialises it to the case of DT model, one should note that (i) a(t) = -*f', and (ii) f (t) = In a(t). direct substitution the unconditional variance of the log of the short rate out to time T therefore becomes
a use ~ of (ii)~(f(t) g = lna(t)) in Equation (924, one can immediately verify that, in the DT case, the unconditional variance is indeed simply given by
(9.23)
~e have therefore reached an interesting conclusio~~ given any ~ean-reverting ~ r o c e sfor s which the reversion speed is e ~ a c t l y e ~to ua the l ~egativ ofthe ~ logarith~ic derivative o f the instantaneous volatili~ with r e s ~ e to ~ tt i ~ (i.e. e a(t) = - ~ I ~ a ( t neither ) ~ ~ t the ~ r~versionspeed nor the past i ~ s t a n t a n e ~ ~ s v o ~ enter the expression for the uncon~itional va~iance, which only ~ e ~ e n on d s the instant~neo~s short-rate volatili~ at the ~ n t a ~ ~~ e . This observation fully explains the DT paradox, and sheds light on the reason why a more satisfactory model like the Black and ~ a r a s i n s (which ~ displays 'true9 mean reversion) is considerably more difficult to calibrate. The reader might however still be puzzled as to 'what went wrong' from the a l g o ~ t ~ m i c point of view: by the end of the BDT constr~ction one has used all the degrees of freedom at one's disposal and all of today's market inputs (bond andcaplet prices) have been correctly recovered. How could one have done anything di~erently
and still retained a log-normal distribution for the short rate? We address this question in the next section.
In order to gain a deeper understanding of the nature of mean reversion in shortrate models, let us look at the algorithmicorigin of the result just obtained. More precisely, let us consider the BDT c o n s ~ c t i o n over two time steps in the cases of a steeply decreasing and a steeply increasing short-rate volatility function. Figures 9.2-9.5 represent the first two steps of a non-recombining (bushy) tree and of a BDT tree with the same time-dependent volatility, and with the same probabilities (112) for both jumps. On the y-axis one can read the loga~thm of the short rate. Given that an up state is linked to its corresponding down state by the relationship rup = rdown exp(2ofdAt)
and that the volatility can depend on the time step, but not on the state, all the
up and down logarithms have the same separation (in log space) at a given
Figure 9.2
A bushy-tree construction for the case of decreasing volatility
down
.3 The corresponding BDT construction for the same decreasing volatility
~ownJup
upJdown
down/down
A bushy-tree construction for thecase of increasing volatility
everion in l~teret-
downlup up/down
down/down
Figure 9.5 The corresponding BDT construction for the same increasing volatility
downldown
ure 9.6 Same as Figure 9.5 but with a purely t i ~ ~ - ~ e p e n d(negative) ent drift (the term e(t) in Equation (9.4))
time step; therefore, both in the bushy and in the BDT tree, the y-axis distance between any two states originating from the same node is given by 2 cr,JAt. It is essential to note that in both trees the c o n s t ~ c t i o must ~ recover not only the total unconditional variance from the origin, but also the conditional variance from each node,
From the pairs of corresponding figures (i.e. from Figures 9.2, 9.3 and Figures 9.4, 9.5) one can immediately appreciate that, in the bushy case, any drift could have been assigned to the short rate, and the const~ctionwould still have been possible. Looking at Figure 9.3, however, which refers to the case of sharply decreasing volatility in the BDT construction, one can see that the only way to ensure that at each node the condition rup = rdown e x p ( ~ ~is ~ ~ fullfilled, and that the tree recombines is to push the two up nodes (labelled uplupandupldown in the figures) towards the two down nodes (labelled down/down and down/up) in Figure 9.3. Simila~ly, looking at Figure 9.S one can see that, if the volatility of the short rate is increasing, the up nodes must bemoved up even further, and the downnodes further down. Note that all the nodes can be movedup or downby the same amount (i.e. by adding a purely time-dependent drift) without affecting the feasibility of the construction, as can be appreciated by comparing Figures 9.5 and 9.6. They cannot be moved in a state-dependent way, however, without compro~ising the BDT recombining const~ction. It is clear from the figures that the moving down of the up nodes and up of the down nodes shown in Figure 9.3 therefore introduces a mean-reverting component into the process for the short rate. Si~ilarly,the pus~ingapart of the nodes shown in Figure 9.5, necessary to ensure recombination, introduces a mean-fleeing component. In the BDT case the recombination requirement therefore strongly limits the possible drifts, over and above the drift imposed by recombination, that can be specified by the user: in particular, this extra drift can at most be time-dependent (the term 0(t) in E~uation (9.4.) algorithmically accounted for by the const~ction in Figure 9.6), and, therefore, cannot introduce I). true mean reversion (since this would require the state va~able, From these considerations it can be seen from yet another angle why a mean reversion (of sorts) can only occur with the BDT algorithm if the volatility is time-dependent (and, more specifically, decaying). Note that the negative aspects in~oduced by the procedure are more insidious than the usual limitations of onefactor or low-dimensionality models: the recombining-lattice geomet~ i~troduce an inextricable link between its reversion speed and the (logarithmic) derivative of the short-rate volatility, The time-decaying volatility needed in the BDT model in order to contain an excessive dispersion of rates does succeed in obtain in^ an ~ ~ c o n ~ i t i distribution ~na2 of rates consistent with the one implied by the cap mean reversion is absent from the model, market. Since an explicit deter~nistic however, for any non-decaying behaviour of the short-rate volatility this can only be obtained at the expense of a lower and l o w e r ~ ~volatility. t ~ ~ e This undesirabl~ feature can have a limited impact for relatively~short-maturity options, but must always be borne in mind by users who extend their analys~s well beyond the common volatility hump observed in most cap markets. Unfortunately, the first sections of this chapter have shown that it isimpossible to remove this undesirable feature and to retain at the same time the original ease of calibration.
everion in l~teret-
The previous section has highlighted that the l i ~ t a t i o n s of the BDT approach discussed above directly stem from the algorithmic prescription for its construction: it is not necessary, for instance, to construct a computational lattice by imposing identical probabilities to the up and down jumps, Alternatively, a procedure based on a tree-branching lattice could be introduced. In either case, the construction and the calibration become more cumbersome, but a very significant advantage is reaped: in computational terms, this can be seen as the gaining of one degree of freedom (e.g. the non-equal probabilities); infinancial and economic terms, as the possibility to specify a true (i.e. volatility-independent) mean reversion. As we saw in the equity/FK case, mean~reversionin the real ~ o r has l ~ no direct bearing on option pricing, since the transfo~ationfrom the real world to the pricing measure effects a drift transformation that completely scrambles whatever dynamics might have been observed in the real world. The important distinction to keep inmind is that, in the interest-rate case, one is typically evolving non-traded quantities. ~o-arbitrageconsiderations therefore do not directly dictate that these non-traded state variables should have a drift equal to the risk-less rate. hen one is dealing with interest rates, one is therefore at liberty to start with an almost arbitrary functional form for the s~ecification of the drift in the risk-neutral world. In particular, this specification could incorporate a mean-reverting feature. As shown in the introductory section, this can have a direct effect on the pricing of claims depending on the driving factor(s). Despite the fact, however, that one can introduce a mean-reverting tern into the ~ s k ~ n e u t rdynamics al of the short rate, it is not a priori obvious why it should be advantageous to do so. This point is absolutely crucial since, despite its irrelevance for option pricing e j a~ ~ does ~ ~ t r e have ~ a in the real world, mean reversion in the r i ~ ~ - ~~ ~ ~rofound9 if often ~isunderstood,impact on the evolution of rate volatilities, In very qualitative terms, if part of the burden to avoid excessive9 dispersion of the rates is shouldered by a mean-reverting term (rather than by a timedecaying volatility, as in the BDT case), then the volatility can be constant, or mildly varying, and the market prices of caplets can still be recovered by the model. This observation, however, poses another question: why should we want to require that the volatility of the short rate be constant, or mildly varying? Or, enerally: since our information about the real-world mean reversion does not imply anything about the drift of the risk-adjusted dynamics, how does one uess the nature of the risk-adjusted mean reversion? The key to a satisfacto~ answer to both questions lies in the observation that the transformation from the real world to the pricing measure affects (via Cirsanovs theorem) the drift but not the volatility of the financial quantities of interest. In other words, our calibrated model will imply that whatever future volatility
structure we are left with after recovering the caplet market prices is the same as the one that we should expect in the real world. In the light of this observation, one possible (and useful) criterion could therefore be: choose the ~ a r a ~ e t eo r fs the risk-~djusted ~ean-reverting process in such a way that the volatili~ f~n~tio~ n e e ~ to e price ~ the ~ a r ~ e t - t ~ a d e ~ are o~ ~rteiao s~ on s a ~tliy ~ ~ e-ho~ogeneou In s. other words, make sure that the future trading universe looks acceptably similar to the universe we observe today. The ultimate justification for this heuristic criterion is that we often observe term structures of volatilities which have been approximately constant over time (see also the discussion in Chapter 11); therefore, unless we have specific views to the contrary, it is reasonable to assume that these feature will remain similarly (appro~imately) constant in the future. ~ e e ~ l eto ss say, if the trader were to have views that the future will not be like the past, then she should use a model that reflects, in the evolution of the relevant volatilities, this particul~ view. This issue is tackled in depth in Chapter 11, where it is argued that one of the main reasons why the modern (BCM) approach is so much more preferable to the traditional (short-rate based) framework is because the former gives the trader a direct way of controlling the evolution of those q u ~ t i t i e s (such as the term structure of volatilities) which have a direct impact on the future re-hedging costs. In the meantime, however, it is worthwhile noting that imposing time homogeneity for some financial quantities can have much wider applicability than the purely interest-rate case: similar conclusions can be drawn, in fact, in the case of the implied volatility smile. We emphasised in that context that reproducing todays smile is only (a small) part of the overall task of a model. Just as important is the models prediction of the f ~ t ~ smiles. re If enough liquid instruments were traded in the market, and exactly priced by a model so as to capture not only the effect of todays smile or of todays cap volatility term st~ucture but also of the evolution of these quantities, then the traders task would truly be one of pure arbitrage. In the case of interest-rate options we shall show in Chapter 10 that the missing liquid instruments are the serial options. Similarly, in the equity/FX world forward-setting options for a great variety of maturities and reset times would have to be lai in-vanilla rather than exotic products. If that were the case, the exotic traderS task would be considerably simpler (or, more likely, her job would disappear): she would not have to worry whether the market is pricing in an unreasonable future evolution for the smile or for the term structure of caplet volatility: she could simply engage in the transactions suggested by her model (calibrated to all the products), resting virtually assured that she could lock in any price discrepancy. True model arbitrages, are, however, exceedingly rare, and, in reality, a given strategy, for which the model suggests a positive present value today, will only prove pro~table if the world were to indeed evolve as the model predicts (or at least not too dissimilarly). We have already analysed two such cases when we dealt with forward-setting options and continuous double barriers in the presence of smiles. We showed
eversion in Interest-
that the re-hedging costs (or the unwinding costs)-and hence the price of the structure-were a function of the assumed sticky or floating nature of the smile. We shall analyse a similar situation in the interest-rate case in the next chapter, by indicating what instruments should be traded in order to lock in any price discrepancies in the c a p l e ~ s ~ a p t i o market. n In the light of this discussion, one can leave the short-rate models with the observation that, apart from all their other shortcomings, they provide the user with an extremely blunt instrument when it comes to making the model produce a pre-chosen future evolution for the term structure of volatilities. Some models, as we have seen, just cannot accomplish this goal; others (like the Black and Karasinski or the Hull and White) have the potential to deliver this result, but offer a very inefficient way to produce the desired effect: I have always felt that driving the term structure of volatilities from the short-rate end is like push in^, rather than pulling, a rope. The next chapter will show how,by means of a type of approach, a more constructive interaction with a rope can be established.
From (9.20) one can write
d In r(t) a(t)In r (t)dt = a(t)b(t) dt

This implies that
+ o(t)dz(t)
But the quantity I n r ( T ) e x p [ ~ ~ a(s)ds] can be written as
and. therefore
"_
"
Recalling that, for any deter~inistic function f ( t ) ,
it then follows that

var(1n r(T)] = E[(ln
(E[ln r(T)])2
Since the early 1980s until the mid 1990s several interest-rate models were vying for the attention of practitioners and academics, and no clear market consensus had emerged as to the favourite or most widely used interest-rate model. Tradeoffs always bad to be made between ease of calibration (or, more often than not, lack thereof), availability of closed-form solutions, realism of distrib~tio~al assumptions, etc. The common feature of almost all these models was that the yield curve was a s s u ~ e d to be driven by the unobservable short rate (plus, sometimes, another equally ~nobservable variable, such as the yield of the consols, or the variance of the short rate). In all cases the user bad to perform, explicitly o r implicitly, a tra~sfor~ation between the input values of the unobservable state variables, and those quantities, such as volatilities of forward or swap rates, that the user could observe in the market. The model, in this respect, acted as the black box that transformed what the user could input (e.g. volatility of the short rate) into what the user would have liked to be able to input (e.g. caplet or swap volatilities).
Part of thematerialpresentedinthischapterwasoriginallypublishedin:TheRiskTamers, a ~ u t u r and ~ s QTC ~~~~1~ supplement, June 1998; The ~ oo f C~o ~ ~ ~~ ~ a a final^^^, t i o ~ n vol. ~ ~ 2, nos. 3 & L 4, 1999; and Risk, March, 1999 (Rebonato (1998b, 1999a, 1999~)). Useful discussions with Mike Sherring are gratefully acknowledged. Thanks are also due to Soraya Kazziah for performing sorne calculations. This chapter contains material that requires a slightly higher familiarity with linear algebra than other parts of the book.Any text dealing with matrix algebra, and the principal compoflents teChniques can be used as a reference (see, for example, Chatfield and Collins (1988a). Chapter 3 of Rebonato (1998a) provides a derivatioll of the results regarding Principal Components Analysis stated without proof in this chapter.
The more recent Heath-Ja~ow-Morton ( ~ J M (l989) ) approach radically changed the set of driving state variables by focusing attention on the infinite number of (instantaneous) forward rates that describe the yield curve. These instantaneous forward rates were still, strictly speaking, unobservable, but they were somewhat closer to what the user had direct access to (i.e. forward L1 rates). In its more general form, however, the HJM modelis hardly more user-friendly, when it comes to calibration of the model to market data, than traditional shortrate-based models. Since the HJM approach in its most general, and useful, form ideally lent itself to Monte Carlo simulation, but to virtually no other computational techni~ue, the calibration problem remained virtually unresolved. A number of papers appeared proposing more or less cumbersome methodologies to specify the volatility functions in such a way that cap prices could be approximated, F u ~ h e ~ o rthe e , fact that technical conditions did not allow the iastantaneo~s forward rates to be l o g - n o ~ a l l y distributed (the modelwould explode) did not make matters any simpler. ~ractitioners decided to avert their eyes from the problem, and pretend that it did not exist; academics proposed several partial and, one has to say, highly arti~cial solutions. ace-Gatarek-~usiela (BGM) (1995) approach brought about (to use but, for once, appropriate tern) a veritable paradigm shift, Now direct market observables (i.e. LIBOR-for~ard or swap-rates) became the building blocks of the new methodology, earning the new approach the title of the market model. The BGM formalism quickly proved to be ideally suited for discrete-look, ndent derivatives products in a multi-factor framework. Acceptance of approach was facilitated by the fact that it shared with the HJM model the fundamental insight, bythen familiar amongst practitioners, that the noarbitrage conditions describing the deterministic part of the evolution of forward rates could be expressed purely in terns of their correlations and volatilities. As a result, the research emphasis became progressively focused on the correct specification of the time-dependent instantaneous volatility of the underlying state variables, i.e. the discrete LIBOR rates. At the same time, it became progressively clear that, in order to achieve decorrelation amongst forward rates, introducing a non-flat volatility for the forward rates provided a more impo~antand more realistic mechanism than invoking a large number of driving factors (see the discussion in Chapter 3). In other words, the attention gradually shifted from the instant~aeo~s to the t e ~ i n a Z correlation amongst rates. In a nutshell, the simultaneous specification of these time-dependent volatilities and correlations became not oae of the problems, but the problem, in the specification of the BGM model. ~alibrating the BGM model simply to the volatilities (and hence either to the caplet or to the European swaption prices, if the forward-rate-based or the swaprate-based implementation is chosen, respectively) is very simple indeed, and, as shown below, can actually be accomplished exactly in many ways. However,
the fact that so many possible solutions exist creates a problem: since each specification of the time-dependent volatilities will give rise to a different degree of t e ~ n a correlation l amongst the forward rates, the question arises as to whether it can be in some way possible to define an optimal simultaneous calibration to cap prices and to the correlation surface. This is the question addressed in this chapter.
The type of problem for which the BGM formalism is perfectly suited is the pricing of discrete-look, path-dependent derivatives in a multi-factor frame~ork, as was analysed in Section 3.4 of Chapter 3. The derivative product to be priced should ideally be discrete-look, because the BGM model evolves a set of discrete LIBOR rates in cQ~ti~UQus time. It should be path-dependent, because the general model is non-Markovian and therefore does not allow mappingonto recombining trees: backward induction is therefore dif~cult with BGM, whilst a forward induction (Monte Carlo) evolution is very easy. Finally, the approach can easily be multi-factor because one does not have to (or, rather, cannot) build recombining trees, which pay an exponential price in the number of state variables (the so-called curse of dimensionality); the computational burden of carrying out a Monte Carlo simulation, on the other hand, grows in an approximately linear fashion with the number of evolved state variables. Examples of securities that can be ideally priced using the combined BGM~onte Carlo approach are therefore trigger swaps (IPS swaps), knock-out caps, one-way floaters, ratchet caps, discrete-sa~pling average-rate caps, For all these applications it is very easy to calibrate the BGM model to caplet volatilities. Why this is the case can be appreciated as follows. Let o-black(Ti) be the implied Black volatility of forward rate4 i of maturity Ti, fi(Ti) (often abbreviated in what follows as fi) and let si(zt, Ti)be the instantaneous volatility at time U for that forward rate. Since we are, for the moment, ignoring the possibility of smiles, the instantaneous and the implied Black volatilities are linked as usual by the following relationship: (10.1)
Foradiscussionofthepricingaspectsofmanyoftheseproducts,andoftheimportance of volatility and correlation in particular, see Chapter 2 of Rebonato (1998a). ~ ~ t a ~t ~ i ~t ~~ exactly ~ ~the i same s , theorems and results derived in what follows apply to the c of forward swap rates. The approach described in Sections 10.2 and 10.3 can therefore be used in order to price European swaptions and fit, in the best possible way, a surface of correlations amongst swap rates.
The evolution of a log-normal forward rate is described by a SDE of the form
whose solution (see Chapter 3) is well known and given by
(10.2) {f } ,U) reflect the no-arbitrage condiIn Equation (10.2) the drifts p i({S}, {p}, tions, and are, in general, a function of the instantaneous volatilities of the forward rates, of the correlation amongst them, and of the forward rates themselves. The notation {S}, {p}, {f } emphasises that the drift can, in general, depend on all the instantaneous volatilities, on all the forward rates and on the full correlation matrix. Once the volatility and correlation functions have been specified, these drifts are therefore uniquely deter~ined, and their functional form and derivation is well known (see, for example, Jamshidian (1997) or Rebonato (l998a)). They are therefore no longer dealt with in this chapter. As for the constants ('loadings') a j k ( U ) , they can be inte~reted as the sensitivities at time U of the jth forward rate to the kth shocks. A total number S of shocks are allowed to perturb the forward rate. If
(with 6ij = 1 if i = j and 0 otherwise), then Equations (10.2) and (10.3) describe the most general o~hogonal l u g - ~ o r ~ forward-rate al s-factor model. Note that if condition (10.3) were not satisfied for a given model (i.e. if E[dzi, dzj] = pij dt), then a rotation of variables could always be carried out so as to ensure orthogonality; the loadings ai,t.(t)have a calendar-time dependence, and are specific to each individual forward rate via the first index. They cannot, however, be of the form aik ( f i , t ) and preserve the log-normal distributional feature for the forward rate f i . Therefore they truly represent the most general specification of an s-factor model consistent with l u ~ " ~ u r ~ forward aZ rates.
As discussed in Chapter 2, calibrating the BCM model to caplet market prices therefore means ensuring that, at each point in time, the square si(U, Ti)' should be equal to the contribution to the instantaneous variance per unit time of the forward rate f i coming from the stochastic term. Let us calculate this contri~ution
ode1
to the instantaneous variance:

2
(10.4)
k=l,s
Therefore
In Equation (10.4) the notation varstoch[.] indicates the component of the variance arising from the Brownian (stochastic) part which, as we know, is the relevant contributor to option pricing. The distinction is not pedantic, because, as we have pointed out above, the drift contains the state variables themselves, and therefore the total variance does not coincide with the quantity vasto&. See (1999b) for a careful discussion of this point. If the relationship above is satisfied for each caplet we can rest assured that the model will recover the cap market prices exactly (within numerical error). A l l that we have to do to ensure that the instantaneous volatility will be exactly recovered by the model is to be certain that the squares of the arbitrary numbers we choose for the loadings, (ajk}, exactly add up to the square of the instantaneo~s volatility itself. If this latter quantity had thenbe chosen in sucha way that its time integral to expiry equalled the square of the implied Black variance ( Q & ~ ~ ~ then the caplet will be correctly priced. How can we ensure that this relationship is satisfied? Let us impose that
a ~ k ( t )= M t ) S i ( %
Ti)
(10.5) (10.6)
with
bik
($1= ~ ~ ~ ( te) ) ik
j=l,k-l
sin(eij(~)), k = 1, S 1
for arbitrary angles O i j . If that is the case, then ( be correctly priced. To see why this is the case, let us begin by calculating the variance of the l o g ~ t h m of the forward rate as var[ln f i ( T i ) ] = E[ln fi(ti)2] E[ln fi(ti)12
76
Define
Then
= 1 Vu, because the coefficients {b}can be recognised as the But "j"k=l,s b i k ( u ) ~ hypersphere. Therefore polar coordinates of a unit-radius (S l)-di~ensional
and
LT
S&,
T)2du = var[ln f(T)] = c&,,(T)~T
Q.E.D.
The special case of S = 2, k = 1 (i.e. a two-factor model) can enhance the intuition. For any forward rate i we can write:
Then
277
where the last equality follows because, for any x,sin2(x) last line is just our condition
k=l.s
+ cos2(x) = 1, But the
In other words, if we choose the quantities bil and bi2 as the sine and cosine of aay a r b i t r a ~ angle, we can rest assured that the caplets will always be exactly priced. For the moment the expression of the condition about the instantaneous volatility interms of arbitrary angles (Equations (10.5)and (10.6)) might seem to provide little, if any, advantage. We shall show later on that the introduction of these new variables (B) is actually computationally very useful, since it allows us to cast a constrained-optimisation problem in terms of an equivalent unconstrained one. Let us now return to the more general multi-factor case, and let us consider any path-dependent option problem such that the expiries and maturities of a set of n forward rates constitute all the dates when p~ce-sensitive events O C C U ~ .As ~ mentioned before, examples of this type of problem could be the discrete trigger events, the discrete sampling times for averages, or the resetting of stochastic strikes for ratchet caps, etc. After each of these price-sensitive events has taken place, the number of forward rates left in the problem is reduced by one. Let then h(i) be the number of forward rates alive at time step i. As discussed in detail in Sections 3.3 and 3.4 of Chapter 3, what any model (im~licitlyor explicitly) produces for this type of problem is a series of discrete covariance elements of the type
If, at each time step, the chosen number of factors is equal to h(i), then the problem of reproducing an arbitrary set of instantaneous volatilities and an ~ b i t r exo~enously ~y assigned correlation matrix can always be trivially accomplished by orthogonalising the above time-dependent covariance matrix. In other words, if we had enough market information about the self-consistent prices of traded instru~ents, we could afford to fit all these prices by ensuring that all the covariance elements above are recovered. A simple example can clarify this very important point.
This restriction is not necessary, and is introduced simply to simplify the notation. A brief reminder: if we start from an IZ x rz conelation matrix (which must be real and symmetric, and we shall assume of rank n ) and orthogonalise it, we can obtain rz distinct eigenvalues, hi, and eigenvectors,ai.The latter are normally called the ~ ~ ~ r co~~orzerzts. z c ~ ~ If~we l build an n x n matrix by juxtaposing these n eigenvectors, and call it matrix A, then one can prove that A Ar (where AT denotesthetransposeof A) is equal to the correlation matrix we started from. See Chapter 2 in Rebonato (1998a) for a discussion of Principal Col~ponents Analysis.
A second importa~t example can further illustrate the importance of the time dependence of the volatilityof each forward rate, namely thecase of serial options presented below,
The examples presented in Sections 10.3 and 10.4 have shown that, in the twolook universe described above, the evolution over two time steps of thetwo forward rates (and, therefore, the pricing of a Z E the discrete-look instruments depending on the realisations of the forward rates at these two times) depends on the distinct elements of one 2 x 2 covariance matrix, and a (degenerate) 1 x 1 matrix:
and (See again Section 3.4, Chapter 3, for a more general discussion.) In market applications, however, the number of price-sensitive events (and, co~espondingly, of forward rates and covariance elements) can easily become prohibitively high: in a quarterly 30-year trigger swap the number (h) of forward rates would be 120, and the number of covariance elements of the order of IO4: there are, in fact, O( lo2)
covariance matrices, each of which contains h( $ h / 2 )distinct elements. The most general description of the l20 variables could be obtained by orthogonalising, at each time step, the covariance matrix describingtheir dynamics, and associatingone o~hogonal factor to each of the eigenvectors. In other words, to retain maximum generality one would, for this example, have to make the su~mation over S in Equation (10.12) run out to 120.Retaining as many factors as residual forward rates is therefore too onerous in practice for most applications. The choice must be made to reduce, sometimes drastically, the number of driving factors. ~ u ~ a r i s i n if g one : retained h(i) factors at each time step, then, after making use of relationships (10.5) and (l0.6), there would be enough degrees of freedom in order to specify any feasible exogenously specified covariance matrix element. nce such liberality in the number of factors is often not affordable, the interting ~uestion however is what happens to the ~ Q ~covariances e Z when the user is forced to retain only ~ ( i) c h(i) factors at each time step:
Clearly, the model covariance elements obtained by retain in^ only $(l) hctors will be diRerent from the desired ones. In p ~ i c u l a r if , only ~ ( i ) h(i) factors are retained, the non-~ia~onal model covarianc~ elements will not coincide with the corresponding arbitrarily specified e x o ~ e n o ~ inputs. s For a given choice of the time-dependent instantaneous volatility the calibration problem is therefore tantamount to specifying the behaviour of the model time-dependent correlation implied by the d y n a ~ c s of the ~ ( i ) h(i) factors. As shown above, the most eneral h-factor implementation of a l o g - n o ~ a l forward-rate BGM model is he matrix { b j k } , j = 1, h(i), k = l , S i .For future reference the rth column vectorin the matrix of elements { b j k } . In general an arbitrary target (market in what follows7) correlation function will not be reproducible with s(i) orthogonal factors. The user is therefore faced with the problem of determining the elements of the matrix { b j k } in such a way that
l. the o~hogonality between the different vectors is retained:
2. each vector is normalised to l :

bjk2 j=l,h(i)
= l Vk 5
s(i);
3. the sum of the coefficients {bjk}acro,~s~actors also adds up to one:

bjk
k=l,s(i)
= 1 v j 5 h(i);
4. the discrepancies between the implied (model) and market correlation matrices are minimised in some precise way to be defined.
Con~itions1 and 2 ensure orthonormality of the vectors, and condition. 3 ensures the correct recovery of a desired instantaneous volatility. The fulfilment of condi. general, optimisi tion 4 is the final goal of the exercise we are u n d e r t ~ n g In the coefficients { b . j k } is a complicated exercise, given the joint constraints ab0 sums over factors, about sums over forward rates and about the model correlation matrix. As far as the ful~lment of condition 3 is concerned, the expression of the {b} coefficients in terms of the polar coordinates of a hypersphere is very useful; the user can, in fact, perform an Unconst~aine~ search over angles, resting assured that the unit radius condition will always be satisfied; if that were not the case, a constraint on the SUM of the squares (numerically more unpleasant than a constraint on the individual value of each optimisation variable) would have to be imposed. Despite the help provided by the representation of the loadings in terrns of ~ n c o n s t ~ angles, a ~ ~ e however, ~ the remaining conditions still appear to impose a heavy computational burden. A simple and very useful theorern is of great assistance in the joint fulfilment of de erata 1 to 4. The gist of the theorem can be explained as follows: if the vectors were indeed o r t h o n o ~ a lwe , h o w that on matrix would simply be given b is an (n x n matrix of rank S.) mmetric matrix, and, as such, it can be the matrix of the resulting eigenvectors. (Note that, si nk S, only S eigenvectors will b is concerned, we do h o w that certainly gives the correlation matrix. Theorem 2 below then proves that Matrix multiplication is obvious1 er operation than matrix orthogonalisation (and the more so the larger the size of the matrix), and the saving in the computational burden becomes crucial when one has to calculate many co~elation ~atrices in the course ofan optimisation. Therefore, the advantage provided by the theorem is that one can rest assured that, as long as the sum of the squares of the elements in each adds up to one, the matrix obtained by simply multiplying rrelation matrix that will be obtained after the orthogonalisation is carried out. These qualitative statements can be made more precise as follows: be an n x S real matrix such that rank ( )= S. Then, since ea1 and symmetric, it can be diagonali ,i.e. there exists
an o~hogonal matrix (n x S ) and a diagonal matrix h ( n x Let

i
E)
such that
is the ith column of
If this is the case, then
is the correlation matrix of
The proof of this theorem can be found in Rebonato (l999a). Let us see in practice how this theorem can be of assistance in the calibration problem. Let us start from a matrix { h p } such that only condition 3 is satisfied:
As shown above, this can always be achieved, without loss of generality, for an arbitrary choice of the angles {e}.From the vectors (0) (and therefore from k } ) let us make use of Theorem 2: let us construct the matrix the ~ a t r {ib j~ } and evaluate the correlation matrix simply by by ju~taposingthe ve multiplying the matri transpose (as Theorem 2 shows one can do). Let us now vary the angles (0) in such a way that the sum of the (suitably weighted) squared discrepancies between the model and market correlation surfaces are ~ n i m i s e d .Since by Theorem l we know how to calibrate exactly to the variances of the various forward rates, calibration to the correlation matrix is all that is left in order to specify completely the most general log-normal forward rate model. But this latter task can always and simply be accomplished by virtue of Theorem 2 since the model correlation matrix { p j , ~is } just given by e precisely, the reasoning goes as follows: after o p t i ~ s i n g the angles {Q}, nce the coefficients { b j k } , in the desired manner to the market correlation function, the resulting vectors are neither n o ~ a l i s e d , nor other. However, after orthogonalising the h(i) x h(i) matri and 2 will be satisfied for the resulting eigenvectors multipli igenvalue. Let us denote by {a$) the elements of the new containing the eigenvectors ~ultiplied by the square root of crucial point is that, by Theorem 2, one can rest assured that, on-o~hogonal, non-normalised vectors, the correlawill be i~entical to the co~elation matrix obtained ing to a ~ a ~correlation ~ e t s ~ ~ a and ceo f ~ ~ l l ~ n con~itions I ~n~ 2 can the~efore be p ~ o ~ t a b t ly a c ~ l sepa~ately. e~ ~urthermore, thanks to the approach described above, a very highly constrained problem (see the first three conditions above, which must all be met at the same time) can
be solved by means of a totally unconstrained optimisation procedure over the angles Thanks to Theorems 1 and 2 the parameterisation of the most general B 6 log-normal forward rate model cantherefore be acco~plished in a fast and numerically efficient way so that, at the same time,
{e}.
1. an arbitrarily specified set of instantaneous volatilities can be recovered; 2. if these instantaneous volatilities are consistent with Equation (lO.l), then the Black volatilities are correctly reproduced within the model (and hence all the caplets are correctly priced); 3, the resulting driving factors are orthogonal to each other; 4. the bestfit (given a set of exogenous quality criteria) to a givenmarket correlation function compatible with the number S of factors can easily be found by means of a t o ~ ~ l~ ly ~ c o ~ s t r optimisation. ai~e~ Note that, at time step i, the market correlation matrix is characterised by h(i) x h(i) elements. Despite the fact that there are h(i) x s(i) terns in the matrix ,in reality only at most h(i) x [s(i) l] elements of the model correlation matrixcan be forced to match exactly8 (if one so wished) the co~esponding elements of the market correlation matrix. That this is the case can be intuitively understood from the fact that there are only h(i) x [s(i)l ] angles that can be changed at will. In particular, for the case of S = 2 one can obtain Theorem 3 below. For a two-factor model the correlation between forward rate i and and 0jl. For s = 2 the correlation pii is given by
j only depends on the difference between the angles
pij
=E
s=1,2
r=1,2
Recalling the definition of {bi,}, the orthogona~ityof the Brownian increments l for the angles, one (dw,}, and dropping the redundant second subscript obtains: bil = COS^^, biz = sint3i and therefore
Itwillbearguedlaterinthischapterthatforcedmatchingofagivensubsetofthemarket correlation matrix can in general be less than desirable, and indeed quite dangerous. This result was, to my knowledge, first obtained by Sherring i n 1995.
All the results reported in this section pertain to the case of a collection of twelve 12-month forward rates. The task was undertaken of fitting simultaneously and exactly to all the market volatilities, and of obtaining the best possible fit to a n (target or market) correlation matrix by using the procedure described As we know, any combination of the angles in Equations (10.5) and (10.6) is, by construction, compatible with the exact recovery of the volatilities of all the individual forward rates. We also h o w , however, that each of these combinations gives rise to a different correlation surface. Given an arbitrarily chosen quality function,*othe various angles can therefore be varied in such a way as to minimise this exogenously specified function. One such quality function, for instance, could be the sum of the squares of the 12 x 12 differences between the whole model and market correlation matrices. The exact procedure and the results are desc~bed in detail below.
{e}
andom numbers were first of all chosen for the 2 x 12 angles @ij ( i = , 2 , ...,12, j = 1,2). The coefficients {bij} (i = 1,2, ...,12, j = i,2,3) were created using Equations (10.5) and (10.6). Let be the 12 x 3 matrix made up by the vectors bij. The mo Relation matrix was then constructed by means of Theorem 2 as {pmod} = .The random numbers were then varied until the sum of the squared discrepancies [ p y d pTketl2 Over the ~ ~ u ~ Z e ~ twas r i reduced to a ~ i n i m u m The . market (target) correlation function was assumed to be given by the following expression:
{e}
p r k e t= LongCurr (1 L u n g c u ~ r ) e ~ p [~tj11 lt~ p = dl d2 max(ti, t j )
where L o ~ g C u r r = 0.3, dl = -0.12 and dz = 0,005. When the optimal vectors {bij}were found, the resulting 12 x 12 matri o~hogonalised,giving rise tonew vectors {aij}. Given the rank of the matrix, only three of the resu eigenvalues were different from zero, The vectors {bi.i},{aij}, the model ) andtarget correlation matrices, the eigenvectors and the eigenvalues resulting from the orthgonalisation are shown either in tabular form {Tables 10.1 and 10.2) or as figures (Figures 10.3 and 10.4). It is very interesting to note that, despite the fact that no orthogonality constraints were imposedin the optimisation, the solution qualitatively turned out to be very similar to what was found using the Principal ~omponents Analysis. Indeed, as shown inFigure 10.3, the rotation induced by the ort~ogonalisation was minor.
lo A quality function i s defined as the sum of the squared differences of specified subsets of the model and the target correlation matrices.
.l
The vectors { b ~ and } {q)
0.802 0.848 0.899 0.948 0,971 0.978 0.977 0.965 0.934 0.899 0.873 0.851
-0.490 -0.391 -0.248 "0.01 4 0.140 0.1 99 0.1 95 0.120 -0.048 -0.1 77 -0.250 -0.305
0.342 0.357 0.361 0.31 8 0.1 96 0.057 -0.085 -0.233 -0.354 -0.401 -0.420 -0.426
0.839 0.874 0.909 0.934 0.944 0.949 0.953 0.954 0.944 0.924 0.906 0.891
0.436 0.432 0.410 0.323 0.1 75 0.02 7 -0.1 11 -0.242 -0.329 -0.351 -0.356 -0.353
0.33 0.22 0.07 -0.1 5 -0.28 -0.3 1 -0.28 -0.1 8 0.01 0.1 5 0.23 0.28
The well-known shortco~ngs of low-dimensionality models in reproducing correlation unctions with positive convexity at the origin are well known (see, ebonato and Cooper (1995) or Rebonato (l 998a), where the implications for pricing are discussed at length), andwere indeed observed again in this study. ~ i t h o u repeating t material presented elsewhere, it will suffice to say that, by creating a too strong co~elation between adjacent forward rates, correlation between distant forward rates, the model correlaand a too tion surface obtained with a small number of factors systematically misprice~ swaptions ('short' swaptions too 'expensive', and 'long' swaptions too 'cheap'). Withthis proviso inmind, the overall agreement found by using the procedure described above is, however, good, and shown more clearly by the model co~elation surface in Figure 10.5.
~~~~
In order to appreciate to what extent the agreement between the model andmarket correlation surfaces is sensitive to the number of factors, a similar investigation was carried out for a four-factor approach. The same procedure described in Section 10.6.1 was followed by adding a further column of l 2 angles to the set used for the three-factor case. After st from totally random angles and optimising again as described above to th e-matrix quality function, the vectors {hij}and {aij} shown i n Figure 10.6 were obtained. Once again, the vectors {hij} thus obtained bear a close resemblance to the first four eigenvectors usually In particular, qualitatively found with the Principal Components Analysis (PCA). counterparts are linked b a it seems that the vectors {hij}and their orth~onalised phase rather than frequency transformation. This would indicate that th of Theorem 3 above can be extended to higher dimensions.
1.000
0.800
0.600
0.400
0.200
-0.200
-0.400
-0.600 l
Figure 10.3 The vectors {bij} and {aij}for the three-factor case (ail-matrixfit)
1.oooo
0.9000
0.8000
0.7000
0.6000
0.5000
0.4000
13 9
11
Fo~ard index
(a)
The market and model correlations between (a) the first, (b) the third, (c) the sixth and (d) the tenth forward rates and all the other forwards obtained using three factors,and imposing an overall best fit to the whole correlation matrix
1 .oooo
0.9000
0.8000
0.7000
0.6000
0.5000
0.4000
1 .oooo
0.9000 0~~000
0.7000
0.6000
0.5000
0.4000
(c)
1~0000 0*9000
0.8000
0.7000
0.6000
t
~
0.5000
0.4000
11
1 3
(d)
el
Q,
$2
c, W
Q, x
-t
It is interesting to compare the 'improve~ent'in the fitting of the same columns of the correlation matrixingoing from three to four factors. As is apparent from Figure 10.7, the greatest changes take place for the first series, i.e. for the correlation between the firstandsecond forward rates and all the others. Overall, however, the improvement is rather limited, confirming previous findings (Rebonato and Cooper (1995)) that the convergence to an exponentially decaying target correlation surface is very slow. In particular, the same qualitative features conce~ing negative convexity at the origin are still observed. The analysis so far has shown the results obtained by using the sum of the squared discrepancies over the correlation matrix as the quality function. For specific option problems, however, it might be argued that specific subsections of the correlation matrix might be more i~portant for pricing, and that an attempt should therefore be made to strive for a closer fit to these particular areas, It is therefore instructive to examine in detail the correlation surfaces obtained by using different quality functions. With four factors there are enough degrees of freedom to match exactly up to three columns of the correlation matrix. In particular, the first three columns could be targeted: this choice of the quality function could be motivated by the desire to capture exactly (or as well as possible) the correlation between the index rates and the residual swap rates to maturity first s h o ~ - m a t u ~ LIBOR ty in the case of a trigger swap. Whilst achieving this goal is indeed possible, as displayed below, if one so did the price one would pay would be an increasingly unsatisfactory fit (see Figure 10.8) to the remaining columns of the correlation matrix (giving rise therefore to a highly unsatisfactory correlation between later indices and the residual swaps). Even more interest in^ is to explore how the whole correlation matrix behaves when use is made of all the available degrees in order to fit as closely as possible to the elements of the matrix itselfwhichinfluence the value of a particular er~udan swaption (a 9NC2,'l was chosen for the example). Despite the fact that the fit tothe desired portion of the correlation surface is successfully achieved (see Figure 10,9), it is clear that the resulting overall surface is completely unrealistic (see Figure 10,10). In addition, the resulting eigenvectors bear hardly any resemblance to the familiar results of PCA; this fact leaves the user without any intuitional insight about the possible inte~retation of the factors as level, slope, curvature, etc. The results pertaining to the erm mud an-swaption fit, in addition, caution strongly against using a low imensionality, forward-rate-based model to price Bermudan swaptions by forci exact pricing of a subset of the underlying European swaptions (perhaps the most valuable of the underlying European swaptions). Note that if the user could afford to match exactly a E E the underlying European swaptions,
" A 9NC2, pronounced 9-non-call-2, ~ e ~ u d swaption a n is a swaption with a maturity of 9 years, which can be called or put on every reset date after year 2.
Oe5 0.4 4 0
10
12
( 4
6 (b)
10
12
0.9
0.8
0.7
0.6 0.5
0.4
three or four factors (Model3F andModelLCF, respectively) and imposing an overall fit to the correlation matrix
.7 omp par is on of the market and model correlations between (a) the first, (h) the fifth and(c) the eighth forward rates and all the other forwards obtained using
1 .oooo
0.9500 0.9000 0.8500

0.8000
3 0.7500
L-
0.7000
Market Modelcl ModelC2 ModelC3
0.6500 0.6000 0.5500 0.5000

1
1 1
13
( 4
Market Model~l ModelC2 ModelC3

0.6500
0.6000
0.5500 0.5000
1
1 1
13
Comparison of the market and model correlations between (a) the first, all the other forwards (b) the third, (c) the fifth and (d) the eighth forward rates and obtained using four factors and imposing an exact fit to the (Modelcl), first the first and second (ModelC2), and the first, second and third (ModelC3) columns of the correlation matrix
rather than a subset, then he would be in a position not to need any reduction in the di~ensionality of the problem, which is the very motivation of this study, and of most practical applications. ~~ap-rate-based models, instead, can reproduce y of factors and any exactly the prices of all the European swaptions for a ~ number implied co~elation (see E~uations (10.4) and (10.7), and Rebonato (3998a) for a drifts in the swap-rate-based framework). detailed discussion of ris~-neutral
2
1.0000
0.9500
*I_
0.9000 0.8~00
tr
0.8000 0.7500 0.7000 9 1

I
l
I
Market ModelCl ModelC2 ModelC3
l1
13
1.oooo
0.9500 0.9000 0.8500

0.8000
0.7500 0.7000
0.6500 0.6000
Along similar lines, for the quality function one could explore the choice of the sum of squared errors along the t~-diagonals of the model and market co~elation matrices. This choice could be seen as an attempt to reflect the importance of the correlation amongst contiguous forward rates in the pricing of i n s t r ~ ~ e nof t s the resettable-cap farnily (instruments, that is, where a stochastic strike on a given forward rate is determined by the reset of the immediately preceding forward rate). Without going into a detailed analysis, the resulting model correlation
10
12
o ~ a r Index d
(4
0.9000 0.8000
0.7000
5 0.6000 ._/
CI
"g 0.~000 B
m
0 0.4000
0.3000 0.2000
0.1000
t
L
,
Forward Index
6
1
10
12
@)
~ i ~1 0 ~ . 9 r Comparison e of the market and model correlations between (a) the first, (b) the fifth and (c) the eighth forward rates and all the other forwards obtained using four factorsand imposing a best fit to the elements of the correlation matrix which affect the value of a 9NC5 Bermudan swaption
1.oooo
0.8000 0.6000
._.
c,
m 0.4000
g!
-0.2000
(c)
matrix (see Figure 10.11) displays, once again, the sho~comingsof imposin over-fitting to particular subsections of the correlation surface. As is clearly visible from Figure 10.11, the effect of i ~ p o s i n g an (almost) exact co~elation along the three main diagonals has the effect of imposing an unrealistic deco~elation a~ongst forward rates farther apart. This, in turn, will produce undesirable tilts and twists in the yield curve which, after a sufficiently high number of resets, will ultimately have a strong effect on the price of the resettable product.
We have shown in this chapter that a simultaneous optimal calibration of a logte to the percentage ~olatiliti~s of the individual normal BCM f o r ~ ~ d - r amodel rates and to the c o ~ ~ l a t i o surface n is not only possible, but easy to achieve. The task can be accomplished in two separate steps: in the first part of the calibration (i.e. to volatilities) caplet prices are exactly recovered thanks to straightforward g e o ~ e t ~ crelationships. al After this step has been carried out, the fitting to the ~ such e ~ a way correlation surface, thanks to a simple theorem, can be o ~ t a i in that the calibration to the volatilities is not spoiled by the second part of the procedure; the resulting optimisation requires a simple matrix multiplication rather than a matrix orthogona~isation;
lmcyl_
o o o o g
0 0 0 0 0 0 0 0 0 0
d d d d d
0 0 0 0 0 j 0 :0 .0 ~O 0O 0 C
d d d d d
I
U
T-
3
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
22
0
p4
w ? z
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
v&Pv?v?z2!
despite the existence of several constraints on the final solution, the optimisation can be acco~plished in a totally unconstrained manner, Three- and four-factor imple~entationsof the approach were analysedin Section 10.6, andthe overall agreement betweentargetandmodel correlation surfaces wasshown to be very good. The difficulties in recovering the right convexity close to the main diagonal are not model-de~endent(see Rebonato and Cooper (1995)) and are inescapable, given the low di~ensionality of the model. Finally, the dangers of over-parameterisation, i.e. of forcing (near) exact fitting to certain portions of the co~elation matrix, were highlighted by ana lysin^ the cases of a trigger swap, a Bermudan swaption and a one-way floater (resettable cap).
It was shown inChapter 10 how to achieve an optimal simultaneous fit to caplet prices and to an exogenous correlation matrix. This approach could be directly applied if one took the market prices of caplets and an econometrically estimated correlation matrix as ones fundamental building blocks, This procedure would be particularly reasonable if one believed, as is often stated, that co~elation functions are relatively stable over time. The approach of Chapter 30, however, does not logically depend on the target correlation function having been estimated on the basis of econometric data: if the trader were to have views as to future correlations, these could be used instead of, or in co~bination with, the results of a purely statistical analysis. More importantly, if one were to assume a ~ a ~ i c ufunctional ~ar form for the instantaneous volatility of forward rates, then the prices of European swaptions would provide direct i n f o ~ a t i o naboutthe implied correlation function (see Case Study 10.1 in Chapter 10, and Section 3.4 in Chapter 3). The trader would still not be able to lock in this correlation structure by trading in liquid inst~ments (see Case Study j0.2 in Chapter 10); but she would be in a position at least to extract the conditional market consensus as to this correlation surface. It is crucial to point out that, in this context, the term conditional hi~hlights that the distillation of the co~elation is contingent upon a given instantaneous volatility surface having been assumed. It is also impo~ant to point out that the method presented in Chapter 10 did not depend for the validity of its implementation on any particular functional form for the instantaneous volatility functions, allowing as it does the recovery volatility ~ e u ~ofs each forward rate over each time step (see of the i ~ s t ~ ~ t ~ the equation below Equation (10.4) in Chapter 10). The problem therefore still remains of how these instantaneous volatility functions have to be specified.
The early implementations of the H J M and BGM models often allowed for flat (time-independent) volatility functions for the various forward rates. As shown below, little, if anything, can be said in favour of this choice apart from numerical simplicity, since the evolution of the term structure of volatilities it implies is one of the most unrealistic. In general, the term structure of volatilities, defined, it will be recalled, as the time-t function that associates an average (Black) volatility ( o ~ ~ to ~ k each ) matu~ty, T , is related to the instantaneous volatility (ar(u)) by the relationship (11.1) What one observes from the market is todays set of Black volatilities (as a functio~ of T ) : ( l 1.2) where the lighter notation oLIac, has been used to denote oiizck. One such function for CBP is shownin Figure 1 1.1. The most noticeable features of this curve are the steep increase inimplied volatilities from thevery short maturities to approximately 1 years, and the fact that, after this maturity, the implied volatilities are ono tonically decreasing with an upward convexity. The fact that this implied volatility curve has a highly structured shape (and that this shape,
12.00%
10.00%
.l The term structure of volatilities observed for the GBP market in January 1999. The dateson the x-axis are the expiry dates of six-month caplets
The l ~ t a ~ t a ~ Vo~ati~ity eo~ o f~ o r ~ a r
as shown later on, is roughly constant over time) has profound implications for
the instantaneous volatility function. In order to see whythis is the case, let us focus attention ontwospecific matu~ties. One can directly read from Figure 11.1 that, in order to price, say, the 10-year 6-month caplet and the 15-year 6-month caplet one would need average volatilities of approxi~ately 20.00% and 16.00%, respectively. If one made the assumption that the instantaneous volatilities were constant (and, therefore, equal to 20.00% and 16.00% throughout the lives of the respective underlying forward rates) the implied volatility for the 10-year caplet in, say, five years time can be easily calculated, and is given by the expression
andwouldobviously be equal to 16.00%.Todays 10-year implied volatility (CT:;::~), however, is, from the same graph, around 20.00%. Therefore theflat volatility assumption implies, in general, a change in todays implied volatility function. It is easy to see, from Equations (1 1.1) and (11.2) and from the discussion above, that the future term structure of volatility implied by a flat instantaneous volatility assumption will look like todays implied volatility curve translated in maturity by as many years as one is looking ahead in the future, to g mid-1998, is shown in An example, still from the GBP market, but r ~ ~ e r r i n Figure 112 , The curves in Figure 11.2 were obtained by enforcing the constant instantaneous volatility assumption. The figure displays todays term structure of volatilities, and the same ~uantityafter one, two and four years, as implied by the flat instantaneous volatility assumption. The dramatic change in shape in the term structure of volatilities implied by the flat volatilityassumption is clearly apparent from this figure. Apart fromthis rather extreme case, however, any choice of instantaneous volatility for the forward rates spanning the yield curve will uniquely determine, via Equation (1 1.l), the evolution of the term structure of volatilities. Therefore the objection frequently raised that, given our poor howledge about the future volatility of a forward rate, Occams razor should be extensively used, and the simplest (i.e. the flat) assumptio~ should be enforced does not stand up to s c ~ t i n y : our knowledge might be poor indeed, but the flat volatility hypothesis is in no way neutral,, implying as it does a very extreme evolution for the term structure of implied volatilities. An informed opinion can (and should) be formed by a trader about the evolution of tradable quantities such as the term structure of volatilities. Avoiding making a choice about instantaneous volatilities allows a black-box model to make those choices that are the job, and privilege, of a trader.
A s we have seen, exactly the same considerations apply to the case of the evolution of todays smile surface: there is no such thing as a neutral choice, and any model will imply a future evolution of the smile that the trader should consciously accept or reject.
l5
yr
( 16.00%)2du ==I
(1 1.3)
25.00%
23.00%
21.OO%
19.00%
17.00%
15.00%
13.00%
0.5
1.5
2.5
3 . 5
4.5
The evolution of the term structure of volatilities im~lied by a flat instantaneous volatilityassum~tion over four years. The implied volatility curve one in years time can be obtained by trans la tin^ the spot curve by one year to the left along the m~turity axis, and similarly for the two- and four-year curves
houses that have seriously embrace fram choices about the ~unctional he in nyof the methods that have been put in place are uite i ilar in spirit, and all attempt to come to terms with the p r o b l e ~ of user a reasonably direct control over the volatilit~evolution. The ap scribed below is in no waythe best, butit is represe~tative of several common form was, to my knowland pr~ctical i~plementa .The specific para~etric edge, first published by e et al. (1997), but several houses were already sing similar, or virtually identical, fu~ctional choices.
should Let us beginby assuming that the instantaneous volatility for all forwar~s be given by the following expression:
The l~sta~ta~eows Vol~tility o f For~ar
q ( t ) = [a
= OT(Z)
+ b(T -t)] exp[-c(T -t ) ]+ d = [a + bz] exp[-cz] + d, t = T t
( l 1.4)
where t is calendar time, and T is the expiry of a generic forward rate, The quantity c therefore represents the residual time to maturity of a particul~ forward. As for the functional form chosen, the presence of a linear term together with a decaying exponential allows for the existence of a hump in the curve, and the asymptotic instantaneous volatility is assumed to converge to a finite value, d. A few of the possible shapes are displayed in Figure 11.3. The functional form (1 1A) proposed above has the advantage of being extremely simple, and of affording, at the same time, a simple inte~retation for some of the parameters, or their combinations. To begin with, in the limit as z goes to zero, instantaneous and average volatilities tend to coincide, and therefore ( 0= ) a d = limT-0 OBl&(l?) gives some in~ication about the ~ e ~ a t i o ~ sa h~ ip the possible range of values for a and d. Furthermore, in the limit as z goes to infinity, the instantan~ousvolatility should converge to the implied volatility of the very long expiry caplet: O T ( O O ) = d = limT+ooOBl,&(?l). Finally, it is easy to show that the maximum of the instantaneous volatility curve occurs for t = (l/c a/b). If the user estimated statistically the portion of the life of a future or forward contract that is associated with its highest volatility, then this third relationship can provide some further useful indication about the plausible range of values for the parameters {a,b, c, d } . The last criterion directly brings us to the justi~cation of the humped shape as a possible candidate for the instantaneous volatility curve. Note that the presence of a hump in the term structure of volatility is not, by itself, a direct reason for imposing a similar shape to the instantaneous volatility curve. A humped term
Figure 11.3 A variety of shapes for the instantaneous volatilitycurve produced by the functional form(11.4)
st~cture of volatilities, for instance, could in principle be obtained if all forward rates experienced the same time-dependent volatility, and if this latter function were to display a maximum for some future time z*.(~ebonato (1998a)explains why this choice is neither realistic nor computationally appealing.) A more compelling financial justi~cationfor the existence of a hump in the i ~ ~ t a ~ t a volatility ~ e o ~ s curve can be given by the observation of the trading dynamics of different forward and future contracts: at the short end of the maturity spect~m futures contracts are pinned by the i m ~ n e n setting t of deposit rates, which are, in turn, influenced by the actions of the monetary authorities. In normal, periods, i.e. when no action from the monetary authorities is anticipated, the unce~aintyand the trading activity in the front contracts therefore tend to decline as they approach expiry. At the opposite end of the maturity spectrum the variation in market expectations about very distant forward rates is not as ~ronounced as it normally is for the matu~ty spanning a few months to one or two years. If correct, this state of affairs would give rise during normal periods to a maximum in the market uncertainty (and therefore, in the volatility of the forward rates) in the intermediate-maturity region, which the functional form (11.4) is supposed to be able to capture. if, on the other hand, the possibility of monetary action is deemed to be significant, then the uncertainty about the reset values of the cash rates which most directly affect the front futures contract can become very high. If this is the case, then the instantaneous volatility should increase sharply at the short end, and the hump would disappear. The functional form of Equation (1 1.4) allows for this possibility as well. These qualitative considerations therefore give a rationale for the functional choice (1 1.4), and the relationships between the coefficients and some quasiobservable market quantities can be of help in estimating a plausible range of values for the free parameters. It should be stressed, however, that these relationships should give at most an indication of the plausible values for the coefficie~ts, but should probably not be used to fix irrevocably three of the four parameters. Not only is the statistical estimation of quantities, such as the ~ a x ~ m u in mthe instantaneous volatility function, always fraught withdifficulties, but, even if these estimates (based on past history) were truly robust, there would still be no guarantee that the (forward-looking) option market has incorporated them into the prices of the liquid traded options. This observation can be ~~ticLllarly i~porta in periods shortly after a major transition in a ~nancial regime, such as the introduction of the Euro. Past history, in this case, can provide very unsatisfactor~ or incomplete information, I therefore show below how to bring into play the (forward-loo~ng) information that the caplet and swaption market does provide. To begin with, given a set of para~eters{a, b, c, d } one can check, for each forward rate, whether the integral of the square of the instantaneous volatility out to the expiry of the forward rate does coincide with the total Black variance, i.e.
whether
~
QB~~~~ =( T )[(a ~ T bt.) exp -(et.)
+ dl2 dz
(11.5)
In general the same choice of parameters for all forward rates will not allow the exact fulfil~ent of condition (1 1.5) for more than one forward rate. To each forward rate one can therefore associate a different scaling factor, kr, defined as
[(a bz) exp -(et.)
+ dl2dt.
and write for the forward-rate-s~eci~c instantaneous volatility function

Q T ( ~ )= k
~ ( [Ia b(T t)l exp[-c(T t)l
+ d)
With this choice of a f o ~ a r d - s p e c i ~ normalisation c factor the caplet condition is therefore ensured to be fulfilled by construction everywhere along the curve. In order to illustrate the procedure, Figure 11.4 shows a mar~et-implied volatility curve ( ~ a ~ V o 1 and ) the model curve (VolEst) obtai~ablefrom Equation (11.5) with a suitable choice of parameters before the no~alisation. Figure l 1.5 then shows the expiry-dependent normalisation factor, K T ,obtained using Eq~ation (1 1. G ) ,necessary to bring about exact pricing of the caplets,
26.00%
24.OoO/o
22.00%
20.00%
18.00%
16.00% 14.00% 12.00% 10.0O~h

0
A market (CBP 27 Nov 98) implied volatility curve (CapVol) and the (11.5) withasuitable choice of modelcurve(VolEst)obtainablefromEquation parameters beforethe normalisation (i.e. with kT = 1for all T)
3
1.060 1.040
. W
ti
U
1.020
1.000
0.980
0.940
0.960
0.900
0.920
0.880
re 11.5 The expiry-dependent normalisation factor, KT, obtained using ~~uatio n (1 1.6), necessary to bring about exact pricing of thecaplets, Note thatthe y-axis is rather compressed, spanning as it does the values from ap~roximately 0.9 to 1.05. The variation of the scaling parameter is therefore rather muted
~nalysing a curve like the one in Figure 11.S is extremely useful, and explains why the parameterisation proposed above can be very effective in controlling the evolution of the term structure of volatilities*In order to see why that should be the case, let us consider the case where all the { k ~coefficients } are exactly unity. From Equation (1 1.1) one can immediately see that, if that were the case, the of residual time to future term structure of volatilitites would look (as a functio~ maturity) exactly as todays, Therefore by looking at the degree of variation of ) across maturities gives an immediate understanding of the likely the { k ~values degree of time homogeneity of the implied volatility curve. To give a concrete example, Figures 1 1.6 and 11.7 show the evolution of the term structure of volatilities obtained from the instantaneous volatility curve in Figure 11.4 after normalisation. As one can readily appreciate, despite the fact that the future term structure shape is fundamenof volatilities is not identical to todayb curve, its ~ualitative tally preserved. Figures 11.6 and 11.7, in particular, can be profitably compared with the implied volatility surface that would be obtained using the flat volatility assumption, as displayed in Figures 11.2 and 1 1.8. Therefore, a choice of parameters {a,b, c, d } such. that todays model term structure of volatilities already bears (i.e. before the nor~alisation factors { k ~are } brought into play) a strong resem~lance to the arke et-implied volatility curve will automatically ensure that } be very close to unity, and the tern structure the correction factors { k ~ will of volatilities reasonably time stationary. One possible criterion for the choice of the {a,b, c, d } coefficients could therefore be the minimisation of thediscrepancy between the curves labelled Market and CapVol in Figure l 1.4, The fact that the functional form proposed in Equation (1 1.4) allows an analytic solution of
"Spot 1 year 2 year "4 year
0.5
l .5
2.5
3.5
4.5
urc! 11.6 The evolution of the term structure of volatilities over one, two and four years obtained from the instantaneous volatility curve i n Figure 11.4 after norma~isation
the integral in (l 1.S) makes the procedure simple and computationally fast (a few on a middle-of-the-range PG). I shall advocate later on that choosing the four parameters {a,b, c, d } usin only this criterion might not be desirable, but certainly the analysis outlined above provides the user with a powerful and practical tool. The next obvious question is: To what extent is it actually possible and desirterm structure of volatilities? The question able to achieve a ti~e-homogen~ous is addressed in the next section by employing a two-pron~ed approach, i.e. from an econometric and from a ~ar~et-implied perspective.
The question as to whether, or to what extent, the term structures of volatilities have indeed been t i ~ e - h o ~ o ~ e n e o inu the s past is fundamentally an empirical one. Whether theywill still displaythis feature in the future, however, adds an essential element of market j~dgement. Figures 11.9 and l l .10 show a small representativ~portion of the historical market-im~liedvolatility out to a maxi^^^ 10-year maturity which has been collected over a period of approximately two years (1996/199"7)for the GBP and FRF. The two figures are representative of the twomain classes of theterm structure ofVolatilities:those
FT-T-
+ J
t :
T-
3 1
that display a relatively stable shape (ofwhich the USD, GBP or DEM-at least before the euro conversion-are typical examples), and those that alternate between two mainshapes (of whichthe FRF was the most typical example before the euro conversion). As for the first type, it should be noted that certain structural features, such as, for instance, the steepness of the implied volatility curve close to the origin, tend to be very constant over time, Even much finer features, such as the existence of a plateau between approximately 1.5 and 2.5 years for the DEM,seem to be preserved, despite the fact that the plateau itselfmight change level (see Figure 1l .9). As for the second type of shape, the pre-Euro FRF volatilities clearly showed the existence of two distinct regimes: periods where the uncertainty about the short end of the yield curve would dra~atically increase, giving rise to implied ~olatilities of the order of 40% and to monotonically decreasing volatility curves; and more 'normal' periods when no action from the monetary authorities was expected by the market agents, and the more common humped shape prevailed. As can be appreciated looking at the far side of Figure 11.10, the change from the humped to the decaying states occurred very abruptly (in a matter of days), and tended to disappear much more slowly (over weeks). For the first (stable) class of term structures of volatilities, a good criterion for the choice of the instantaneous volatility parameters would be to ensure that the average (over trading dates) i ~ p l i e d volatilities for the different matu~ties are indeed recovered by the model future implied volatility curves. The resulting DEM o in Figure 11.11, and average curves are shown for the GPB and ~ r e - E u ~ one can readily note the qualitative s i ~ l a r i t y between the obtained average curve and the sample of curves presented in Figure 1l .9.
27.00%
6
CI
19.00% 17.OOYo
15.00% 13.00% 1 1.ooo/o
9.00% 7.00%
Figure l l .l l
The average (over approximate~ytwo years of trading dates-l 996, 1997) of the implied volatilities at different maturities observed in the market for the GBP, DEM and FRF
The ~ ~ s t a ~ t ~Volatility ~ e o u sof Forwar
317
The same procedure can, of course, be followed in the case of the unstable (i.e. FRF-like) term structures of volatilities, but in this case looking at the average of radically different regimes is obviously much less meaningful. Needless to say, this is only a starting point, and, if the trader had extra information (or simply different views) about the evolution of the implied volatility curve, then he could always choose a set of parameters that embody these views. This way of looking at the volatility p~ameterisation actually constitutes one of the greatest strengths of the BCM approach: by choosing a particular instantaneous volatility function, and therefore a future tern structure of volatilities, the exotic trader directly takes a position in the quantity (the instantaneous volatility) that constitutes the dif~cult-to-hed~e component of the present value of the exotic trade. This state of affairs can be compared with the predicament of a trader who has to use one of the many models that take unobservable ~uantities, such as the volatility of the all-driving short rate, as input. Note carefully that changes in the overall level of the volatility (which can be adequately captured also by the black-box models) are unlikely to hurt the trader significantly, as long as she has kappa-hedged herself against rigid movements in the implied volatility curve. But, asthe discussion in the previous chapter has shown, there are not enough liquid inst~ments tolock in with suitable hedges the finer but still very important cont~bution to the value of an exotic trade originating from the apportion in^ of the i~stantaneous volatility over the option life. It is therefore exactly in this latter dimension that the superiority of the BGM approach can be most easily appreciated. A further important comment regarding this proposed approach is that ~ Q S ~ term structure of volatilities can sometimes give some indication of the extent to which the market believes that it can remain unchanged in shape and level over time. More precisely, one can easily show (see, for example, Rebonato (1998a)) that, if the market operates in an arbitrage-free, log-normal Black world and the quantity :(T) = c ~ ~ ~ ~ ~ k is ( l not i )a ~ strictly T increasing function of maturity then the future term structure of volatilities cannot have the same shape and level as todays. In other words, if e ( T 1 ) > 6 ( 7 2 ) , with 7 7 1 772, then, even if we changed the functional form for the instantaneous volatility and added as many para~eters as we want, we would still not be able to find a normalisation vector { equal to I for all the forward rates. Therefore, the failure of the quantity :(F) to be strictly increasing does not imply any arbitrage possibility (even in a perfect Black world), as hi~hlighted in Chapters 1 and 2, but does imply that either the market does not believe in the forward rates being log-nor~allydistributed, or believes that the level and/or shape of the implied volatility curve will change in the future. Interestingly, the mostnotable violations of the condition df/dT >0 were observed in recent years for European currencies inthe run-up to the Euro conversion. Duringthisperiod observation of cap market practice in the less liquid currencies (say, ITL and ESP) and comparison with the same-matu~ty
quotes provided for the DEM indicated that traders tended to subscribe to a more norma~square-rootprocess for the forward rates, than to a log-normal process ain the brief discussion in Chapter 4, Section 4.5). f One couples this observation with the fact that the rates inmost currencies were expected to move ow^ towards convergence, one can readily explain why term structures of volatilities with locally decreasing e ( T ) were observed.
In the previous section one possible criterion (i.e. imposing that the vector { k ~ } should be as constant as possible) was proposed in order to choose the functional form and pin down the parameters of the instantaneous volatility function. As mentioned in Section 11.1, however, if one couples this approach with a similar choice for the instanta~eous correlation function, then one can bring into play the com~ined info~ation from the cap and swaption markets. If one regards a swap rate as a linear combination of forward rates (see Case Study 10.1 in Chapter 10) one can write a verygood approximation for the model average volatility to expiry of a swap rate made up of n forward rates as follows:
(11.7)
where
SR is the swap rate

f i is t o ~ ~ value y ' ~of the ith forward rate in the swap rate
is the model percentage volatility of the swap rate TeXp is the expiry of the swap rate Texp(i) is the expiry of the ith forward rate gi(Tex~(i) U) is the instantaneous volatilit~at time U of the T e x p ~ i ~ - e x p forward rate p( Texp(i),Texp~*i), U ) is the time-u instantaneous correlation between the forward rate expiring at time Texp(i) and the forward rate expiring at time Texp(j) wi are the weights in the linear combination SR = ,n w i f i
x,.=l
Several approximations have been used in arriving at Equation (11.7), which, in~identally,is slightly di~erentfrom the co~espondingequation in Case tudy 10.3 in Chapter 10. The mildest of the a~proxil~ations is that both the swap rate and the forward rates are jointly log-normally distributed. The impact of this assumption is extremely limited, as has been shown in Rebonato (1999b). F u ~ h e ~ o rgiven e , the notation employed, the instantaneous volatility for each
forward rate has implicitly been assumed to be a forw~d-rate-specificfunction e ~ c ethe expiry of the forward rate (whence the subscript i ) of the ~ ~ ~ e rbetween and the current time, U: crf(Texp(i) U). This expression has been used simply because this happens to be the case for the particular functional form discussed in the previous section. More generally, one would simply write crf(Texp(i), U), and the results would not be affected. More serious is the approximation that the weights w ican be treated as constants. Fortunately, their volatility is typically so much lower than the volatility of the forward rates that the assumption is ~ ~ disturbingly, c i e , the forward rates easily justi~able. Finally, and, ~ r i ~ ~ more have been assumed to be fixed at todays value. It is not easy to give an a priori justi~cationfor this assumption. However, if one carries out a Monte Carlo simulation of the forward rates, constructs the resulting swap rate, calculates the volatility of its logarithm and compares this value with the quantity obtainable from Equation (1l.7), then the discrepancies turn out to be very small indeed. ~ntuitively,the reason for this lucky state of affairs probably lies in the fact that the forward rates, which obviously change stochastically over time, describe (in the pricing measure) paths ending so~etimes above and someti~es below todays projected value, and the effect, possibly large on a path-by-path basis, almost exactly cancels out in the end. Equation (l 1.7) therefore provides a convenient link between the swap rate volatilities implied by a set of chosen instantaneous volatilities and correlations on the one hand, and the corresponding market value on the other. As me~tioned before, if we assumed to know withcertainty the instantaneous volatility function, and that the cap/swaption markets were perfectly coherent, then the swaptionimplied volatility matrixwoulduniquely pin down the u ~ n o w n correlation function (more precisely, would allow us to determine the covariance elements in Section 3.4 of Chapter 3). Instantaneous volatility functions are, however, not God-given; norshould one automatically assume that the cap and swaption markets should be perfectly in line with each other. Therefore a combined approach, where considerations about the time-homoge~eity of the t e r n s t ~ c t u r e of volatilities are analysed together with the resulting model swaption implied volatility m a t ~ x is , in practice probably the most profitable. If this approach is taken, then the user would not expect to match the market and model swaption volatility matrices almost perfectly everywhere, but would try to modify the choices of instantaneous volatility andlor correlation whenever the discrepancies show a strong systematic bias. As a rule of thumb, a good parameterisation in a liquid market should produce overall agreement within one kappa between model- and market-implied volatility surfaces. Tables l 1.111.3 and Figures l 1.12- 11-15 below provide some examples of good and obviously poor matches, In particular, Figures l 1.14 and 11.15show that the effect of moving from a flat to a time-dependent instantaneous volatility is very important. It is very impo~ant to point out that one can observe empirically that allowing for a time-depende~t volatility actually has a much stronger impact on the
32
Vol~tility of ~ o r ~ a r ~
g
9
c: . -
._.
U
l8 8 8
7 P -
$ 8 8
z
8
I
g uc! 4 g
2
CO
0
W
a ,
resulting swaption volatility m a t r i ~ than the effect produced by reasonable but pronounced changes in the instantaneous correlation. Therefore it mi fair to say that swaption prices are a blunt instrument in order to pin down the correlation function (see also Sidenius (1998) who concurs about this point). n the one hand, this might prompt the trader to conclude that, if she had to choose between a one-factor model with realistic instantaneous volatilities, and a multi-factor approach with flat instantaneous volatilities, she would probably opt for the f o ~ e r I. would concur with this choice, but, at the same time, this conclusion should not lull the user into the belief that instantaneous correlatio~ hardly matters at all: what appears to be the case for swaptions might not be the case at all for other exotic inst~uments, and the neglected correlation might turn out to be much more important for the pricing of a particular exotic instrument. The trader should always check, when pricing exotic products, whether the correlation impact is indeed as benign as it appears to be. Finally, the analysis of the modeland market-implied swaption volatility matrices clearly indicates that the swaption market (at least in all the major currencies) seems to trade strongly at odds with a flat instantaneous volatility assurnption. See, for instance, Figure 11.15. This observation therefore indicates that the market consensus indeed concurs with the analysis presented in the previous section, and prices in a significant time dependence of the instanta~eous volatility functions.
In this chapter the link has been highlighted between the (non-directly-observa~le) time dependence of the instantaneous volatility of forward rates and the evolution of todays (observable) term structure of volatilities. I have argued that, despite the fact that this link undoubtedly poses further and serious constraints on the modeller, this way of looking at option pricin~is actually extremely powerful and helpful, and constitutes one of the strongest points in favour of the approach. A specific separable function form for this crucial instantaneous volatility function, o ~ g i n a l published by race et al. (19973, was then analysed, and it was shown how the parameters characterising this description can be n ~ o w e d down (if not really pinned down uniquely) by the joint requirements of a good fit to the swaption-implied volatility function, and of an approximately tirnestationary behaviour for the term structure of caplet volatilities. ~ c o n o ~ e t rdata ic can also supply useful indications by giving an idea of the most likely ranges for the various parameters. Furthermore, the functional form for the instantaneous volatility suggested in Section 11.2 has the advantage of allowing for a very transparent link between statistically- and arke et-accessible quantities, and some of the parameters.
Market examples were then presented, highli~htin~ the fact that (within a range of reasonable values) the instantaneous correlation plays a relatively minor role in d e t e r m i ~ i n the ~ implied volatility of swaptions, and that, therefore, this between different market does not provide a very accurate tool to disc~minate possible functional forms or paramete~sations for the correlation surface, This does not mean, unfo~unately, that instantaneous corre~ation does not matter for any exotic instru~ent.
Abramowitz, M. and Stegun, 1 . A. (1972) ~ a n d ~ o f o aoat~he ma tical Functions, Applied Mathematics Series, Vol. 55. Washington: National Bureau of Standards; reprinted 1968 by Dover Publications, New York. Ames, W. F. (1977) ~ u m e r i c a l ~ e t h o ~ s . ~ o r ~ a r t i a l ~ i ~ e r e n tNew ial E~ York: Academic Press. Amin, K.I. (1993) Jump diffusion option valuation in discrete time, J~urnal o f ,no, 5, 1833- 1863. Bates, D. (1 996) Jumps and stochastic volatility, ~ e v i o e f Financial ~ Stu~ies, 69- 107. er, M. andRennie, A. (1996) Financial Calculus. Cambridge: Camb~dge University Press. Benjamin, J. R. and Cornell, C. A. (1970) ~ r o ~ a ~ Statistics~ i l i ~ , and ~ e c i s i o n for Civil Engineers. New York: McGraw-Hill Publishing Company. Beumee, J. G. B., Hilberink, B., Patel, S. and Walsh, P. (1999) Hedging Contingent Claims on risky Financial Instruments, Riskcare Publication (to appear in FOW). Bjork, T. (1998) bitra rage Theory in Continuous Erne. Oxford: Oxford University Press. Black, F. (1976)Thepricing of commodity contracts, Journal o f Financial Econornics, 3, 167- 179. Black, F., Derman, E. and Toy, W. (1990) A one-factor model of interest rates Journal, and its applicationtoTreasurybond options, Fin~ncial ~nalysts 330-339. Black, F. and Scholes, M. (1973) The ons and corporate liabilif a ~olitical l Econo~ics, ties, J ~ u ~o Boyle, P. P. (1977) Options: a Monte ch ,Journal o f Financial Boyle, P. P,, Evnine, J. and Gibbs, S. (1994) Valuationofoptionson several underlying assets, Working paper, University of Waterloo, Ontario, Canada. Brace, A., Gatarek, D. and Musiela, M, (1995) The market model of interest rate dynamics, Working paper, School of Mathematics, University ofNew South Wales, Australia,
~tten-Jones, M. andNeube er, A. (3996) Arbitragepricingwithincomplete markets, Appl. at^. Fin, atf field, C. and Collins, A. J, (1989) ~ntroductionto ~ultivaraite Analysis. Chapman & Hall. and Talbot, J. (1999) The yeddollar exchange rate in 1998: views from the options markets, an^ o f ~nglan~ ~uarterly 68-77. M. (1979) Option pricing: a simpli~ed Cox, J. C., Ross, S. A. and ubinstein, approach, Journal o f ~inancial ~cono7 ~ ,i 229-263. cs, a volatilityterm st~ucture, rouhy, M. and Galai, I) (1995)Hedgingwith Journal o f ~erivatives, (1990) Application of the e loss processes, ~nsurance:
ani, I. (1998) Stoc~astic implied trees: arbitrage pricing with , ~nternational Journal o f Theor~tical and Ap~lied ,I. and Zou, J. (1 996) The local volatility surface: Analysts ~ o ~ r n a l , in index option prices,~inancial
.U. (1990) Prices in Financial ~ a r ~ e t s .

n a ~ i Asset c Pricin~ eo^, 2nd
g, J. and Whaley,
(1 998) ~mp~ied volatility functions:
cingandhedgingwithsmiles,aribas
and Options Research Team, London, cing with a s ~ i l eRisk, ,
Capital ~ a r k e t s
(1998)
tion on Pricing For~ulas. New York:
A.and Morton, A. (1989) Bond pric paper,revised edition, ture of interest rates: a new met~odology, Wor~ing Cornel1 ~niversity, u r e Other s ~ e r i ~ a t i Securities, ve 2nd edn, ull, J (1993) ~ p t i o n s ~ F ~ tand wood Cliffs: Prentice Hall International Editions. F. (1997) Libo and swap market models and measures, Journal of a~shidian, Finance and Stoc~astic~, ~ o ~P.~ (1997) n , V a l ~ and e Risk. ~ h i c ~ e s tJohn e ~ : Wiley. aratzas, I. and Shreve, S.E. (1991) ~ r o w n i a n ~ u t i and o n Stochastic ~ a l c u l ~ s , 2nd edn. New York: Sp~nger-Verlag.
. (1993) A ~ v a n c e E~gineering ~ ~ u t h e ~ ~ t i7th c s edn. , Chichester:
Lal~berton, D. and Lapeyre, (199 1) ~ntroduction au calcul s t o c ~ a s t i ~ ~ e h la finance, Ecol ational des Ponts et Chausskes. appliqu~ D. and Rebonato, R. (1999) Closed-form solutions for option irfenderes~, pricing in the presence of volatility smiles: A density-function approach, submitted to Journal o f Com~ut~tion~2 Fin~nce. erton, R.C . (1973) heory of rational option pricing, ~el2 ~ournal o f Econo~nt mics an^ ~ u n ~ g e mSciences Merton, R, C. (3990) Continuous.Oxford: Blackwell. Naik, V, (1993) Option valuation and hedging tegies with jump in the volatility of asset reteurns, Journul o f~ina~ce Neftci, S. N.(1996) An Intro~~ction to the ath he^ New York: A c a d e ~ c Press. Nelson, D, B. and a ~ a s w a m K. ~ , (1990) Simp binomial appro~imations in Review o f Fi~uncia2 Stu~ie~, Oksendal, B. (1995) Stoch~stic ~i~erential E~uat Verlag. Plislsa, S. R. (l 997) ~ ~ t r o ~ ~ c tot ~ i o~n t ~ e ~ a t i c Press, ~. H., Teukolsky, S. A.,Vetterling, W. ~ ~ ~ e ~ Recipes i c a in l ~ ~ R ~ ~ Art N o -f ~ scient^^ h e C o ~ ~ u t i2nd n ~ edn. , Cambridge: Cambridge University Press, . (1997) A class of arbitr models, A ~ ~ p l i e ~ ~ a t h Fin e~atic~l Rebonato, R. (1998a) Interest Rate ~ p t i o Wile . (1998b) The a e of innocence, ~ u t u r e su n ~ ~ ~ t i~ o o~r sl ~ ,
.(1999a) ~alibrating the BGM model, Risk,
. (1999b) Onthe pricing implications of the joint log-normal it^ for the cap and swaption markets, accepted for publication Jo~rnal
taneous calibration of mu1 two-factor interest ce, mean reversion and short rate volatility th in el and of tree-based Net Exp models in general, Reiner, E. (1998) Understanding skew and smile behaviour in the context of jump processes and applying these results to the pricing and hedging of exotic options, Working paper presented at the Global Derivatives Conference, Paris, 28 April 1998.
ubinstein, M.(1994) Implied b i n o ~ i a ltrees, J o ~ ~ n a of Z~i~ance, 771-818. Schoen~uc~er, P. J. (1996) he term s t ~ c t u r e of defaultable bonds, or^^^ d isc cuss ion) Paper No. 384, Dept. of Statist~cs,Faculty of Economics, of Bonn. ~niversity irirne. s ~iami: Kolb P u b l i s h i ~ ~ S h i ~ ~D. o ,C. (1992) Finance in C u n t i n ~ o ~ Co~pany. Shreve, S, (1997)Stochastic calculus for derivatives, Course Notes for 1997 k train in^ Course, London. us, J. (1998) LIBOR m a r ~ e models t in practice, sub~itted to ~ o ~o f~ a o ~ ~ u t a t i o n ~inance. al ~ i l ~ o tP. t ,(1998) ~ e r i v ~ t i v eChichester: s. John Wiley.
9NC2 Bermudan swaption, 293 arbitrage, 3, 13, 18, 23-7, 80-2, 105, 129-30, 226-7 50 B JN, 229 implied volatility, 16 1 86, 2 16 20 strategy, 232-3 Debreu prices, 10 112 Arrow asymmetric smiles, 144-50, 247-50 at the money (ATM), 109, 318, 157-62, 202-5, 216-20, 226, 248-9, 261 average volatility, 15, 17, 29 backward (Kolmogorov) equation, l33 -5 basket options, 51 Troy model BDT see Black- Deman Bermudan swaptions, 293-301 Bernoulli distribution, 260 Bessel functions, 100 BCM see Brace-Gat~ek-Musiela approach binary options, 74-7 bino~al distribution bushy-tree approach, 263 -6 mean reversion, 260, 263-6 BJN see Britten-Jones and Neuberger Black and Karasinski model, 256, 262 Black market volatility, 15, 17, 34, 70, 113- 14 Black and Scboles model, 24, 42, 57, 73-7, 93, 131-5, 203, 226, 237, 244 hedging pe~ormance,3 74 plain-vanilla options, 30-7 risk aversion, 226 spot and forward processes, 8- 13 Black volatility, 15 ~iack-Derman-Troy (BDT) model, 253-67
continuous time equivalent, 2613 discrete time case, 259-61 lattices, 25966 paradox, 256 8 short-rate unconditional variance, 63 259 boxes see continuous double barriers B r a c e - C a t ~ e k - ~ u s i e l a(BGM) model, 254, 304-6, 327 8 1 case studies, 278 concept, 272- 3 conclusions, 298-301 construction, 273-7 dimensionality reduction, 28 1-5 four-factor model, 287-98, 301 numerical results, 286-98 optimal calibration, 27 1-301 three-factor model, 286-8, 289-91, 30 1 and Neuberger (BJN), B~tten-Jones 229 -50 case studies, 243-9 computational results, 243-9 developments, 249-50 discussion, 249-50 numerical implementation, 236-42 optimal hedge, 235-42,249 50 problems, 249 Brownian motion, 9, 19, 52, 98-9, 177, 195, 217-28, 257, 275, 285 bushy-tree approach binomial approach, 263-6 mean reversion, 263 -6 calls, 11-12, 47, 107-9, 132-3, 150-66, 169, 186-7 BJN, 236-42, 243-9
calls (cont.) degenerative hedging strategy, 176-7
instantaneous correlation,4, S 1-70, terminal correlation, 4, 5 1-70, 272-3 Grouchy-Galai const~ction, 44-50, 230,
272, 303, 318-28
GB2,199-206 caplets, 13, 17, 77, 86, 198, 207-13, 256-61, 273-81, 298, 304-11, 327 caps, 73, 273, 317-28 case studies, 18-22 BGM, 2'788 1 BJN, 243-9
caplets and a two-period swaption,
249
definitions, 3 -28 degenerative hedging strategy, 176-7 deltas, 8, 11, 21-2, 31-50, 74-7, 93-5,
176-7, 205-6,213- 18
DEM market, 316-18 density direct modelling, 189-214
278 -9
continuous double barriers, 174- 8l degenerative hedging strategy, 176-7 delta hedging, 74-7 European options, 58-60 forw~d-setting strike, 18 -22
GB2,197-214
quoted option price usage, 214 risk-neutral estimation, 1959 Derman and Kani (DK) model, 100-27,
135-7, 173-4
natural pay-off, 24 path-dependent options, 6 18 Quanto (Dim FRAs, 26-7 serial options, 280- l smiles, 74-8, 174-81 static portfolio replication, 77 l 8 1 terminal correlation, 57-68 t~o-pe~o swaption d and caplets, 278-9 cash-or-nothing binary options, 77-8 CEV see consta~t elasticity variance Gholesky decomposition, 69 closed-form solutions, direct density modelling, 189-2 14 complete markets jump-diffusion models, 89-90, 195 option pricing frame~ork,2293.5 constant elasticity variance (CEV), 79-80, assumptions, 130-5 concept, 1049 hedging performance, 173-4 impleme~tation results, 113 -27 numerical aspects, 10913, l20,
126-7, 135-9
problems, 109- 13, 120 detel~inistic discounting, concept, 6,
9- 13
dimensionality reduction, direct density modelling, closed-form solutions, 189-214 discount bonds, 7-8, 11, 25 discount factor, concept,23-4 discrete time BDT, 2596l financial theory, 10.5 DK see Derman and dominating strategy, 229 Dupire model, 100- l , 174 economic agents, concept, 2 16 17 effective volatility theory, 99, 100, 13 1 equities, 4-9 asy~etric smiles, 144-50, 247-50 option volatility,89 smiles, 83-5, 144-50, 186, 215-16,
100
constant volatility, 30-3, 184 continuous double barriers, 173, 174- 86 case studies, 177-81 static replication, 177- 8 1 unwinding costs, 182-6 continuous time, BDT, 261 -3 convergence
BJN,243-4
portfolio variance, 37-4 l co~elations BGM, 27230l imperfectly correlated variables, 52-7 i~portance, 3, 23, 27-8, S 1
219-50
spot quantity hedging, 6-9 Euro, 317-18 European options, 9, 58-60, 77-8,
176-7, 194, 230-2

European swaptions, 293 -301 exotics, 126, 228-9, 268, 327-8
explicit finite di~erences method, 104-5, 110- 11, 130 extraction, future local volatility, 129-87 floating smiles, 80-2, 120, 126-7, 132, 184, 216-50 F o k k e r - ~ l ~equation, c~ 133-5 foreign exchange (FXj, 87, 140-4, 186, 215-16, 245-50, 304-6, 311-18 option volatility, 8-9, 11, 13 symme~ricsmiles, 140-4, 247 -9 forward concept, 4-9 forward ( F o ~ e r - ~ l a n c kequation, ) 133 -5 forward price concept, 4-6 DK model problems, 1 10 processes, 6-7, 24, 250 future volatility, 4- 14 present volatility, 14 14, 194 volatility, 8 forward rate, 5-6, 68 BGM approach, 254, 271 -301 Black market volatility, 17 concept, 4-6 future vo~atility, 4-6, 14 instanta~eous volatility, 16, 303-28 spot quantities, 5-6 sticky smiles, 78-80 stochastic variance, 16- 17 volatility, 4-6, 14, 16, 17, 303-28 forward rate agreements (FRAs), 25-7 f o ~ ~ d ~ s e tstrike, t i ~ ghedging, 18-22 frame-of-reference-f~ee concept, 24 FRAs see forward rate a ~ r e e ~ e n t s market, 31 1, 315-17 fully stochastic volatility models, 88-9, 98-9, 100, 215 fundamental concepts, 3-28 future volatility, 4-6, 126 efficient local extraction, 129 87 forward processes, 8- 14 fo~ard rate, 4-6, 14 spot processes, 4-6, 13, 14 futures contracts, case studies, 18 FX see foreign exchange gammas, 21-2, 34-41, 204-6, 213, 217- 18 GB2 See Generalised Beta of the second kind GBP market, 304-6, 311- 16 Generalised Beta of the second kind (GB2j, 197-214
Girsanov theorem, 3-4, 27, 267 Green S function see Arrow -Debreu prices Heath- Jarrow--Morton (HJM) approach, 272, 304 hedging, 6-14,44-50,74-8, 226-9 BJN, 235 -42 Black and Scholes model p e r f o ~ a n c e , 174 constant Volatility, 30-3 degenerative hedging strategy, 176-7 forward-setting strike, 18-22 mean-reve~in process, ~ 41 -4 optimal hedge, 235-42, 249 performance compa~sons, 173 -4 lai in-vanilla options, 30-50 spot quantities, 6-9 time-dependent volatility, 34-41 HJM see Heath- Jarrow -Morton approach Ho and Lee model, 256 Hull and White model, 256 imperfectly correlated variables, stochastic evolution, 52-7 implied volatility, 15-17, 74-80, 93, 113-27, 131-5, 189-250,273, 28 304 acceptable prices, 167, 170- 3 arbitrage, 16 186, 2 16- 20 inversio~,11 local volatility links, 139-74 no-arbitrage conditio~s, 161-86 in the money, 13, 20-1, 109, 150-3, 160, 9 248 incomplete markets, option pricing framework, 229 35 instant~eous co~elation, 4, S 1-70, 272, 303, 318-28 instantaneous short rate, logarithm variance, 269-70 insta~taneous volatility, 15, 29, 43-5, 52-7, 273-85 8 conclusions, 327fitting, 3 1 1-27 forward rates, 16, 303-28 functional form, 306-13, 327 shape variety, 307 swaption-market i n f o ~ a t i o n 3 , 18 -27 1l 19 time-homo~eneity, 3 interest rates, 4-9, 13, 206-13, 253-328 BDT, 256-8
3 interest rates (cont.) BGM, 254, 271 -301 instantaneous volatility speci~cation, 28 303 mean reversion, 253-70 monotonic (interest-rate) smiles, 150-4, 160- 1 smiles, 85-7, 150-4, 160-1, 186 interest-rate forward contracts, 4-8 interest-rate options, volatility, 8 -9 Itos lemma, 11, 222-3, 225, 254-5 joint evolution, terminal correlation, 57 -68 jump-diR~~sion, smiles, 89-91, 195, 215-50 kappas, 20-1, 74-7, 80-1, 93-5, 191-2, 204-5, 216-18, 250 knock-out caps, 273 5 Kolmogorov equation, 133LIBOR-in-a~ears FRAs, present values, 25 -6 local volatility, 14, 101, 105, 113-26 DE( model problems, 1 1 1 12 efficient future extraction, 129-87 implied volatility links, 139-74 Monte Carlo simulation, 1536l martingale, concept, 24-7 mean reversion, 28, 41-4 DT, 253-67 general interest-rate models, 267-9 interest-rate models, 253-70 real-world situation, 29-50, 253, 267-8 ris~-adjustedsituation, 29-50, 253-8, 267 8 66 short-rate lattices, 259 true role, 267-9 Merton model, 2 19, 229 ~irfendereski-Rebonatoapproach, 126, 189, 196 mixed jum~-diffusion processes analytical description, 220-9 smiles, 2 15-50 money-market account, 23 monotonic (interest-rate) smiles, 150-4, 160- 1 Monte Carlo simulation, 19, 30- 1, 37, 52-70, 126, 130-2, 193, 250, 319
61 i ~ p l i e ~ l o volatility ~al links, 153 jump process, 224-5 ~ulti-factor models, 4

natural pay-off, concept, 24-7 numeraire, concept, 23 -7 Occams razor, 305 one-touch barriers see continuous double,.. one-way floaters, 273, 301 optimal calibration, BGM, 271 -301 optimal hedge, BJN, 235-42, 249 OTC see over-the-counter options out of the money, 13, 36-7, 109, 150-3, 159-65, 205-6, 216-20, 226 over-the-counter (OTC) options, 73 overview, ix-xiii partial diRerentia1 equation (PDE), 105, I 1 1, 133, 175 -8 path-dependent options, 61 PCA see principal components analysis PDE see partial differential equation plain-vanilla options, 9, 1 1 12, 17, 20, 25,44-50, 73-7 BJN case study, 243-9 constant volatility, 30-3 future local volatility extraction, 129-87 GB2, 206- 13 hedging, 30-50 41-4 mean~reve~i processes, n~ smiles, 93-5, 100, 104 tim~-dependentvolatility, 34-4 1 Poisson process, 92, 221, 225-6, 249 present values FRAs, 25-7 FRAs, 25-6 LIBOR-in-a~ears Quanto (Dim FRAs, 26-7 present volatility forward processes, 14 spot processes, 14 price future volatility, 4-6 spot quantities, 5-6 volatility, 4-6 pricing engine, 101, 130 pricing measure, real world, 22-8 principal components analysis (PGA), 286-7 probabilities, 189-214
DE; model, 105l 3 CB2, 197-214 puts, 11-12, 109, 132-3, 159-66, 199-206
169,
Quanto (Diff) FRAs, present values, 26-7 ~ i k o d ~derivative, m 34 Radon random-amplitude jump-diffusion models, smiles, 90- 1 ratchet caps, 273 real world mean reversion, 29-50, 253, 267-8 8 pricing measure, 22 50 variance, 29references, 32932 relative prices, concept, 23-4 resettable caps, 296, 298, 300 residual volatilities, concept, 231 restricted-stochastic-volatility models, smiles, 91-3, 97-127, 174, 186, 195, 206 7 risk, 22-3, 226-9 mean reversion, 29-50, 253 -8, 267-8 smiles, 93 -5 variance, 2950 ris~-neutraldensity function, estimation, 195-9 ris~-neutral SDE, 10-11, 255 Rubinstein model, 100-1, 174 SDE see Stochastic Differential Equation serial options, case studies, 280-1 ShimkoS approach, 2 14 66 short-rate lattices, mean reversion, 259 short-rate uncondition~variance, BDT, 63 259 sho~-rate-basedinterest-rate models, 70 254 smiles, 13, 73-95, 129-87 a s y ~ e t smiles, ~ c 144-50, 247-50 case studies, 74-8, 174-81 closed-form solutions, 189 -2 14 concept, 78 continuous double barriers, 173, 174-86 2 14 direct density modelling, 189 empirical facts, 83 -7 equities, 83-5, 144-50, 186, 215-16, 219-50 floating smiles, 80-2, 120, 126-7, 132, 184, 216-50
foreign exchange, 87, 140-4, 186, 215-16, 245-50 fully stochastic volatility models, 88-9, 98-9, 100, 215 interest rates, 85-7, 150-4, 160-1, 186 ju~p-diffusion, 89-91, 195, 215-50 mixed jump-di~usion processes, 215-50 models, 82, 87-93, 97- 127, 174, 186, 195, 206-7, 215-16, 219-50 monotonic (interest-rate) smiles, 150-4, 160- 1 no-arbitrage conditions, 161 -86 pl~n-vanilla options, 93 -5, 100, 104 rest~cted-stochastic-volatilitymodels, 91-3,97, 98-9, 100-27, 174, 186, 195, 206-7 risk derivatives, 93 -5 steepness, 162-7, 245 sticky smiles, 78-80, 120, 126-7, 132, 184 stochastic volatility models, 88-9, 91 -3, 97- 127, 215- l 6 symmet~c smiles, 140-4, 247-9 tales, 78-82, 216-20 127 tree methodologies, 97 unwinding costs, 182 6 spanning, 7 spot processes, 24, 62, 250 future volatility, 4-6, 13, 14 Monte Car10 simulation, 19 present volatility, 14 time-dependent volatility, 8- 14 volatility, 4-6, 8-14, 194 spot quantities forward rates, 5-6 hedging forward contracts, 6-9 price, 5 -6 spot rate, future volatility, 4-6, 14 spread options, 51 static portfolio replication, 17586 sticky smiles, 78-80, 120, 126-7, 132, 184 Stochastic Di~erential Equation (SDE), 9- 11, 16, 255, 257-8 stochastic evolution imperfectly correlated variables, 52-7 terminal correlation, 57-68 stochastic variance, 16- 17 stochastic volatility models, sniles, 88-9, 91 -3, 97- 127,215- 16 strangles, 2 12
33
strike, 5-6, 18-22, 73, 107, 117-18, 130-3, 153-66, 191 -6 DE;model problems, 11l 12 swaptions, 5l , 70, 73, 256, 278-87, 293-308, 318-28 sy~etric smiles, 140-4, 247-9 term structure of volatility, 17, 268-9, 303 -28 terminal correlation, 4, 5l -70, 272-3 case studies, 57-68 European options, 58-60 joint evolution, 57-68 8 properties, 67 stochastic variable joint evolution, 57-68 time-dependent volatility, 17, 52-70, 113-23, 136-8, 144, 151-2, 163, 215-18, 277-81 hedging, 34-41 lai in-vanilla options, 34-41 spot processes, 8- 14 term structure of volatility, 303-28 time-homogeneity, term structure of volatility, 3 1 1 19 total variance, 15, 16-18, 44-50, 250 tree methodologies, smiles, 97- 127 trigger swaps, 273, 281, 301 underlying, 3-8, 13-14, 57-74, 91-3, 130-5, 195-6, 214
unwinding costs, con~inuous double barriers, 182-6 Vailliant brackets, 27 variance, 3, 10, 12, 16-17, 28, 226-9, 250 logarithm of the instantaneous short rate, 269-70 real-world situation, 29-50 risk-adjusted situation, 29-50 short-rate BDT,259-63 volatility average volatility, 15, 17, 29 Black market volatility, 15, 17, 34, 70, 113-14 constant volatility, 30-3, 184 effective volatility theory, 99, 100, 131 forward processes, 8- 14, 194 forward rate, 4-6, 14, 16, 17, 303-28 importance, 3, 23, 27-8 interest-rate options, 8-9 local volatility, 14, 101, 105, 11 1-26, 129-87 present volatility, 14 price, 4-6 residual volatility concept, 23 1 spot processes, 8- 14, 194 stochastic volatility models, 88-9, 91-3, 97-127, 215-16 term structure, 17, 268-9, 303-28 terminology, 4

Great - Volatility and Correlation in The Pricing of Equity, FX, and Interest-Rate Options (Wiley, 2005) 2

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Great - Volatility and Correlation in The Pricing of Equity, FX, and Interest-Rate Options (Wiley, 2005) 2

Uploaded by

Copyright:

Available Formats

This Page Intentionally Left Blank

In the ~ r i of~ Equity, ~ nFX~ and ~

Risk ~ a n a g e ~ eand n t Analysis. Vol. l : ~easuring and ~ o d e l l i n g Financial Risk

Derivat~ves D e ~ y s t ~ e Using d : Structured Financial Products

Les Clewlow and Chris Strickland

Derivatives for Decision ~ a k e r s : S t r a t e ~ i c ~ a n aIssues ge~ent

Dictiona~ o f Financial Eng~neering

Volatility and Correlati~n in the Pricing o f E q u i ~FX , and ~nterest-Rate Opt~~ns

Credit Derivatives: A Guide to instru~ents and Ap~licat~ons

Chichester * New York * ~ e i n h e * i Brisbane ~

This Page Intentionally Left Blank

1.1 1.2 1.3 1.4 1. S 1.6 1.7 1.8

2.5 Hedging a Plain-Vanilla Option With Finite Re-Hedging Intervals 44

stant~neo~s and Termina~ ~orre~ations

ethodolo~ies for Smiley

This Page Intentionally Left Blank

This Page Intentionally Left Blank

This Page Intentionally Left Blank

This Page Intentionally Left Blank

This Page Intentionally Left Blank

PVi = Ai[FS(to,T i ) XilP(t0, Ti)

Ai[S(to)exp(r(t0, ri)(Ti to)) XiIP(to, Ti)

By'spothedging'astrategy underlying asset are taken.

is meantwhereby spot (asopposedtoforward)positionsinthe

Ti+l, then one can simply take positions in

S(T) = FS(0,T)ertp --$uZVgT ~ ~ v g J T ~ )

and, since at expiry, F ( T , T) = S(T), one can write

F S ( T , T) = FS(0, T)exp (--&FZ~~T oavgJTc.)

d F S ( t , T ) / F S ( t ,T) = a(t) dz(t)

the log-normal form

r a ~ ~ o d e rnuEIeraire n approach.) Traders

o ~=tf(~a"g(O ~ -+N ) ) EE f (var[S(O -+ N ) ] )

2)] = var[S(O + 2)] var[S(O " + l)]

ut we also know that

,in what follows, we are

n what follows the i

term which depended on th

with in Chapter 10, Section 10.4,

. The asset B is called a ~ ~ ~ e r ~ i r e . . The ~uantityl / is called the ~ i sfac~or. ~ o

andDefinition 5, one can write

were the case, if one hedged the

f the option with a ( d i ~ f e ~ e same vol~tilityas when the trade

realised volatility) than in

0000 0000 NVcO\L) NVooQ

me~n-revert in^ process (dS = ~ e v S ~ ~ ( ~ e S)dt v ~ e +vaSdz(t))

34.98588 10.51 709

0.7201 0.837356 1 83 0.8 0.6 0.4 0.2 0

Then, integrating both sides, we obtain

x ( T ) is normally distributed with mean x(0) and variance

a process of the form

Applying Ito's lemma to the function h ( f ) one readily obtains

Integrating Equation (3.4) from time 0 to time i ' one obtains

identify the terms in Equation (3.4) as

qua ti on (3.3j with the co~espondingterms in

CThas been extended in an obvious way to

with E[dz(t), dw(t)] = 0 It is straightforward to verify that, indeed,

(See, for example, Rebonato (1998a)J

. Integrate Equations (3.9) and

(3. IO) to obtain (3.l l )

that the From the discussion above we already know

. The correlation we seek to evaluate is given by

From the de~nition of covariance one can write

with the time-dependent weight, h(t), given by