KILIN ForwardVolatilitySensitiveDerivatives

Numerical advances in pricing forward volatility sensitive equity derivatives
Fiodar Kilin
Centre for Practical Quantitative Finance Frankfurt School of Finance & Management
A thesis submitted for the degree of Doktor rerum politicarum Supervisor: Prof. Dr. Uwe Wystup Submission date: July 7, 2009
Acknowledgements
I am deeply grateful to my supervisor, Professor Uwe Wystup for many helpful suggestions, important advice and constant encouragement during this work. I also wish to thank Professor Rolf Poulsen, Professor Wolfgang Schmidt, Professor Robert Tompkins, Professor Eckhard Platen, Dr. Bernd Engelmann, Dr. Peter Schwendner, Dr. Michael Dirkmann, Dr. Matthias Fengler, Dr. Friedrich Hubalek, Dr. Antonis Papapantoleon, Morten Nalholm and Martin Keller-Ressel for useful comments and discussions. I wish to express my gratitude to the sta of the Frankfurt School of Finance & Management that have provided an excellent research environment. The nancial support of Quanteam AG is gratefully acknowledged.
Contents
List of Figures List of Tables 1 Introduction 2 Models 2.1 2.2 2.3 Heston model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bates model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Barndor-Nielsen&Shephard model with the Gamma-Ornstein-Uhlenbeck latent state . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4 2.5 Levy models with stochastic time . . . . . . . . . . . . . . . . . . . . . . Bergomi model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 6 7 11 11 13 17 18 21 22 24 28 30 v vii 1 3 3 4
3 Pricing exotic options in stochastic volatility models 3.1 3.2 3.3 3.4 Model choice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Calibration and pricing of vanilla options . . . . . . . . . . . . . . . . . Monte-Carlo pricing of exotic options . . . . . . . . . . . . . . . . . . . . Analytical pricing of exotic options . . . . . . . . . . . . . . . . . . . . .
4 Accelerating the calibration of stochastic volatility models 4.1 4.2 4.3 4.4 Characteristic functions . . . . . . . . . . . . . . . . . . . . . . . . . . . Pricing methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Caching technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Numerical experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . .
iii
CONTENTS
5 Forward-start options in the Barndor-Nielsen&Shephard model 5.1 5.2 Derivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Numerical examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
35 36 39 51 51 53 56 58 65 67
6 On the cost of poor volatility modeling The case of cliquets 6.1 6.2 6.3 6.4 Forward volatility and forward skew . . . . . . . . . . . . . . . . . . . . Cliquet options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Price comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hedge performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7 Conclusion Bibliography
iv
List of Figures
5.1 5.2 5.3 5.4 5.5 5.6 Price of a forward-start call in the Barndor-Nielsen&Shephard model, sensitivity to comovement . . . . . . . . . . . . . . . . . . . . . . . . . . Price of a forward-start call in the Barndor-Nielsen&Shephard model, sensitivity to mean-reversion rate . . . . . . . . . . . . . . . . . . . . . . Price of a forward-start call in the Barndor-Nielsen&Shephard model, sensitivity to exponential law parameter . . . . . . . . . . . . . . . . . . Price of a forward-start call in the Barndor-Nielsen&Shephard model, sensitivity to Poisson intensity . . . . . . . . . . . . . . . . . . . . . . . Price of a forward-start call in the Barndor-Nielsen&Shephard model, sensitivity to initial latent state . . . . . . . . . . . . . . . . . . . . . . . Price and theta of a forward-start call in the Barndor-Nielsen&Shephard model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.1 6.2 6.3 Histogram of relative cumulative hedging errors of a call spread cliquet. Hedging with recalibration using delta and short-term vega . . . . . . . Histogram of relative cumulative hedging errors of a call spread cliquet. Hedging with constant parameters using delta and short-term vega . . . Histogram of relative cumulative hedging errors of a call spread cliquet. Hedging with constant parameters using delta and parallel shift vega . . 64 63 62 42 42 41 41 40 40
LIST OF FIGURES
vi
List of Tables
3.1 Models examined and recommended in some of the existing literature on model risk. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1 Comparison of three methods for pricing vanilla options in stochastic volatility models. Grid sizes that are needed to obtain some benchmark accuracy levels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 Comparison of three methods for pricing vanilla options in stochastic volatility models. Average calibration time . . . . . . . . . . . . . . . . 5.1 5.2 5.3 5.4 5.5 5.6 Price of a forward-start call in the Barndor-Nielsen&Shephard model, sensitivity to comovement . . . . . . . . . . . . . . . . . . . . . . . . . . Price of a forward-start call in the Barndor-Nielsen&Shephard model, sensitivity to mean-reversion rate . . . . . . . . . . . . . . . . . . . . . . Price of a forward-start call in the Barndor-Nielsen&Shephard model, sensitivity to exponential law parameter . . . . . . . . . . . . . . . . . . Price of a forward-start call in the Barndor-Nielsen&Shephard model, sensitivity to Poisson intensity . . . . . . . . . . . . . . . . . . . . . . . Price of a forward-start call in the Barndor-Nielsen&Shephard model, sensitivity to initial latent state . . . . . . . . . . . . . . . . . . . . . . . Price and theta of a forward-start call in the Barndor-Nielsen&Shephard model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.1 6.2 6.3 6.4 Input parameters for pricing and hedging tests . . . . . . . . . . . . . . Input parameters for pricing and hedging tests . . . . . . . . . . . . . . Calibrated Heston parameters. . . . . . . . . . . . . . . . . . . . . . . . Calibrated Barndor-Nielsen&Shephard parameters. . . . . . . . . . . . 49 56 56 57 57 48 47 46 45 44 33 32 14
vii
LIST OF TABLES
6.5 6.6 6.7 6.8 6.9
Calibrated VGSA parameters. . . . . . . . . . . . . . . . . . . . . . . . . Comparison of theoretical cliquet values. Scenario 1. . . . . . . . . . . . Comparison of theoretical cliquet values. Scenario 2. . . . . . . . . . . . Comparison of theoretical cliquet values. Scenario 3. . . . . . . . . . . . Comparison of theoretical cliquet values. Scenario 4. . . . . . . . . . . .
57 58 58 59 59 59
6.10 Comparison of theoretical cliquet values. Scenario 5. . . . . . . . . . . .
viii
GLOSSARY
Glossary
This glossary contains abbreviations for the models used in this thesis, the full name and a reference to a page, where they are rst dened.
Bates Bergomi
4 7 36
BNS (Barndor-Nielsen&Shephard) latent state) Heston 3 5
BNS-GOU (Barndor-Nielsen&Shephard model with the Gamma-Ornstein-Uhlenbeck CEV (Constant elasticity of variance) 8 6
NIG-CIR (Normal Inverse Gaussian with the Cox-Ingersoll-Ross stochastic clock) NIG-GOU (Normal Inverse Gaussian with the Gamma-Ornstein-Uhlenbeck stochastic clock) 6 6 VG-CIR (Variance Gamma with the Cox-Ingersoll-Ross stochastic clock VG-GOU (Variance Gamma with the Gamma-Ornstein-Uhlenbeck stochastic clock) 6 6 VGSA (Variance Gamma with stochastic arrival)
ix
GLOSSARY
Introduction
The main purpose of this thesis is to describe new numerical and analytical methods for pricing equity derivatives in stochastic volatility models. We introduce an innovative method of high-speed calibration for this class of models. We also develop an ecient method of pricing forward-start options. Using this method we derive a readyto-implement formula for this type of contracts. The described results are supported by numerical examples and description of thorough tests. Before describing the main results, we provide a literature overview that motivates introducing the new calibration and pricing techniques. We analyse applications of the introduced methods in ane stochastic volatility models (e.g., the models of Heston, Bates, Barndor-Nielsen&Shephard) and the Levy models with stochastic time change. Furthermore, we illustrate the use of these methods for model risk analysis and evaluating hedging performance of stochastic volatility models. Among the contracts which we consider are forward start options, variance swaps, volatility swaps, options on realized variance, reverse cliquets, accumulators and Napoleons. The obtained results are important in both practical and academic applications. Attractive benets from using the described innovative methods in the quantitative analysis departments of investment banks are speed and accuracy of calculations, stability of calibrated model parameters in time direction, stability and reliability of greeks. From the point of view of academic research the introduced techniques signicantly accelerate experiments that require numerous recalculations of prices in dierent models and under dierent market conditions.
1. INTRODUCTION
The advantages of stochastic volatility models1 over local volatility models are the possibility to produce a realistic forward smile, the leverage eect, the volatility clustering, the possibility to t the whole market implied volatility surface with ve or six parameters only, the possibility of hedging second order greeks with respect to volatility (i.e. volga and vanna). A natural question is often posed: Why are stochastic volatility models not used extensively in equity derivatives modeling, although they possess a lot of theoretical advantages? We point out following reasons for this situation: (1) In calm markets (e.g. 2004-2007), the local volatility model performs quite well and it is easier to handle. This statement is corroborated by tests described in Engelmann et al.(2006b). Mercurio & Morini(2009) also show that that the dynamic and hedging behaviors of local volatility and stochastic volatility models are not qualitatively as dierent as it is assumed in current market wisdom. (2) Application of stochastic volatility models often leads to calibration bias in cases of steep implied volatility skew. This bias negatively aects the quality of the valuation of down-and-out puts and makes static hedge strategies for reverse-knock-out options even more complicated. This problem is described in details in Engelmann et al.(2006a). (3) Greeks in a local volatility model are intuitive. Sensitivities in the stochastic volatility models are often not directly linked to observable market parameters. In giving these reasons we do not intend to advocate the local volatility model. However we would like to point out that practitioners still have not made their nal choice between local volatility and stochastic volatility models. This fact motivates further model risk research which is also a component of this thesis. This thesis is organized as follows: In Chapter 2 we describe stochastic volatility models that are used in this thesis. In Chapter 3 we describe the steps needed to price exotic options in stochastic volatility models and provide an overview of existing numerical techniques typically used to accomplish these steps. Chapter 4 introduces an innovative method of accelerating the calibration of stochastic volatility models. A formula for pricing forward-start options in the Barndor-Nielsen&Shephard model is derived in Chapter 5. An application of our numerical methods in assessment of model risk and hedging experiments is given in Chapter 6. Chapter 7 concludes.
Some authors use the terms stochastic volatility model and Heston model as synonyms. In this thesis we we use the notion stochastic volatility for all stochastic volatility models, not only for the Heston model.
1
Models
In this chapter we describe stochastic volatility models that are used in this thesis. A further overview of the most promising stochastic volatility models can be found in Schoutens et al.(2004) and in Cont & Tankov(2003). The popular models we consider are representative of three dierent approaches to model volatility clustering, namely, diusion models with stochastic volatility, nonGaussian Ornstein-Uhlenbeck-based models and Levy models with a stochastic timechange. Specically, we consider the models by Heston(1993), Bergomi(2005), BarndorNielsen & Shephard(2001) and Carr et al.(2003). Here we outline the risk-neutral dynamics in these models and describe the inuence of each model parameter on the form and dynamics of the implied volatility surface.
2.1
Heston model
The risk-neutral dynamics in the Heston model is dSt = rdt + t dWt , S0 0, St where
2 2 t , 0 0, dt = ( t )dt + t dW
(2.1)
(2.2) (2.3)
t ] = dt. Cov[dWt , dW
The parameter denotes the mean-reversion speed. The reciprocal of this parameter =
1
separates short and long maturities in the sense that asymptotic expressions
2. MODELS
for the ATM (at-the-money in the sense of strike being at the forward price) implied volatility and skew are valid for t (short-term asymptotics) and t (long-term asymptotics). The long-run instantaneous variance has a major impact on the longterm implied volatility surface1 - the long-term ATM implied volatility is approximately proportional to , the long-term skew (the slope of the implied volatility curves) is ap proximately inverse proportional to . The volatility of variance creates convexity in the implied volatility curves for each maturity and controls the dynamics of short-term implied volatilities. The correlation parameter also has two dierent objectives in this model: it creates the skew for each maturity and measures the correlation between the spot and and the instantaneous variance. The initial value of the instantaneous volatility is 0 . The state variable is not an observable in theory. However, in practice, this variable can be observed in liquid option markets using extrapolation in the implied volatility surface to the zero maturity and ATM strike. Such an extrapolation is justied by the short-term asymptotics of the implied volatility surface generated by the Heston model.
2.2
Bates model
This extension of the Heston model introduces jumps in the underlying process. The model equation is dSt = (r J )dt + t dWt + Jt dNt , St S0 0, (2.4)
where Nt is a Poisson process with intensity > 0. The process Nt is independent of t in Equation (2.2). Jt denotes the percentage jump size. It is lognormally, Wt and W identically and independently distributed over time ln(1 + Jt ) N ln(1 + J )
2 J , 2 2 J
(2.5)
t The volatility process t follows the SDE (2.2) and its driving Brownian motion W satises (2.3).
The asymptotic expressions for the ATM implied volatility and skew in the Heston model can be found in Gatheral(2005) and Bergomi(2004).
1
2.3 Barndor-Nielsen&Shephard model with the Gamma-Ornstein-Uhlenbeck latent state
2.3
Barndor-Nielsen&Shephard model with the GammaOrnstein-Uhlenbeck latent state

1
This model introduces simultaneous up-jumps in the volatility and down-jumps in the underlying price. The risk-neutral dynamics of the log-spot is
2 d(ln St ) = (r k [] t /2)dt + t dWt + dJt ,
< 0,
(2.6)
with the latent state following the process

2 2 dt = t dt + dJt ,
(2.7)
where Jt =
Nt
xn .
n=1
(2.8)
Nt is a Poisson process with intensity a, xn is an i.i.d. sequence, each xn follows an exponential law with mean 1/b. The cumulant function of J1 is k (u) = ln E (exp(uJ1 )) = au(b + u)1 . (2.9)
In contrast to the Heston model, the short-term skew in the Barndor-Nielsen&Shephard model is not explained by the dependency between the underlying and the latent state processes. In this model the short-term skew is generated only by the possibility of a jump in the underlying. Therefore the short-term skew is controlled by a triplet of parameters {a, b, }. Since the comovement parameter is assumed to be negative, this model produces a negative short-term skew. The long-term skew in the BarndorNielsen&Shephard model is generated by the superposition of two eects: jumps in the underlying and dependency between the two state processes St and t . Therefore the long-term skew is more sensitive to the comovement parameter than the short-term skew. The Poisson intensity parameter a has both static and dynamic objectives in this model. Its main static objective is to control the level of the long term ATM-volatility.
2 is stationary and has a marginal law that The reason for this fact is that the process t
follows a Gamma distribution with mean a and variance a/b. The dynamic objective
Barndor-Nielsen&Shephard model with the Gamma-Ornstein-Uhlenbeck latent state is a special case of the general Barndor-Nielsen&Shephard model. The risk-neutral dynamics for the general case is given by (5.4)-(5.5)
1
2. MODELS
of the parameter a is to control the dynamics of short-term implied volatilities. An intuitive comparison between the dynamic objectives of the Poisson intensity a in the Barndor-Nielsen&Shephard model and the volatility of variance in the Heston model can be seen here. However, it should be taken into account that t in the BarndorNielsen&Shephard model is not the only source of volatility since the term dJt also aects the returns.
2.4
Levy models with stochastic time
This class of models introduces stochastic volatility eects by making the time stochastic. The risk-neutral underlying process is modeled as St = S0 exp(rt) exp(XYt ), E [exp(XYt )|y0 ]
t
(2.10)
where Xt is a Levy process, Yt is a business time Yt = ys ds, (2.11)
and yt is the rate of time change. A possible choice for Xt is the Variance Gamma (VG) process or the Normal Inverse Gaussian process (NIG). Typical examples of the rate of time change yt are the CIR stochastic clock dyt = ( yt )dt + yt dWt , or the Gamma-Ornstein-Uhlenbeck process dyt = yt dt + dJt , where dJt is dened as in (2.8). A representative of this class of models is Variance Gamma with the CIR stochastic clock also known as Variance Gamma with Stochastic Arrival (VGSA). In this model Xt is a Variance Gamma process dened by three parameters: drift and volatility of the Brownian motion and the variance of the Gamma process that subordinates this Brownian motion. The VGSA short-term asymptotics is controlled by the three Variance Gamma parameters only. The short-term skew is mainly dened by the drift parameter1 . The subordinator variance creates convexity in the implied volatility
Of course, this is only true if is positive. If is zero the Variance Gamma process becomes a Brownian motion, i.e. produces no skew.
1
1/2
(2.12)
(2.13)
2.5 Bergomi model
curves for short maturities. The CIR stochastic clock prevents the long-term skew from attening too quickly1 . Therefore the lower the CIR-long-term-mean parameter , the higher the similarity between the long- and the short-term skews. The reciprocal of the CIR-mean-reversion rate separates short and long maturities in the same sense as the corresponding parameter in the Heston model. The dynamics of short-term ATM implied volatilities is mainly controlled by the CIR-volatility .
2.5
Bergomi model
When pricing exotic options the shape of future implied volatility surfaces should be taken into account. Bergomi(2004) shows the importance of this issue for cliquet options. The above described models do not accurately capture the market dynamics of the implied volatility surfaces. Motivated by these observations an option pricing model where the dynamics of the variance swap variances is modeled directly was suggested by Bergomi(2005). Here we outline the denition of this model. The dynamics of forward variances is modeled for discrete time intervals [Ti , Ti+1 ], where Ti = t0 + i, i = 0, ..., N . The time step is typically equal to a reset period of a cliquet option2 that we need to price and hedge. A set of the forward variance processes is dened as (t) =
Ti
(Ti+1 t)Vt
Ti+1
(Ti t)VtTi
0 t Ti ,
(2.14)
where VtT is the implied variance swap variance observed at time t for maturity T . Initial values of the implied variance swap variances calculated from this input using Equation (2.14). The dynamics of each forward variance process Ti (t), i = 0, ..., N d Ti (t) = Ti (t) ek1 (Ti t) dUt + ek2 (Ti t) dWt , Cov[dUt , dWt ] = dt.
1 i VtT , i = 0, ..., N 0
are used as are
an input to the Bergomi model. Initial forward variances
Ti (t0 ), i = 0, ..., N is modeled as
(2.15) (2.16)
The main disadvantage of Levy models without stochastic arrival is that the implied volatility skew attens too quickly. See, e.g., Chapter 13 in Cont & Tankov(2003). 2 Cliquet options are discussed in Section 6.2.
2. MODELS
Ti (t) is a random process in the time interval [t0 , Ti ]. At time Ti the variance swap variance for the interval [Ti , Ti+1 ] is known to be VTii+1 = Ti (Ti ) . The solution of the SDE (2.15) is T (t) = T (0) exp At where At = ek1 (T t) Xt + ek2 (T t) Yt , (2.19) 2 Bt , 2 (2.18)
T
(2.17)
2 Bt = e2k1 (T t) E Xt + 2 e2k2 (T t) E Yt2 + 2e(k1 +k2 )(T t) E [Xt Yt ] ,
(2.20)
dXt = k1 Xt dt + dUt , and dYt = k2 Yt dt + dWt . Over the interval [Ti , Ti+1 ] the risk-neutral dynamics of the underlying is dS = rt dt + i S i 1 dZt , S parameters i and i are recalibrated when t reaches Ti so that
CEV i (Ti ) = IV (STi , Ti+1 Ti , F, i , i ) ,
(2.21)
(2.22)
(2.23)
where r is the risk-free interest rate and the dividend yield is assumed to be zero. The
(2.24)
CEV CEV i (Ti ) = IV (STi , Ti+1 Ti , 1.05F, i , i )IV (STi , Ti+1 Ti , 0.95F, i , i ) ,
(2.25)
CEV (S , , K, , ) is the Black-Scholes implied vowhere F = STi er(Ti+1 Ti ) and IV i i Ti
latility calculated from the price of a call option in the CEV model (2.23) calculated at the time Ti . Here K and are strike and time to maturity of this option. The idea behind this approach is to have the model parameters and set by the responsible trader, where resembles a general level of the skew, which is called a risk reversal in
2.5 Bergomi model
the market and is inverse proportional to the overall volatility level. This way the skew becomes stochastic in addition to the volatility. The underlying process is correlated with both factors that drive the forward variance process via Cov[dZt , dUt ] = SX dt, Cov[dZt , dWt ] = SY dt. (2.26) (2.27)
As noted in Bergomi(2005), the model is currently dicult to use because of a lack of suitable calibration instruments. This will change if and once options on forward ATM or variance swap volatilities become standard products, thus qualitatively expanding the set of calibration instruments.
2. MODELS
10
Pricing exotic options in stochastic volatility models

Pricing exotic options involves several steps. First of all one has to choose a model that adequately reects the risks of a particular exotic option. Typical problems caused by an inadequate model choice are mispricing, unsatisfactory hedging, errors in estimation of market and credit risk. The second step is the calibration. This step requires fast evaluation of vanilla option and possibly other calibration instruments. At the nal stage Monte-Carlo simulation, solving dierential equation or some analytical method is required to calculate the price of the exotic option for the calibrated model parameters.1 In this chapter we provide an overview of the techniques that can be used in practice to complete these steps. Section 3.1 deals with the model risk issues, Section 3.2 deals with vanilla pricing and calibration, Section 3.3 describes Monte-Carlo simulation in stochastic volatility models, Section 3.4 provides some formulae for analytical pricing of exotic options that are used in the subsequent parts of this thesis.
3.1
Model choice
Investment banks use mathematical models for pricing and hedging exotic equity derivatives. These models do not necessarily reect the market dynamics adequately. Therefore the exotic contracts are exposed to model risk. The term model risk can be interpreted as the risk arising from the use of an inadequate model(Hull & Suo(2002)).This
1
In this thesis the valuation procedures for exotics are carried out via Monte Carlo simulation.
11
3. PRICING EXOTIC OPTIONS IN STOCHASTIC VOLATILITY MODELS
denition can be explained as follows. Suppose an exotic option should be hedged with liquid hedging instruments. If the dynamics of market prices of the hedging instruments contradicts model assumptions, the initial hedging strategy cannot be completed. It turns out that the initial price of the exotic option was false because this price was calculated based on assuming the existence of a hedging strategy, which cannot be executed in practice. Estimation of model risk for exotic options is not unambiguous. There are at least four approaches to accomplish this task. The rst approach is to test whether modeled stochastic processes t historical market data well. The second approach is to calibrate a model at one point in time and compare model and market prices at a later time. This approach has been used in Dumas et al.(1998), Gupta & Subrahmanyam(2000) and Driessen et al.(2000). The third approach consists of a straightforward comparison of exotic option prices in dierent models. Schoutens et al.(2004) report signicant price dierences between the models of Heston(1993), Bates(1996), Barndor-Nielsen & Shephard(2001) and Carr et al.(2003) for barrier options, cliquets and variance swaps. Eberlein & Madan(2007) also use this approach and provide numerical justication of high model risk for cliquets, options on realized variance and options on volatility. The fourth approach is to assume that the true data-generating process is produced by a complex model that takes into account specic risks of a particular class of exotic contracts. The pricing and hedging performance of this model can be compared with the pricing and hedging performance of the model being tested. This approach is used in Hull & Suo(2002), Andersen & Andreasen(2001), Longsta et al.(2001) and we also used it in Chapter 6, where we deal with cliquet options and use the Bergomi model as a benchmark model to test model risk of the Heston model, the BarndorNielsen&Shephard model and Variance Gamma model with stochastic arrival. Another possibility to deal with the model choice is to analyze and compare theoretical and heuristic features of dierent models and analytically identify factors that determine the risk of using these models. Bergomi(2004) used this approach to show a large model risk when pricing cliquet options in the Heston and jump/Levy models. An obvious disadvantage of this approach is that it does not quantify the model risk. Such an analysis is, however, a good starting point for any of the four model risk estimation approaches described above. Indeed the theoretical reasoning in Bergomi(2004) is validated by a pricing experiment that we present in Chapter 6 of this thesis.
12
3.2 Calibration and pricing of vanilla options
It is necessary to point out that all these approaches do not guarantee the absence of model risk in case of positive test results. If the test results, however, show that the model prices or hedge performance are too far from the benchmark levels, we can be sure that there is a substantial model risk. Another important aspect of the model risk issue is the sensitivity of hedge strategies to model choice. This problem is analyzed in Nalholm & Poulsen(2006). They carry out a simulation study and compared performance of dynamic and static hedge strategies for barrier options in Black-Scholes, Constant Elasticity of Variance (CEV), Heston, Merton and Variance Gamma models. They show that the standard deviation of hedge errors produced by static hedge strategies are more model risk sensitive than those errors arising from delta hedging the options. Poulsen et al.(2007) analyse whether the performance of risk-minimizing hedge strategies1 is sensitive to model risk. They quantify four sources of error: wrong martingale measure, parameter uncertainty, wrong data-generating process and greeks which are not executable with market products. They have found that only one of these sources, namely wrong greeks, has considerably negative eect to the performance of risk-minimizing hedge. Some papers dealing with the model risk issues conclude with recommendations for model choice in practice. Table 3.1 summarizes their arguments in favor of chosen models. The main contribution of this thesis in the analysis of model risk issues is a recommendation for model choice for pricing cliquet structures. We recommend to use the Bergomi model, which will be supported by our analysis in Chapter 6.
3.2
Calibration and pricing of vanilla options
A typical input for the calibration of stochastic volatility models is a set of market prices of vanilla options. During the calibration procedure a number of recalculations of model prices for this set of options is required. These recalculations are the most time-consuming part of the calibration algorithm. Therefore it is important to be able to calculate model prices of vanilla options as fast as possible. A commonly used
1
A locally risk-minimizing hedge is described in El Karoui et al.(1997) and Bakshi et al.(1997)
13
Paper Gatheral(2005) Cont& Tankov(2003)
Compared models local volatility, Heston Heston, exponential Levy, Bates, BNS, Levy with stochastic clock Heston, Bates
Instruments barrier options, cliquets, general analysis
Recommended models Heston Bates
Main argument dynamics of the volatility surface exibility of term structure, accuracy of calibration overparameterization of the Bates model hedge performance dynamics of the volatility surface
Detlefsen(2005)
barrier options, forward-start options general analysis cliquets
Heston
Mercurio& Morini(2009) Bergomi(2004)
local volatility, SABR Heston, Levy, Levy with stochastic clock, Bergomi
both models are unapt to hedging Bergomi, Levy with stochastic clock
Table 3.1: Models examined and recommended in some of the existing literature on model risk.
14
3.2 Calibration and pricing of vanilla options
method to accomplish this task is to use a semi-analytical formula. Raible(2000) shows that the price V of an option with payo w(ST ) can be calculated as V ( ) = where v (x) = w(ex ), = ln erT S0 , (3.2) (3.3) eRrT 2

eiu L [v ] (R + iu) (iR u) du,
(3.1)
St is the underlying price, T is the maturity of the option, r is the risk-free interest rate, the dividend yield is assumed to be zero, L [v ] (z ) is the bilateral Laplace transform of v for z C L [v ] (z ) =

ezx v (x) dx,
(3.4)
( ) is the characteristic function of XT ( ) = EQ (eiXT ), where XT = ln (3.5)
ST S0
rT,
(3.6)
Q is the risk-neutral measure, R is a real constant such that x eRx |v (x)| is bounded and integrable and the moment generating function mgf (u) of XT satises mgf (R) < . Applying the approach of Raible(2000) to the valuation of European call options reproduces the formulas of Attari(2004) and Carr & Madan(1999), which are discussed in Chapter 4. The purpose of the calibration procedure is to nd model parameters that reproduce the prices of traded options. This can be formally stated as an optimization problem. The objective function in this problem is some distance between the market and model prices. The objective function is minimized with respect to model parameters. Typical possibilities to specify the distance between the market and the model are the root mean squared error of absolute or relative weighted dierences of prices or implied volatilities. Detlefsen & Hardle(2006) point out that these dierent measures of the calibration error give rise to dierent sets of model parameters and that the resulting
15
values of exotic contracts vary signicantly. Their experiments have shown that the calibration error measures can be separated in two groups: Calibrations on relative prices, absolute implied volatilities or relative implied volatilities result in similar values of exotics. Calibrations on absolute prices lead to exotic option prices that are quite dierent from the prices of the rst group. The authors of this test have chosen the same weight to all points of the same maturity. Moreover, the weights have been assigned in such a way that all maturities have the same inuence on the objective function. Possible choices of the optimization methods for the calibration of stochastic volatility models are described in Maruhn(2009), Mikhailov & Nogel(2003), Ben Hamida & Cont(2005), Madsen et al.(2004) and Feoktistov(2006). We use a combination of the Dierential Evolution and the Levenberg-Marquardt optimizer in all numerical experiments that are described in this thesis and require model calibration. The Dierential Evolution algorithm is used to nd a good initial guess for the Levenberg-Marquardt optimization routine.1 Several runs of the Dierential Evolution2 result in a small set of dierent initial values for the Levenberg-Marquardt algorithm. For each of these initial values a local optimum is found using the Levenberg-Marquardt optimizer. Thus a set of guesses for the global optimum is obtained. We apply an objective function to this set and nd the global minimum. The main contribution of the Dierential Evolution step to the described two-step optimization algorithm3 is to provide the features of global optimization. If we used the Levenberg-Marquardt step only, it would be a local optimization algorithm that does not suit our purpose of calibrating the stochastic volatility model. The main contribution of the Levenberg-Marquardt step is to increase the accuracy and the speed of the optimization algorithm. Numerical experiments show that if we use the Dierential Evolution step only, a lot of computation time is spent when the optimum is already reached. At this point of the algorithm the Dierential Evolution produces unnecessary jumps near the optimum because of the heuristic nature of this optimizer. If at this point we turn on the Levenberg-Marquardt algorithm, the optimum is found in one or two iterations because of the deterministic nature of
1 2
Maruhn(2009) proposes a technique that allows to avoid the Dierential Evolution step. Typically three to ten runs. 3 The rst step is the Dierential Evolution, the second step is the Levenberg-Marquardt optimizer.
16
3.3 Monte-Carlo pricing of exotic options
this optimizer. The Dierential Evolution algorithm is described in Price et al.(2005).1 The Levenberg-Marquardt algorithm is described in Gill & Murray(1978).2
3.3
Monte-Carlo pricing of exotic options
Recently a few papers on ecient discretization of the continuous-time stochastic volatility processes have emerged. Most of them deal with the Heston model. Kahl & Jackel(2005a) apply an implicit Milstein scheme for the Heston variance process V (t), (t) for the stock process S (t). Their scheme coupled with a particular discretization S is
(t + ) = ln S (t) (V (t + ) + V (t)) + V (t)ZV ln S 4 1 (t + ) V (t) (ZX ZV ) + 1 (ZV 2 1), + V 2 4 (3.7) (t) + + V (t)ZV V 2 2 + 1 4 (ZV 1)
(t + ) = V
1 +
(3.8)
where ZS and ZV are standardized Gaussian variables with correlation . Andersen(2007) points out that this scheme can lead to negative values of V (t) if 4 2 . (t) drops A possible solution to this problem is to use Euler discretization whenever V (t + )+ and V (t)+ , rather than V (t + ) below zero, and simultaneously using V (t), in (3.7). Andersen(2007) proposes, however, using a quadratic exponential and V approximation of the density near zero, rather than just truncating. Broadie & Kaya(2006) propose a simulation scheme of the Heston model that consists (t + ) from the non-central chi-square disof the following three steps: (1) Sample V (t + ) and V (t) draw a sample tribution; (2) Conditional on V
t+ V t
(u) du using
t+ t
1 2
V (u) dWV (u) =
t+
V (t + ) V (t) +
t
V (u) du ;
(3.9)
An open source code is available from http://www.icsi.berkeley.edu/ storn/code.html An open source code is available from http://quantlib.org.
17
(3) Draw a sample of log-spot conditional on V (t + ) and ln S (t + ) = ln S (t) + (V (t + ) V (t) ) t+ 1 + V (u) du + 1 2 2 t
t+ V t
(u) du: (3.10)
t+ t
V (u) dWV (u).
(3.11)
This algorithm is bias-free by construction. Modications of this algorithm that are necessary to calculate accurate greeks are described in Broadie & Kaya(2004). Recent papers of Zhu(2008), van Haastrecht & Pelsser(2008) and Smith(2007) also propose Monte-Carlo simulation schemes for the Heston model and report computing times and accuracy comparable with the schemes of Andersen(2007). In Chapter 6 we use a simulation scheme of Andersen & Brotherton-Ratclie(2005) which is based on a moment-matched log-normal approximation
(t)2 +(t)ZV (t + ) = + (V (t) )e e 1 2 V ,
(3.12)
where (t)2 = ln 1 +
1 2 2 1 2 V (t) (1
e2 )
(t) )e )2 ( + (V
(3.13)
3.4
Analytical pricing of exotic options
Some types of exotic options can be priced in a model-independent way under certain assumptions. If a particular stochastic volatility model fullls these assumptions, this is a very convenient way to price the option in this model. An example of such a situation is pricing variance swap under diusion assumptions. Since the underlying process in the Heston model is a diusion, variance swaps can be priced in this model using the replication formula even without calibration. A variance swap is a forward contract on the realized annualized variance. Its payo is some notional amount times
2 2 R0 ,T Kvol ,
(3.14)
2
where
2 Rt =u 2 i ,tj ti+1 <tn <tj
Stn ln Stn1 St ln j Sti
Mti ,tj ,
(3.15)
Mti ,tj
1 = ji
(3.16)
18
3.4 Analytical pricing of exotic options
u 2 is an annualization and rescaling factor1 and Kvol is a volatility strike. Since this is a forward contract, this payo can be negative. There exist two versions of the variance swap payo. In the second version the term Mti ,tj is not substracted2 . The properties of the second version of the variance swap are used in the denition of the Bergomi model. The oating leg of a variance swap can be approximated as a total variance of the underlying. The total variance, in turn, can be replicated using an innite strip of European options under assumption of no jumps. The fair value of total variance WT in this case is3
4 T 2 S dt = 2 t 0
EQ where
p(k ) dk +
0
c(k ) dk ,
(3.17)
c(y ) =
(S0 ey ) C , S0 ey (S0 ey ) P , S0 ey
(3.18)
p(y ) =
(3.19)
(K ) and P (K ) denoting undiscounted call and put prices. C However, if the calibrated parameters of the Heston model are already available, a simpler formula can be used to calculate fair value of total variance, namely
T
EQ
2 S dt = t
1 eT 2 (t ) + T. T
(3.20)
The fair strike of a volatility swap can be also approximated analytically. The volatility swap is a forward contract on the realized annualized volatility. Its payo is R0,T Kvol . (3.21)
Brenner & Subrahmanyam(1988) show that the Black-Scholes formula for the at-themoney-forward call with volatility can be approximated by S0 T BS . C ( ) 2
1 2
(3.22)
252 Typical value of the annualization and rescaling factor is u 2 = 10000 j . i The impact of the term Mti ,tj on the price is negligible. Its omission makes the payo additive. 3 The derivation of this formula is given in Gatheral(2005). 4 This formula assumes zero interest rates and the absence of dividends. Thorough analysis of the eect of dividends on the valuation of variance swaps is given in Buhler(2009)
19
Feinstein(1989) applied this result in order to approximate the fair strike of the volatility swap. Indeed S0 impl (F0 ) T C BS (impl (F0 )) = EQ (C BS (R0,T )) 2 S0 EQ (R0,T ) T S0 R0,T T )= , EQ ( 2 2
(3.23)
where F0 is the forward price and Equation (3.22) is used in both cases when the approximation sign occurs. Therefore the fair strike of the volatility swap is approximately EQ (R0,T ) impl (F0 ). (3.24)
This approximation is valid only at the inception time of the volatility swap. Carr & Lee(2007) propose a synthetic portfolio that results in a robust replication strategy and gives the possibility to approximate values of volatility swap at all times between inception and maturity. An ecient method of calculating prices of forward-start options in the Heston model is described in Lucic(2003). We use it in Chapter 6 when calculating prices of callspread cliquets. A number of analytical formulae for option pricing under assumption of continuity of the payo function and/or the underlying process is available (e.g. Raible(2000), Borovkov & Novikov(2002)). Eberlein et al.(2008) analyze conditions under which the valuation formulae hold in a general setting, i.e. for discontinuous payos and for variables that might not possess a Lebesgue density.
20
Accelerating the calibration of stochastic volatility models

When implementing a calibration algorithm for an option pricing model with known characteristic function of the assets return1 , one has to choose a method for pricing vanilla options. In this chapter we compare the following methods: (1) Direct integration, (2) Fast Fourier Transform (FFT), (3) Fractional FFT. Before choosing one of these techniques, it is important to consider all possible ways of improving accuracy and calculation speed of each of these methods. These improvements can include mathematical modications as well as implementation techniques. In this chapter we compare optimized implementations of the calibration algorithm based on each of the above mentioned valuation methods. It helps to identify the factors which are most important for accuracy and speed of calibration. We show that using an additional cache technique makes the calibration with the direct integration method at least seven times faster than the calibration with the fractional FFT method. In the existing literature unoptimized versions of direct integration are criticized. Carr & Madan(1999) point out the inability of the direct integration method to harness the computational power of FFT. Lee(2004) and Carr & Madan(1999) point out the numerical instability of the direct integration method in case of using a decomposition of an option price into probability elements. However the modication of the direct integration method described in Attari(2004) is free from this instability. In the present
1
E.g., Heston, Bates, Barndor-Nielsen&Shephard models or Levy models with stochastic time.
21
4. ACCELERATING THE CALIBRATION OF STOCHASTIC VOLATILITY MODELS
chapter we compare systematically all advantages and disadvantages of the three valuation methods. A special attention is paid to the possibility of a simultaneous valuation of a set of options and to the eciency of the applied numerical integration methods. The outline of this chapter is as follows. Section 4.1 lists the characteristic functions of stochastic volatility models that are used in our numerical experiments. Section 4.2 describes the pricing methods compared in this chapter. In Section 4.3, we describe the caching technique that accelerates the calibration with the direct integration method. Section 4.4 elaborates the details of a numerical experiment that compares the speed of calibration for the compared methods.
4.1
Characteristic functions
The methods discussed in this chapter can be applied to calibrate a bundle of models. The only model-specic element of the calibration algorithm is a calculation of the characteristic functions ( ) dened by (3.5). The derivation of exact expressions for the function ( ) in stochastic volatility models can be found in Gatheral(2005), Cont & Tankov(2003) and Zhu(2000). This subsection lists the characteristic functions of the variable XT 1 for the models used in our calibration tests. Heston model. The characteristic function of XT in the Heston model is given by (4.33). Bates model. The characteristic function of XT in the Bates model is ( ) = exp{1 + 2 3 }, where 1 = 2 (( i d)T 2 ln( 1 gedT )), 1g (4.2) (4.1)
2 2 2 = 0 ( i d)
1 edT , 1 gedT
(4.3)
2 3 = J iT + T ((1 + J )i exp(J (i/2)(i 1)) 1),

1
(4.4)
The variable XT is dened by (3.6).
22
4.1 Characteristic functions
d and g are given by (4.34) and (4.35). Barndor-Nielsen&Shephard model with the Gamma Ornstein-Uhlenbeck latent state. The characteristic function of XT in this model is given by ( ) = exp{1 + 2 }, where
2 1 = (b )1 T 1 (u2 + iu)(1 eT )0 /2,
(4.5)
(4.6)
2 = a(b f2 )1 (b ln(
b f1 ) + f2 T ), b iu
(4.7)
f1 = iu 1 (u2 + iu)(1 eT )/2, f2 = iu 1 (u2 + iu)/2.
(4.8)
(4.9)
Levy models with stochastic time. The characteristic functions of XT in the Levy models with stochastic time can be composed from the characteristic exponent of the Levy component and the characteristic function of the rate of time change according to the formula (ic( )) , (4.10) (ic(i))i where the characteristic exponent of the Levy process Xt is the logarithm of the char( ) = acteristic function of the value of the process at time t = 1 c( ) = ln E [exp(iX1 )] . The characteristic function of the CIR stochastic clock is ( ) = where = 2 22 i. (4.13) exp(2 t/2 ) exp(2y0 i/( + coth(t/2))) , (cosh(t/2) + sinh(t/2)/ )2/2 (4.12) (4.11)
The characteristic function of the GOU stochastic clock is ( ) = exp iy0 1 (1 et ) + a i b b ln b b i1 (1 et ) it . (4.14)
23
The characteristic exponent of the Variance Gamma process is c( ) =

2 2 ln(1 + 1 2 i ) ,
(4.15)
where and are drift and volatility of the subordinated arithmetic Brownian motion, is variance of the Gamma subordinator. The characteristic exponent of the Normal Inverse Gaussian process is 1 1 + 2 2 2i c( ) = ,
(4.16)
where and are the drift and volatility of the subordinated arithmetic Brownian motion and is variance of the Inverse Gaussian subordinator.
4.2
Pricing methods
In this section we describe the pricing methods and point out their limitations. Direct integration. The direct integration method implies computing vanilla call option values using one-dimensional numerical quadrature, for example Gaussian quadrature.1 For this method Attari(2004) obtains an ecient formula2 1 C (S0 , T, K ) = S0 erT K (I + ), 2 where 1 I=
+ 0 ( )) ( )) (Re(( )) + Im( ) cos(l(K )) + (Im(( )) Re( ) sin(l(K )) d, 1 + 2 (4.18)
(4.17)
l(K ) = ln and K denotes the strike price.
KerT S0
(4.19)
The advantages of this formula in comparison with the formula of Heston(1993) are: 1.) Formula (4.17) contains only one integral instead of two.
The rst analytical formula for pricing vanilla options in stochastic volatility models with known characteristic function was obtained in Heston(1993). Bakshi & Madan(2000) extend this approach and point out its theoretical advantages. 2 An equivalent formula has been obtained by Lewis(2001). This formula can also be used with the direct integration method.
1
24
4.2 Pricing methods
2.) The integrand in (4.17) has a quadratic term in the denominator. This gives a faster rate of decay. The implementation of this method should control the branches of the complex logarithm that appears in the characteristic function. It slightly complicates the implementation of this method, but does not aect the accuracy and the speed of the calculations. One possible solution of this problem is a reimplementation of the complex logarithm routine with storing the returned value and the branch number at the previous step of the algorithm. An alternative solution is described in Kahl & Jackel(2005b). Lord & Kahl(2006) and Albrecher et al.(2007) show that for some models this problem can be solved by using an appropriate representation of the characteristic function. Fast Fourier Transform. tion can be expressed as ek
+ 0
Carr & Madan(1999) suggest a transformation of the
vanilla pricing formula that allows to use the FFT technique. The value of a call op-
C (S0 , T, k ) =
eiku (u) du
(4.20)
where k denotes the log of the strike price, is a damping parameter and erT (u ( + 1)i) , 2 + u2 + (2 + 1)ui
(u) =
(4.21)
where
> ( ) = EQ (eix )
(4.22)
is the characteristic function of the log price x = ln(ST ). The integral in (4.20) is approximated using an integration rule
+ 0 NF F T 1
eiku (u) du
j =0
eikuj (uj )wj ,
(4.23)
uj = j,
(4.24)
25
where NF F T is the number of grid points and the weights wj implement the integration rule. The crucial limitation of the FFT method is that the grid points uj must be chosen equidistantly. This limitation prohibits the use of the most eective integration rules such as the Gaussian quadrature. The FFT pricing method simultaneously computes the values of the integral approxiFT mations (4.23) for the set of log-strikes {km = ( NF 2 ) + m, m = 0, . . . , NF F T 1}.
The simultaneous calculation for all strikes is not an exclusive advantage of the FFTbased methods, because a slightly modied direct integration method also has this advantage. This simple modication is described in the next section. The second important restriction is that the grid spacings must satisfy the condition = 2 . NF F T (4.25)
If this condition is satised, the sums in (4.23) can be expressed in the form
NF F T 1 NF F T 1 NF F T 1
eikuj (uj )wj =

j =0 j =0
eijm hj =
j =0
i( N 2
FFT
)jm
hj ,
(4.26)
which allows the application of the FFT procedure invoked on the vector h = {hj = ei(
N j ) 2
(uj )wj , j = 0, . . . , NF F T 1}.
Fractional Fast Fourier Transform. Chourdakis(2005) has shown how the method of Carr & Madan(1999) can be accelerated using the fractional FFT algorithm. This algorithm rapidly computes sums of the form
N 1
Dk (h, ) =
j =0
ei2kj hj
(4.27)
for any value of . The fractional FFT method can be applied without the need to impose the restriction (4.25). However, the fractional FFT method does not overcome the crucial limitation of the FFT method because the grid points ui still must be chosen equidistantly. Fractional FFT is implemented by invoking three FFT procedures, i.e.,
1 Dk (h, ) = (eik )N k=0
2
1 Dk (Dj (y )
Dj (z )),
(4.28)
26
4.2 Pricing methods
where
N 1 N 1 y = ((hj eij )j =0 , (0)j =0 ),
2
(4.29)
1 i (N j ) N 1 z = ((eij )N )j =0 ), j =0 , (e
(4.30)
Dk (h) denotes the FFT sum

N 1
Dk (h) =
j =0 1 Dk (h) is the inverse FFT sum
ei N kj hj ,
(4.31)
1 Dk (H ) =
1 N
N 1 j =0
ei N kj Hj ,
(4.32)
and
denotes element-by-element vector multiplication.
The fractional FFT pricing method is faster than the FFT pricing method, because the absence of the restriction (4.25) allows the use of sparser grids. This eect is more important in terms of computing time than the disadvantage of using three FFT routines instead of one.1 The accuracy of the prices calculated with the FFT or the fractional FFT methods strongly depends on the choice of the damping parameter . Using the same value of the damping parameter in all pricing situations would be fatal for the calibration procedure. In particular, if we use a reasonable grid size NF F T < 4096 there is no value of that leads to an acceptable pricing error for all possible parameter values. Lord & Kahl(2007) provide an example of two pricing inputs with non-overlapping sets of acceptable damping parameters. Only an extremely ne FFT grid will result in overlapping sets of acceptable damping parameters. But ne FFT grids are impractical because they slow down the calibration. Therefore the recommendations of Lee(2004) and Lord & Kahl(2007) for the choice of are not just an additional improvement of the FFT-based methods but a necessary requirement for the implementation of these methods.
1
See Chourdakis(2005)
27
4.3
Caching technique
The most time-consuming part of the computation is the evaluation of the characteristic function ( ). For example the characteristic function of XT in the Heston model1 1 gedT )) 1g (4.33) (4.34) (4.35)
( ) = exp{2 (( i d)T 2 ln(

2 2 + 0 ( i d)
1 edT }, 1 gedT
d = ((i )2 2 (i 2 ))1/2 , i d g= , i + d
contains two complex exponents,2 one complex logarithm and one complex square root. Therefore an extremely important requirement for an eective implementation of the calibration algorithm is the following: The number of evaluations of the characteristic function should be as low as possible. If the calibration algorithm uses the direct integration method to compute the values of vanilla options, a caching technique should be used to avoid unnecessary recalculations of the characteristic function. If the caching technique is not used, the calculation of the values of vanilla options at each iteration of the optimization algorithm includes the following steps: 1. Loop over expiries of the vanilla options. 2. Loop over strikes of the vanilla options. 3. Loop over the points i , i = 1, . . . , U that are used to evaluate the integral in (4.17) numerically. 4. Evaluate the characteristic function in i . 5. Evaluate the integrand in i . 6. Calculate the value of the vanilla option. However, the value of the characteristic function does not depend on the strike. If we use the same grid i , i = 1, . . . , U for all options and run the described algorithm,
The literature provides two specications for the characteristic function of XT in the Heston model. The rst one is used in Heston(1993). The second one can be found in Schoutens et al.(2004) or in Gatheral(2005). We use the second specication. For justication of this choice see Albrecher et al.(2007) 2 We do not count identical repeated terms.
1
28
4.3 Caching technique
we recalculate the same values of the characteristic function at each step of the strikeloop. We can use the following modication of the algorithm in order to avoid these recalculations: 1. Loop over expiries of the vanilla options. 2. Loop over strikes of the vanilla options. 3. Loop over the points i , i = 1, . . . , U that are used to evaluate the integral in (4.17) numerically. 4. If we are at the rst step of the strike-loop, evaluate the characteristic function in i and save this value in the cache. 5. If we are not at the rst step of the strike-loop, read the value of the characteristic function in i from the cache. 6. Evaluate the integrand in i . 7. Calculate the price of the vanilla option. The numerical evaluation of the integral in (4.17) requires a choice of the numerical upper integration limit. Suppose a maximum tolerable truncation error is given. Then the numerical upper integration limit depends on the maturity T and the strike K of the vanilla option: = (T, K ). We can still use the same -grid for all T and K - we just dene the index of the last integration point as an integer-valued function U (T, K ) that satises the condition U (T,K ) (T, K ) < U (T,K )+1 . (4.36)
The grid at step 3 of the algorithm can now be dened as i , i = 1, . . . , U (T, K ). In most cases the function U(T,K) is an increasing function of K. It leads to a dierent number of loop iteration at step 3 for dierent K . Therefore, we have to modify the described algorithm once more in order to take this fact into account. We can use a reverse order of strikes or we can control at each point i whether the characteristic function has been already evaluated at this point. We can also combine these two solutions. In this case the algorithm is: 1. Loop over expiries of the vanilla options. 2. Loop over strikes of the vanilla options. Use a reverse order of strikes. 3. Loop over the points i , i = 1, . . . , U (T, K ) that are used to evaluate the integral in (4.17) numerically.
29
4. If the value of the characteristic function in i is still not in the cache, evaluate it and save this value in the cache. 5. If the value of the characteristic function in i is already in the cache, use this precomputed value. 6. Evaluate the integrand in i . 7. Calculate the value of the vanilla option. There is a further possibility to accelerate this algorithm. Some terms of the characteristic function do not depend on T . These terms can be precomputed before starting the loop over expiries of the vanilla options. For example, we recommend to compute the term (4.34) only once and store it, because it contains a time-consuming square root operator. Dobranszky(2009) shows how to achieve better convergence of the calibration algorithm using control variates for the Fourier inversion.
4.4
Numerical experiment
The FFT-based methods of pricing vanilla options are very popular because they simultaneously give option values for a range of strikes. This simultaneous calculation saves computing time because the characteristic function need not to be recomputed for dierent strikes. However, the direct integration method also has this useful feature - we just have to use the caching technique described in the previous section. Therefore the possibility of simultaneous pricing for dierent strikes cannot be considered as a criterion for comparison of pricing methods.1 We have to dene other criteria for the comparison. The rst of these criteria is the speed of the numerical integration method. Obviously, there are a lot of techniques of simple numerical integration in the general case that are both faster and more accurate than integration using FFT. They are designed to minimize the number of integrand evaluations. One of these techniques is the Gaussian quadrature formula. This section shows that the grid for the numerical integration
The FFT algorithm reduces the number of multiplications in the required NF F T summations from 2 an order of NF F T to that of NF F T ln2 NF F T (Carr & Madan(1999)). However the computing time required for these multiplications is negligible in comparison with the time required for the evaluations of the characteristic function. Therefore we concentrate on the number of the calculations of the characteristic function only.
1
30
4.4 Numerical experiment
(4.17) with six-point Gaussian quadrature is at least seven times more economical than the FFT-grid in (4.20). The second criterion is the rate of decay of the integrand. The integrand in (4.17) decays at a quadratic rate. This is the main reason why we use the pricing formula from Attari(2004) rather than the formula from Heston(1993). The rate of decay of the integrand in (4.20) is also quadratic. Therefore this criterion does not indicate any advantages of Formula (4.20) relative to Formula (4.17). As we have already pointed out, the number of the evaluations of the characteristic function is the main factor driving the calibration time. We carry out a numerical experiment to compare the inuence of this factor in each pricing method. Then we conduct a second numerical experiment where we compare the calibration time directly. First, we notice that the number of evaluations of the characteristic function during the calibration procedure is equal to Ngrid NM , where NM denotes the number of maturities in the set of vanilla options which are used to calibrate a model, Ngrid is equal to NF F T for the FFT-based pricing methods,1 and Ngrid =
NM t=1
maxj U (Tt , Kj ) NM
(4.37)
for the direct integration method. Note that Ngrid in (4.37) is approximately equal to the average size of the numerical integration grid at the last strike of each calibration maturity. In order to compare the performance of the above-mentioned pricing methods, we dene some benchmark accuracy levels. We then estimate grid sizes Ngrid that lead to the desired accuracy. The results are summarized in Table 1.2 These results are based on the following numerical experiment. Values of 100 vanilla options (10 maturities from 0.1 to 5.0 years, 10 strikes for each maturity) are calculated with 100 random (but reasonable3 ) sets of parameters of the Heston model. These calculations are performed with dierent grid sizes. The results of the calculations with extremely
Application of the fractional FFT method results in a smaller grid size and in fewer evaluations of the characteristic function. 2 If the direct integration method is used, Ngrid is not necessary an integer value (see (4.37)). However we report only an integer part of Ngrid for a more natural interpretation. 3 The parameters are drawn from the following ranges: long-run variance [0.01, 1.0], meanreversion rate [1.0, 4.0], volatility of variance [0.01, 10.0], short-term volatility 0 [0.1, 1.0], correlation [1.0, 1.0].
1
31
Accuracy (implied volatility basis points)
Grid size for the FFT method
Grid size for the fractional FFT method
Numerical integration grid size for the direct integration method 96 126 162 582
2.0 1.0 0.2 0.02
4096 4096 8192 16384
1024 2048 2048 4096
Table 4.1: Grid sizes that are needed to obtain some benchmark accuracy levels ne grids are used as the benchmark for accuracy estimations.1 Suppose we obtaine a value p1 with a reasonable grid size and a price p2 with extremely ne grid size. To estimate the accuracy of the price p1 we compute Black-Scholes implied volatilities (p1 ) and (p2 ) for both prices p1 and p2 . The accuracy is dened as the absolute dierence | (p1 ) (p2 )| between these implied volatilities. Table 4.1 reports the minimum grid sizes that lead to the desired accuracy in all 10000 pricing situations.2 The same experiment is carried out for the Bates model, the Barndor-Nielsen&Shephard model and four models based on time-changed Levy processes (NIG-CIR, NIG-GOU, VG-CIR, VG-GOU).3 The results are very similar to the results reported in the Table 4.1. For all these models the grid for the fractional FFT method must be at least seven times ner than the grid for the direct integration method to obtain the same accuracy in both methods. Our second numerical experiment compares the speed of the calibration directly. We select 100 random business days from January, 2000 to November, 2006. For each of these days we use historical market data on DAX vanilla option prices as an input to the calibration routine. Each calibration input contains 80 to 155 options with 8 to 12
It is also checked that these benchmark values are identical for all three pricing methods. We also test the benchmark values against Monte-Carlo simulations. Occasionally there are some reference prices in the literature. In these cases we also compare our prices with these references (Table 3 in Lord, Koekkoek and van Dijk (2006), Figure 6 in Kahl & Jackel(2005b)). 2 10000 pricing situations correspond to 100 options and 100 parameter sets. 3 The description of all these models can be found in Schoutens et al.(2004).
1
32
4.4 Numerical experiment
Model Heston Bates BNS VG-CIR VG-GOU NIG-CIR NIG-GOU
FFT 466 620 405 540 522 546 521
Fractional FFT 239 316 208 281 269 280 273
Direct integration 15 20 13 17 17 18 17
Table 4.2: Average calibration time (in seconds). dierent maturities. We run the calibration procedure with the three dierent vanilla pricing methods applied to seven dierent models. Each calibration consists of three runs of the Dierential Evolution algorithm and three runs of the Levenberg-Marquardt algorithm. The grid sizes are set as in the second line of Table 1 (accuracy = 1.0 basis points). The average calibration time1 is compared in Table 4.2. This table shows that the calibration with the direct integration method is approximately 16 times faster than the calibration with the fractional FFT method. It corresponds to the ratio of grid sizes: 126 points for the direct integration and 2048 for the fractional FFT. However, we should take into account that we are extremely inexible in the choice of the FFT grid - the number of the FFT grid points must be a power of two. Therefore the dierence in calibration speed between the fractional FFT method and the direct integration method highly depends on the desired accuracy. For example, if the desired accuracy is 0.02 basis points, the direct integration grid size is approximately 7 times smaller than the fractional FFT grid (the last line of Table 4.1). It results in a corresponding ratio of calibration times for these methods.
The computations were done on a Centrino Pentium M, 1.5GHz CPU
33
34
Forward-start options in the Barndor-Nielsen&Shephard model

In this chapter we derive a semi-analytical formula for pricing forward-start options in the Barndor-Nielsen&Shephard model. In terms of computational time, this formula is equivalent to one-dimensional integration. The pricing of forward-start options is one of the problems where realistic modeling of volatility dynamics is of particular importance. The Barndor-Nielsen&Shephard model tries to meet this requirement. The general case of this model is introduced in Barndor-Nielsen & Shephard(2001). An empirical performance of the BarndorNielsen&Shephard model is analyzed in Tompkins(2001). Possible patterns of implied volatility surfaces generated by this model are described in Cont & Tankov(2003). A possible modication of this model and parameter estimation technique is described in Hubalek & Posedel(2008). We use the ane1 property of the Barndor-Nielsen&Shephard model in order to derive a formula for pricing forward-start options. Applications of the ane processes in derivative pricing have been studied in Due et al.(2003) and Keller-Ressel(2008). The forward-start option pricing in other stochastic volatility models has been studied in Lucic(2003), Kruse & Noegel(2005) and Bloch(2008). In Section 5.1 we derive a semi-analytical formula for pricing forward-start options
1
The denition of ane stochastic volatility models is given in Keller-Ressel(2008)
35
5. FORWARD-START OPTIONS IN THE BARNDORFF-NIELSEN&SHEPHARD MODEL
in the Barndor-Nielsen&Shephard model. Section 5.2 provides numerical examples based on this formula.
5.1
Derivation
The payo of a forward-start option is max ST k, 0 , ST0 (5.1)
where T is a maturity, T0 is a forward-start time and k is a relative strike. Some techniques of pricing plain vanilla options require only the knowledge of the characteristic function of the logarithm of the underlying1 . Since forward-start options can be seen as options on the underlying UT =
ST ST0 ,
these techniques can be extended
to forward-start options. The only modication that must be made is to replace the characteristic function of the logarithm of the underlying (u) = EQ (eiu ln ST ) by the forward characteristic function f wd (u) = EQ [e
iu ln
ST ST 0
(5.2)
].
(5.3)
Keller-Ressel(2008) has derived a general expression for the forward characteristic function of the ane models in terms of the solutions of the generalized Riccati equations. In this chapter we use this result to derive an implementable formula for the forward characteristic function of the Barndor-Nielsen&Shephard model. The risk-neutral dynamics in the general case of the Barndor-Nielsen&Shephard model is
2
d(ln St ) = dt +
Vt dWt + dJt ,
(5.4)
dVt = Vt dt + dJt ,
(5.5)
where Jt is the Levy subordinator that drives the model, is the drift that is determined by the martingale condition for St , is the mean-reversion rate, is the comovement
Examples of such techniques are described in Chapter 4. Barndor-Nielsen&Shephard model with the Gamma-Ornstein-Uhlenbeck latent state (2.6)-(2.7) is a special case of the general Barndor-Nielsen&Shephard model.
2 1
36
5.1 Derivation
parameter. In general, the cumulant generating function of ane models is dened by some functions (t, u, w) and (t, u, w) and has the form K (u, w) = ln EQ [eu ln St +wVt ] = (t, u, w) + V0 (t, u, w) + u ln S0 , (5.6)
for all u, w C, where EQ [eu ln St +wVt ] < . The forward characteristic function of ane models has the form
ST ST 0
f wd (u) = EQ [e
iu ln
] = EQ [eiu ln ST0 EQ [eiu ln ST |FT0 ]] =
EQ [exp((T T0 , iu, 0) + VT0 (T T0 , iu, 0))] = exp((T T0 , iu, 0))EQ [exp(VT0 (T T0 , iu, 0))] = exp[(T T0 , iu, 0) + (T0 , 0, (T T0 , iu, 0)) + V0 (T0 , 0, (T T0 , iu, 0))]. In the Barndor-Nielsen&Shephard model (t, u, w) = 1 2 (u u)(1 et ) + et w, 2
t
(5.7)
(5.8)
(t, u, w) =
0
F (u, (s, u, w)) ds,
(5.9)
where F (u, w) = (w + u) u(), (5.10)
and (u) is the cumulant-generating function of the Levy subordinator Jt that drives the model. In this chapter we consider an example of the Barndor-Nielsen&Shephard model where the latent state follows the Gamma-Ornstein-Uhlenbeck process. In this case the Levy subordinator Jt is a compound Poisson process
Nt
Jt =
n=1
xn ,
(5.11)
where Nt is a Poisson process with intensity a and each xn follows an exponential law with mean 1 b . The cumulant-generating function of Jt is (u) = au . bu (5.12)
37
We calculate the integral in (5.9) analytically in order to accelerate the pricing algorithm. Equations (5.10) and (5.12) yield F (u, (s, u, w)) = where 1 2 (u u) w (b + u), 2 1 2 B = u2 + u2 2 + (u u)(b + u), 2 1 2 C= (u u) w, 2 1 2 D = b u (u u). 2 A= rule yields a(es Re(A) + Re(B ) + i es Im(A) + Im(B ) ) (b )(es Re(C ) + Re(D) + i [es Im(C ) + Im(D)]) a = s (b )(e Re(C ) + Re(D))2 + (es Im(C ) + Im(D))2 [(es Re(A) + Re(B ))(es Re(C ) + Re(D))+ (es Im(A) + Im(B ))(es Im(C ) + Im(D))+ i{(es Im(A) + Im(B ))(es Im(C ) + Im(D)) (es Re(A) + Re(B ))(es Im(C ) + Im(D))}]. Let us introduce an auxiliary integral G(, , , , , , p, x) := e2px + epx + dx, , , , , , , p, x R e2px + epx + (5.16) (5.15) a(Aes + B ) , (b )(Ces + D) (5.13)
(5.14)
Separating real and complex parts in (5.13) and applying the complex numbers division
F (v, (s, v, w)) =
that can be evaluated analytically, in fact,

+2e arctan(
px ) 4 2
x G(,, , , , , p, x) = + 2 + 1 2 ln( + epx p + e2px )
p 4 2 (5.17)
C = const. + C,
38
5.2 Numerical examples
Combining (5.9) and (5.15) yields (t, v, w) = where a [G1 (t) G1 (0) + i{G2 (t) G2 (0)}] , (b ) (5.18)
G1 (x) = G(Re(A)Re(C ) + Im(A)Im(C ), Re(A)Re(D) + Re(B )Re(C ) + Im(A)Im(D) + Im(B )Im(C ), Re(B )Re(D) + Im(B )Im(D), (Re(C ))2 + (Im(C ))2 , 2Re(C )Re(D) + 2Im(C )Im(D), (Re(D))2 + (Im(D))2 , , x), (5.19)
G2 (x) = G(Im(A)Re(C ) Re(A)Im(C ), Im(A)Re(D) + Im(B )Re(C ) Re(A)Im(D) Re(B )Im(C ), Im(B )Re(D) Re(B )Im(D), (Re(C ))2 + (Im(C ))2 , 2Re(C )Re(D) + 2Im(C )Im(D), (Re(D))2 + (Im(D))2 , , x). (5.20)
Substituting (5.8) and (5.18) into (5.7) yields an analytic formula for the forward characteristic function of the Barndor-Nielsen&Shephard model.
5.2
Numerical examples
In this subsection we illustrate how the price and greeks of a forward-start call depend on each of the Barndor-Nielsen&Shephard parameters. The solid lines in Figures 5.15.5 show the prices of a forward-start call for a range of values of one parameter when all the other parameters are kept constant. The ve gures correspond to ve BarndorNielsen&Shephard parameters. The strike of the option is 1.0, the maturity is 3 years,
39
Figure 5.1: Comovement scenario. = 1.0, b = 50.0, a = 0.5, V0 = 0.2
Figure 5.2: Mean-reversion rate scenario. = 5.0, b = 50.0, a = 0.5, V0 = 0.2
40
Figure 5.3: Exponential law parameter scenario. = 5.0, = 1.0, a = 0.5, V0 = 0.2
Figure 5.4: Poisson intensity scenario. = 5.0, = 1.0, b = 50.0, V0 = 0.2
41
Figure 5.5: Initial latent state scenario. = 5.0, = 1.0, b = 50.0, a = 0.5
Figure 5.6: Forward start time scenario. = 5.0, = 1.0, b = 50.0, a = 0.5, V0 = 0.2, = 2.0
42
the forward-start time is 1 year, the interest rate is set to 0.0 and there are no dividends. Dashed lines show the values of greeks corresponding to particular parameters. Corresponding numerical values are reported in Tables 5.1-5.5. Figure 5.6 and Table 5.6 show price and theta for dierent forward start periods. To calculate theta we set T = + T0 , keep constant and calculate the derivative with respect to T0 . Reported values illustrate an important feature of theta in the BarndorNielsen&Shephard model. Zero values of delta, gamma1 and absence of interest rates does not necessarily lead to zero value of theta. This feature can be explained by the PIDE of the Barndor-Nielsen&Shephard model2 . This PIDE contains not only delta, gamma and theta, but also rst-order and second-order sensitivities to latent state.
1 2
Delta and gamma of forward-start options are approximately equal to zero. This PIDE can be found in Hilber(2005) and Schwab et al.(2007).
43
Comovement -10 -9.5 -9 -8.5 -8 -7.5 -7 -6.5 -6 -5.5 -5 -4.5 -4 -3.5 -3 -2.5 -2 -1.5 -1 -0.5
Price 0.10775 0.10470 0.10166 0.09866 0.09568 0.09276 0.08988 0.08707 0.08434 0.08171 0.07918 0.07679 0.07454 0.07247 0.07060 0.06895 0.06756 0.06646 0.06567 0.06523
Sensitivity to comovement -0.00612 -0.00608 -0.00604 -0.00598 -0.00590 -0.00580 -0.00568 -0.00554 -0.00537 -0.00516 -0.00492 -0.00464 -0.00432 -0.00395 -0.00352 -0.00304 -0.00250 -0.00190 -0.00124 -0.00052
Table 5.1: Comovement scenario. = 1.0, b = 50.0, a = 0.5, V0 = 0.2
44
Mean-reversion rate 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95 1.05 1.15 1.25 1.35 1.45 1.55 1.65 1.75 1.85 1.95
Price 0.10863 0.10189 0.09624 0.09159 0.08781 0.08483 0.08256 0.08092 0.07986 0.07929 0.07917 0.07942 0.07999 0.08082 0.08184 0.08301 0.08427 0.08560 0.08697 0.08837
Sensitivity to mean-reversion rate -0.07332 -0.06171 -0.05132 -0.04200 -0.03364 -0.02613 -0.01940 -0.01338 -0.00804 -0.00333 0.00075 0.00420 0.00705 0.00931 0.01101 0.01224 0.01304 0.01355 0.01385 0.01400
Table 5.2: Mean-reversion rate scenario. = 5.0, b = 50.0, a = 0.5, V0 = 0.2
45
Exponential law parameter 5 15 25 35 45 55 65 75 85 95 105 115 125 135 145 155 165 175 185 195
Price
Sensitivity to exponential law parameter -0.03135 -0.00696 -0.00285 -0.00149 -0.00090 -0.00060 -0.00042 -0.00031 -0.00024 -0.00019 -0.00015 -0.00012 -0.00010 -0.00009 -0.00008 -0.00006 -0.00006 -0.00005 -0.00004 -0.00004
0.30915 0.16042 0.11562 0.09491 0.08325 0.07587 0.07081 0.06713 0.06436 0.06219 0.06044 0.05902 0.05783 0.05682 0.05596 0.05521 0.05456 0.05399 0.05348 0.05302
Table 5.3: Exponential law parameter scenario. = 5.0, = 1.0, a = 0.5, V0 = 0.2
46
Poisson intensity 0.01 0.26 0.51 0.76 1.01 1.26 1.51 1.76 2.01 2.26 2.51 2.76 3.01 3.26 3.51 3.76 4.01 4.26 4.51 4.76
Price 0.04570 0.06357 0.07980 0.09420 0.10698 0.11847 0.12894 0.13859 0.14759 0.15606 0.16406 0.17167 0.17893 0.18590 0.19259 0.19904 0.20527 0.21131 0.21715 0.22283
Sensitivity to Poisson intensity 0.07381 0.06851 0.06119 0.05417 0.04834 0.04374 0.04012 0.03722 0.03486 0.03287 0.03118 0.02972 0.02843 0.02729 0.02627 0.02535 0.02451 0.02374 0.02304 0.02239
Table 5.4: Poisson intensity scenario. = 5.0, = 1.0, b = 50.0, V0 = 0.2
47
Initial latent state -10.0 -9.5 -9.0 -8.5 -8.0 -7.5 -7.0 -6.5 -6.0 -5.5 -5.0 -4.5 -4.0 -3.5 -3.0 -2.5 -2.0 -1.5 -1.0 -0.5
Price 1.18497 1.14325 1.09053 1.03630 0.98930 0.95457 0.93191 0.91630 0.90020 0.87681 0.84257 0.79737 0.74245 0.67840 0.60512 0.52277 0.43212 0.33458 0.23256 0.13130
Sensitivity to initial latent state -0.06705 -0.09739 -0.11020 -0.10368 -0.08257 -0.05642 -0.03597 -0.02908 -0.03765 -0.05711 -0.07979 -0.10049 -0.11899 -0.13728 -0.15578 -0.17336 -0.18878 -0.20059 -0.20601 -0.19345
Table 5.5: Initial latent state scenario. = 5.0, = 1.0, b = 50.0, a = 0.5
48
Forward-start time 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1.0
Price 0.09442 0.09331 0.09224 0.09121 0.09022 0.08927 0.08835 0.08746 0.08661 0.08579 0.08501 0.08425 0.08352 0.08282 0.08215 0.08151 0.08089 0.08029 0.07973 0.07918
Theta -0.02244 -0.02164 -0.02087 -0.02011 -0.01937 -0.01866 -0.01796 -0.01729 -0.01664 -0.01600 -0.01538 -0.01478 -0.01420 -0.01364 -0.01309 -0.01257 -0.01205 -0.01156 -0.01108 -0.01061
Table 5.6: Forward start time scenario. = 5.0, = 1.0, b = 50.0, a = 0.5, V0 = 0.2, = 2.0
49
50
On the cost of poor volatility modeling The case of cliquets

In this chapter we conduct pricing and hedging experiments in order to check whether simple stochastic volatility models are capable of capturing the forward volatility and forward skew risks correctly. As a reference we use the Bergomi model that treats these risks accurately per denition. Results of our experiments show that the cost of poor volatility modeling in the Heston model, the Barndor-Nielsen&Shephard model and a Variance-Gamma model with stochastic arrival is too high when pricing and hedging cliquet options. This chapter is structured as follows: Denitions of forward volatility and forward skew are introduced in Section 6.1. Examples and properties of cliquet options are discussed in Section 6.2. We conduct a price comparison in Section 6.3. The performance of hedge strategies based on simpler popular models is analyzed in Section 6.4.
6.1
Forward volatility and forward skew
Analysis of inuence of implied volatility dynamics on pricing and hedging exotic options should take into account the sensitivity of a particular product to forward volatility and forward skew. However, we should be careful when using the terms forward volatility and forward skew. Sometimes these terms are used without dening them. It can be confusing since there are alternative non-equivalent denitions of forward volatility. We should distinguish between these denitions. We should also distinguish
51
6. ON THE COST OF POOR VOLATILITY MODELING THE CASE OF CLIQUETS
between forward-vol-sensitive options and forward-skew-sensitive options. We may give the following non-equivalent denitions of forward volatility (forward implied volatility): Denition 1. Forward volatility is the implied volatility for a relative strike k and maturity T2 that is observed at a future time-point T1 . This value is unknown today. It is also model-independent: one needs no model to observe this value - one needs to wait only until time T1 . We call this forward volatility the future volatility. Denition 2. Forward volatility is the Black-Scholes volatility implied from a price of a forward-start call computed in another model. In other words: At time t compute the price CF of a forward-start call maturing at T2 and relative strike k , which will be
T1 ,T2 xed at T1 , in some model. Find the volatility f (t, k ), such that the Black-Scholes T1 ,T2 price of this forward-start call with the volatility f (t, k ) is equal to CF . The value T1 ,T2 f (t, k ) is forward volatility. This value is known today and it is model-dependent.
We will refer to this forward volatility as the forward-start-vanilla-implied forward volatility. A similar denition of forward volatility is given in Keller-Ressel(2008). Denition 3. Forward volatility is the forward variance swap volatility process
T1 ,T2 f (t) =
T1 ,T2 (t) =
(T2 t)VtT2 (T1 t)VtT1 , T2 T1
where VtT1 and VtT2 are the implied variance swap variances, t < T1 < T2 . The starting
T1 ,T2 value f (0) of this process is known today. This denition is convenient for modeling
and is used in the Bergomi model described below. We will refer to this forward volatility as the variance-swap-implied forward volatility. Although these denitions are not equivalent, all these denitions are suited for the denition of the term forward-volatility-sensitive option and reect the same qualitative aspect of the implied volatility dynamics. Denitions 1 and 2 can be adapted for the denition of forward skew. We can dene intuitively: f orward skew = or exactly: f orward skew =
T1 ,T2 df (t, k ) T1 ,T2 T1 ,T2 f (t, k2 ) f (t, k1 )
k2 k1
d ln k
,
F (t,T ) k= F (t,T2 ) 1
52
6.2 Cliquet options
where f is the forward volatility and F (t, ) is the forward price for maturity calculated at time point t. This reects the slope of the smile curve as a function of the strike and is related to the risk reversal quotes in the market. A simple example of a forward-volatility-sensitive option is a forward-start call. A simple example of a forward-skew-sensitive option is a forward-start call spread. Further examples of forward-volatility- and forward-skew-sensitive options are introduced in the next section.
6.2
Cliquet options
A cliquet option is a derivative that pays o some function of a set of relative returns of an underlying. Typically this function incorporates local or global caps and oors, minimum or maximum functions, sums and xed coupons. The relative returns are typically calculated on a monthly, semi-annual or annual basis. Wilmott(2002) shows an example where the sensitivity of a cliquet option to a deterministic volatility is negligible in comparison with the real sensitivity of this option to volatility dynamics. To show this fact, he considers a globally oored, locally capped cliquet and analyzes a model where the actual volatility is chosen to vary in such a way as to give the option its highest or lowest possible value. This model exploits the property that an increase in volatility leads to an increase of the option value when gamma is positive and to a decrease of the option value when gamma is negative. Schoutens et al.(2004) compare prices of cliquet options in seven stochastic volatility models calibrated to the implied volatility surface of the Eurostoxx 50 index. They observe a price range of more than 40 percent amongst these models. They show that dierent models can produce almost the same marginal distribution of the underlying, but at the same time totally dierent cliquet prices. They demonstrate how the negrain structure of the underlying process inuences the exotic option values. Here we introduce denitions of particular cliquet options that are referred to in this chapter. Reverse Cliquet. The payo of a reverse cliquet is
N 1
max 0, C +
i=0
ri
(6.1)
53
where ri =
ri
STi+1 STi , STi = min(ri , 0),
0 = T0 < T1 < T2 < ... < TN , C > 0. In this chapter, all numerical examples for the reverse cliquet are calculated using the same contract specication as in Bergomi(2005). The length of the reset period Ti+1 Ti is one month. The number of reset periods is N = 36, therefore the maturity is three years. The maximum possible payo is equal to the coupon C = 50%. This option is called reverse cliquet because the nal payo depends on negative returns only. This option is both forward-volatility- and forward-skew-sensitive. The forward-skewsensitivity of the reverse cliquet can be explained intuitively. If
N 2
C+
i=0
ri > 0,
(6.2)
the value of the reverse cliquet in last period is equal to the value of a call-spread option. Therefore the value of the corresponding forward-start call spread has an eect on the value of the whole structure. Consequently, the price of the reverse cliquet depends on the dierence of two forward-start-vanilla-implied forward volatilities with two strikes corresponding to strikes of this forward-start call spread. This dierence is the forward skew according to Denition 2 in Section 6.1. Napoleon. This contract consists of several building blocks. The payo of each building block, which is settled individually, is max(0, C + min ri ),
i=0,N 1
(6.3)
where ri = 0 STi+1 STi , STi T0 < T1 < T2 < ... < TN ,
C > 0. In this chapter we consider an example of a Napoleon option that consists of three building blocks. Each building block has N = 12 reset periods. The length of each
54
6.2 Cliquet options
reset period Ti+1 Ti is one month. Therefore the maturity of this contract is three years and the possible payments occur at the end of each year. The maximum possible payment at the end of each year is equal to the coupon C = 8%. This contract type is analyzed in Bergomi(2004), Bergomi(2005) and Gatheral(2005). Numerical experiments in Bergomi(2005) show that this option is extremely forwardvolatility-sensitive but almost forward-skew-insensitive. Accumulator. The payo of an accumulator is
N 1
max 0,
i=0
max(min(ri , cap), oor) ,
(6.4)
where ri = STi+1 STi , STi 0 = T0 < T1 < T2 < ... < TN .
In this chapter we consider an example of an accumulator where the oor is set to -1%, the cap is 1%, each reset period Ti+1 Ti is one month and the maturity of the option is three years(N = 36). This option is forward-skew-sensitive. This contract can also be forward-volatility-sensitive but only in cases of strong forward-skew. The intuition for this behavior is similar to the intuition for skew- and volatility- sensitivity of a standard one month call-spread. In the Black-Scholes model this call-spread has negligible vega. However, the absolute value of vega of the call-spread increases in the models that takes skew into account and in the presence of skew. Call spread cliquet. This option pays at the end of each reset period [Ti , Ti+1 ] the amount STi+1 k1 , 0 max STi STi+1 k2 , 0 . STi
max
(6.5)
A call spread cliquet can be seen as a portfolio of forward start call spreads. As in the previous examples we consider N = 36 monthly reset periods Ti+1 Ti . The strikes are set to k1 = 0.95 and k2 = 1.05. This option inherits forward-skew-sensitivity from its forward-start call-spread building blocks.
55
6.3
Price comparison
In this section we compare the theoretical prices of cliquet options in the Bergomi model and in calibrated versions of the discussed simpler models. We take the Bergomi model to be the true data-generating process. This assumption allows us to gauge the dierences between this recent model and simpler popular models. The parameter values we use are reported in Tables 6.1-6.2. Initial implied variance swap variances are
2 generated in the Heston model with the parameters 0 Heston , Heston and Heston . The
meaning of these parameters is described in Subsection 2.1 about the Heston model. Scenario 1 2 3 4 5 1.0 1.0 1.0 1.4 1.5 0.1 0.5 0.2 0.3 0.3 k1 5.0 5.0 4.8 6.0 6.0 k2 0.4 0.4 0.4 0.25 0.25 0.2 0.2 0.2 0.0 0.0 -0.2 0.2 0.2 0.1 0.2
Table 6.1: Values of parameters , , k1 , k2 , and . Scenario 1 2 3 4 5 -0.06 -0.06 -0.06 -0.07 -0.0625 SX -0.6 -0.6 -0.6 -0.7 -0.7 SY -0.3 -0.3 -0.3 -0.35 -0.35 Heston 0.04 0.09 0.09 0.05 0.09
2 0 Heston
Heston 1.0 1.0 1.0 1.0 1.0
0.04 0.04 0.04 0.04 0.04
2 Table 6.2: Values of parameters , SX , SY , Heston , 0 Heston .
We use parameters reported in Tables 6.1-6.2 for the Bergomi model to generate ve implied volatility surfaces. Then we calibrate the simpler models to these implied volatility surfaces and compare price dierences. The calibrated parameters for the simpler models are reported in Tables 6.3-6.5. We use the calibration algorithm described in Chapter 4. Of course, the calibrated values of mean-reversion rate, long-run variance and shortterm volatility in the Heston model1 should not coincide with the corresponding param1
See Table 6.3.
56
6.3 Price comparison
eters that we use to generate initial implied variance swap variances. We cannot expect such a correspondence, because we use the set of vanilla options as a calibration input. In the absence of jumps1 a set of prices of vanilla options with all possible maturities and strikes contains more information than a set of prices of variance swaps with all possible maturities. Scenario 1 2 3 4 5 0.308888 0.0901768 0.0909267 0.053176 0.091009 1.524 1.0 1.0 3.53349 1.12328 2.57848 0.231635 0.149329 0.521236 0.277384 0 0.251838 0.194986 0.193474 0.201211 0.194058 -0.659371 -0.514877 -0.609845 -0.67917 -0.605952
Table 6.3: Calibrated Heston parameters. Scenario 1 2 3 4 5 -5.17388 -2.04324 -2.86624 -2.7184 -1.99725 0.0403766 0.83257 0.804772 2.0589 0.905518 b 11.3553 49.3421 93.3514 47.182 36.6755 a 9.98816 4.00691 7.78661 1.53604 2.85488 0 0.164977 0.174382 0.169629 0.159252 0.168391
Table 6.4: Calibrated Barndor-Nielsen&Shephard parameters. Scenario 1 2 3 4 5 -0.19297 -0.526388 -0.76172 -0.348248 -0.413335 0.256868 0.19289 0.184905 0.197137 0.193394 0.273051 0.0210463 0.0153841 0.0573394 0.033752 1.47153 1.62598 1.58304 3.27292 1.7588 1.56324 0.941867 0.912081 0.712477 0.964997 3.98489 0.726751 0.500934 1.76894 1.01945 y0 1.0 1.0 1.0 1.0 1.0
Table 6.5: Calibrated VGSA parameters. Calibrated values of the mean-reversion rate in the Heston model for scenarios 2 and 3 are exactly equal to 1.0. That is explained by the parameter bounds that are used to calibrate the Heston model. For the mean-reversion rate we use the bounds
1
Both Heston and Bergomi processes do not have jumps.
57
[1.0; 4.0] which correspond to the time interval between three months and one year. Values of cliquet options in all described models are reported in Tables 6.6-6.10. These values are calculated using a Monte-Carlo simulation with one million paths using antithetics as a variance reduction technique and a path generation of the variance process based on Andersen & Brotherton-Ratclie(2005) in order to prevent the variance process to take negative values1 . Using these values we generate 20 gures (5 scenarios x 4 instruments) that allow us to compare between dierent models. In 16 of these 20 gures the best approximation of the reference (Bergomi) price is the value calculated in the Heston model. However, in more than half of the pricing experiments the relative dierence between reference price and value calculated in the Heston model is more than 3.5%, which is unacceptable for practical applications. Bergomi Napoleon Reverse Cliquet Accumulator Call Spread Cliquet 0.0595 0.0258 0.0666 2.0191 Heston 0.0458 0.0401 0.0796 1.9389 BNS 0.0331 0.0203 0.0424 1.9277 VGSA 0.0727 0.0880 0.1821 2.0165
Table 6.6: Comparison of theoretical cliquet values. Scenario 1.
Bergomi Napoleon Reverse Cliquet Accumulator Call Spread Cliquet 0.0153 0.0031 0.0228 1.7776
Heston 0.0148 0.0025 0.0231 1.7929
BNS 0.0176 0.0048 0.0252 1.8087
VGSA 0.0274 0.0144 0.0446 1.8716
6.4
Hedge performance
Practical implementations of the hedge strategies corresponding to the theoretical values analyzed in Section 6.3 necessarily involve a number of approximations. The two main sources of discrepancies between theoretical and realized hedge performance are
In our experiments the highest standard error for reverse cliquets was 0.00002, for Napoleons 0.00004, for accumulators 0.00004, for call spread cliquets 0.0002.
1
58
6.4 Hedge performance
Heston 0.0113 0.0016 0.0216 1.7830
BNS 0.0132 0.0020 0.0239 1.8004
VGSA 0.0254 0.0115 0.0410 1.8619
Heston 0.0279 0.0110 0.0382 1.8658
BNS 0.0374 0.0245 0.0480 1.9050
VGSA 0.0433 0.0422 0.0887 1.9323
Heston 0.0163 0.0034 0.0254 1.8068
BNS 0.0216 0.0084 0.0286 1.8294
VGSA 0.0305 0.0199 0.0555 1.9003
59
discrete rehedging and the practice of recalibrating the model used to compute the hedge ratios. In this section we analyze the magnitude of the discrepancies caused by these two sources. We concentrate on the hedge performance of the Heston model only. There are two reasons for this choice. First of all, the results of the previous section have shown that the dierences between cliquet values in the Heston model and in the reference (Bergomi) model are less than the corresponding dierences for other models. Secondly, our analysis of the hedge performance requires a huge number of recalculations of cliquet values. It would be extremely time-consuming, if each value recalculation was done using the Monte-Carlo method. Fortunately, there is a much faster method to calculate the values of call spread cliquets in the Heston model. Lucic(2003) shows how to calculate the value of a forward-start option in the Heston model in a few milliseconds. Since a call spread cliquet can be seen as a portfolio of forward start call spreads, the same pricing method can be applied to calculate values of call spread cliquets. The most natural way to evaluate the performance of a hedge strategy is to consider the realized prot-and-loss distribution. We construct such distributions by simulating the following implementations of the hedge strategies prescribed by the Heston model: 1. Hedging with constant parameters (HCP): The Heston model is calibrated to the initial implied volatility surface and the subsequent hedge adjustments are done holding the calibrated parameters constant through time. This is in line with the assumptions underlying the Heston model. 2. Hedging with recalibration (HR): The Heston model is recalibrated after each increment in the Bergomi model. This is in line with standard practice. To construct such distributions, we conduct simulation experiments where each iteration has the following structure: 1. Simulate an increment in a realization of the Bergomi model. 2. If relevant, calibrate the simpler models to the resulting implied volatility surface. 3. Adjust hedges according to the, possibly recalibrated, Heston model. The hedges are adjusted on a weekly basis. 4. At expiry, record hedge errors.
60
We investigate the performance of two dynamic hedging strategies 1. Delta and short-term vega (DSV), 2. Delta and parallel shift vega (DPV), where short-term vega is the sensitivity of the option price to the parameter 0 of the Heston model, parallel shift vega is the sensitivity of the option price to a simultaneous shift of the parameters 0 and so that 0 = + . (6.6)
The comparison between HCP and HR implementations is done on the basis of the DSV strategy. The DPV strategy is analyzed using the HCP implementation. Histograms of absolute cumulative hedging errors in the resulting three experiments (HR-DSV, HCPDSV, HCP-DPV) are shown in Figures 6.1-6.3. The total number of hedge scenarios is 100 in the experiment HR-DSV, 299 in HCP-DSV, 2782 in HCP-DPV. The underlying path is generated using parameters of scenario 5 (see Tables 6.1-6.2). The option that is hedged is a 36-periods 95-105% call spread cliquet with monthly resets. The value of the call spread cliquet in the Heston model at the issue date was 1.8068 in all experiments. Relative hedging errors are shown in percent of the cliquet value. Relative frequencies are shown in percent of the total number of experiments. The experiment HCP-DPV shows better performance than HR-DSV and HCP-DSV.1 The strategy HCP-DSV performs the worst. However, all experiments show that the hedging error can be unacceptably high. This observation conrms the assertion that the risk of using Hestons model for hedging cliquet options is too high.
We use the variance of the realized relative hedging errors to compare the performance of the strategies.
61
Figure 6.1: Experiment HR-DSV. Histogram of relative cumulative hedging errors of a call spread cliquet.
62
Figure 6.2: Experiment HCP-DSV. Histogram of relative cumulative hedging errors of a call spread cliquet.
63
Figure 6.3: Experiment HCP-DPV. Histogram of relative cumulative hedging errors of a call spread cliquet.
64
Conclusion
We have shown that an ecient implementation of the direct integration method results in a sizable speed up of the calibration of stochastic volatility models. This method even outperforms the calibration with the fractional FFT. The simultaneous pricing of options with dierent strikes is not an exclusive advantage of the FFT methods compared to the direct integration method, because an application of a cache technique leads to simultaneous pricing of options with dierent strikes in the framework of direct integration. Taking this into account we argued that the pricing methods dier in two aspects only: the numerical integration technique and the pricing formula. The combination of these factors results in higher calculation speed of the direct integration method in comparison to the FFT and fractional FFT methods. Specically: (1) Gaussian quadrature is a much faster numerical integration technique than the FFT, (2) the transformed pricing formula of Attari (2004) provides approximately the same rate of decay of the integrand in comparison with the main formula of the FFT method. As we have pointed out in Chapter 4, the direct integration method is frequently criticized in the literature. However this critique is valid only if we consider an unoptimized implementation of the general formula. The use of the modied pricing formula and the caching technique makes the direct integration method the best choice for practical applications. The second innovative result which is obtained in this thesis is the pricing formula for forward-start options in the Barndor-Nielsen&Shephard model. The pricing of this type of options in the general form of the Barndor-Nielsen&Shephard model re-
65
7. CONCLUSION
quires numerical integration1 for each evaluation of the forward characteristic function. Together with the integration in the Fourier space this yields computational times equivalent to two-dimensional integration. However, if we use a particular form of the Barndor-Nielsen&Shephard model with the latent state following the GammaOrnstein-Uhlenbeck process, an analytical expression for the forward characteristic function is available. It reduces the computational time for pricing forward-start options to a time equivalent to one-dimensional integration, i.e. just a few milliseconds on a standard PC. This particular form of the Barndor-Nielsen&Shephard model is not the only one where such an analytical simplication for the forward-start options is available. Each choice of the latent state dynamics however requires its own formal analysis similar to the one described in Chapter 5. This is a possible direction of further research in the area of ecient pricing in the Barndor-Nielsen&Shephard model. Hedging experiments described in this thesis show that the cost of poor volatility modeling in popular stochastic volatility models (Heston, Barndor-Nielsen&Shephard, Levy with stochastic clock) is too high. A possible cause of relatively high cliquet hedging errors produced by these models is that they are developed to t vanilla prices and to control forward volatility and forward skew simultaneously. A lot of modeling and numerical eort is spent to reach all these targets in one model. However, to our knowledge, today there exists no model that attains both these targets simultaneously. Therefore, we would recommend to change the requirements for the model that should be used for pricing and hedging cliquet options. Specically, we recommend to abandon the requirement to t vanilla prices. This would facilitate direct modeling of forward volatility and forward skew. When pricing and hedging cliquets, accurate modeling of forward volatility, forward skew and vega-hedging are incomparably more important than tting vanilla prices. Therefore we expect better hedging performance if we do not complicate models of smile dynamics by the requirement to t vanilla prices. Bergomis model is a good example of this approach. In our further research we plan to investigate whether it is possible to obtain acceptable cliquet hedging performance using direct forward smile modeling without tting vanilla prices.
See formula (5.9).
66
Bibliography
Albrecher, H., Mayer, P., Schoutens, W. & Tistaert, J. (2007). The little Heston trap. Wilmott Magazine , 8392. Andersen, L. (2007). Ecient simulation of the Heston stochastic volatility model. Available at SSRN: http://ssrn.com/abstract=946405 . Andersen, L. & Andreasen, J. (2001). Factor dependence of Bermudan swaption prices: fact or ction? Journal of Financial Economics , 62, 333. Andersen, L. & Brotherton-Ratcliffe, R. (2005). Extended Libor market models with stochastic volatility. Journal of Computational Finance , 9, 140. Attari, M. (2004). Option pricing using Fourier transforms. Available at
http://ssrn.com/abstract=520042 . Bakshi, G. & Madan, D. (2000). Spanning and derivative-security valuation. Journal of Financial Economics , 55, 20538. Bakshi, G., Cao, C. & Chen, Z. (1997). Empirical performance of alternative option pricing models. Journal of Finance , 52, 20032049. Barndorff-Nielsen, O. & Shephard, N. (2001). Non-Gaussian OrnsteinUhlenbeck-based models and some of their uses in nancial economics. Journal of the Royal Statistical Society , B 63, 167241. Bates, D. (1996). Jump and stochastic volatility: Exchange rate processes implicit in Deutsche Mark options. Review of Financial Studies , 9, 69107. Ben Hamida, S. & Cont, R. (2005). Recovering volatility from option prices by evolutionary optimization. Journal of Computational Finance , 8(4), 4376.
67
BIBLIOGRAPHY
Bergomi, L. (2004). Smile dynamics. Risk , 9, 117123. Bergomi, L. (2005). Smile dynamics II. Risk , 10, 6773. Bloch, D.A. (2008). Expanding forward starting options. Available at SSRN: http://ssrn.com/abstract=1138162 . Borovkov, K. & Novikov, A. (2002). On a new approach to calculating expectations for option pricing. Journal of Applied Probability , 39, 889895. Brenner, M. & Subrahmanyam, M. (1988). A simple formula to compute the implied standard deviation. Financial Analysts Journal , 44(5), 8083. Broadie, M. & Kaya, O. (2004). Exact simulation of option greeks under stochastic volatility and jump difusion models. in R.G. Ingalls, M.D. Rossetti, J.S. Smith and B.A. Peters(eds.), Proceedings of the 2004 Winter Simulation Conference . Broadie, M. & Kaya, O. (2006). Exact simulation of stochastic volatility and other ane jump diusion processes. Operations Research , 54, 217231. Buhler, H. (2009). Volatility and dividends - volatility modelling with cash dividends and simple credit risk. Available at SSRN: http://ssrn.com/abstract=1141877 . Carr, P. & Lee, R. (2007). Realized volatility and variance: Options via swaps. Risk , 5, 7683. Carr, P. & Madan, D. (1999). Option valuation using the Fast Fourier Transform. Journal of Computational Finance , 3, 463520. Carr, P., Geman, H., Madan, D. & Yor, M. (2003). Stochastic volatility for Levy processes. Mathematical Finance , 13, 345382. Chourdakis, K. (2005). Option pricing using the fractional FFT. Journal of Computational Finance , 8(2), 118. Cont, R. & Tankov, P. (2003). Financial modelling with jump processes . Chapman & Hall/CRC. Detlefsen, K. (2005). Hedging exotic options in stochastic volatility and jump diusion models. Master thesis .
68
BIBLIOGRAPHY
Detlefsen, K. & Hardle, W.K. (2006). Calibration risk for exotic options. Available at http://sfb649.wiwi.hu-berlin.de . Dobranszky, P. (2009). Option pricing using numerically evaluated characteristic functions. Presentation. Conference on Numerical Methods in Finance . Driessen, J., Klaassen, P. & Melenberg, B. (2000). The performance of multifactor term structure models for pricing and hedging caps and swaptions. Working Paper, Department of Econometrics and CentER, Tilburg University . Duffie, D., Filipovic, D. & Schachermayer, W. (2003). Ane processes and applications in nance. Annals of Applied Probability , 13, 9841053. Dumas, B., Fleming, J. & Whaley, R. (1998). Implied volatility functions: Empirical tests. Journal of Finance , 53(6), 20592106. Eberlein, E. & Madan, D. (2007). Sato processes and the valuation of structured products. Available at SSRN: http://ssrn.com/abstract=957167 . Eberlein, E., Glau, K. & Papapantoleon, A. (2008). Analysis of valuation formulae and applications to exotic options in Levy models. Available at http://arxiv.org/abs/0809.3405 . El Karoui, N., Peng, S. & Quenez, M. (1997). Backward stochastic dierential equations in nance. Mathematical Finance , 7, 177. Engelmann, B., Fengler, M., Nalholm, M. & Schwendner, P. (2006a). Static versus dynamic hedges: An empirical comparison for barrier options. Review of Derivatives Research , 9, 239264. Engelmann, B., Fengler, M.R. & Schwendner, P. (2006b). Better than its reputation: An empirical hedging analysis of the local volatility model for barrier options. Available at SSRN: http://ssrn.com/abstract=935951 . Feinstein, S. (1989). The Black-Scholes formula is nearly linear in sigma for at-themoney options: Therefore implied volatilities from at-the-money options are virtually unbiased. Federal Reserve Bank of Atlanta . Feoktistov, O. (2006). Dierential Evolution. In search of solutions . Springer.
69
BIBLIOGRAPHY
Gatheral, J. (2005). The volatility surface: A practioners guide . Wiley Finance. Gill, P. & Murray, W. (1978). Algorithms for the solution of the nonlinear leastsquares problem. SIAM Journal on Numerical Analysis , 15(5), 977992. Gupta, A. & Subrahmanyam, M. (2000). An examination of the static and dynamic performance of interest rate models in the dollar cap-oor markets. Working Paper, Weatherhead School of Management, Case Western Reserve University . Heston, S. (1993). A closed-form solution for options with stochastic volatility with applications to bond and currency options. Review of Financial Studies , 6, 32744. Hilber, der nancial N. (2005). stochastic Sparse with wavelet methods for option Zurich. pricing on Available unat
Levy
volatility jump
models.
Presentation. ETH
Workshop
modelling
processes.
http://www.quam.polytechnique.fr/AMAMEF/abstracts/hilber.pdf . Hubalek, F. & Posedel, P. (2008). Joint analysis and estimation of stock prices and trading volume in Barndor-Nielsen and Shephard stochastic volatility models. Available at arXiv: http://arxiv.org/abs/0807.3464 . Hull, J. & Suo, W. (2002). A methodology for assessing model risk and its application to the implied volatility function model. Journal of Financial and Quantitative Analysis , 37, 297318. Kahl, C. & Jackel, P. (2005a). Fast strong approximation Monte-Carlo schemes for stochastic volatility models. Working Paper, ABN AMRO and University of Wuppertal . Kahl, C. & Jackel, P. (2005b). Not-so-complex logarithms in the Heston model. Wilmott Magazine , 94103. Keller-Ressel, M. (2008). Moment explosions and long-term behavior of ane stochastic volatility models. Available at http://arxiv.org/abs/0802.1823 . Kruse, S. & Noegel, U. (2005). On the pricing of forward starting options in Hestons model on stochastic volatility. Finance and Stochastics , 9, 233250.
70
BIBLIOGRAPHY
Lee, R. (2004). Option pricing by transform methods: extensions, unication and error control. Journal of Computational Finance , 7(3), 5186. Lewis, A. (2001). A simple option formula for general jump-diusion and other exponential Levy processes. Available at http://www.optioncity.net . Longstaff, F.A., Santa-Clara, P. & Schwartz, E.S. (2001). Throwing away a billion dollars: the cost of suboptimal exercise strategies in the swaptions market. Journal of Financial Economics , 62, 3966. Lord, R. & Kahl, C. (2006). Why the rotation count algorithm works. Available at http://ssrn.com/abstract=921335 . Lord, R. & Kahl, C. (2007). Optimal Fourier inversion in semi-analytical option pricing. Journal of Computational Finance , 10(4), 130. Lucic, V. (2003). Forward start options in stochastic volatility models. Wilmott Magazine , 7274. Madsen, K., Nielsen, H. & Tingleff, O. (2004). Methods for non-linear least squares problems . Technical University of Denmark. Maruhn, J. (2009). A projection-based algorithm for the calibration of nancial market models. Presentation. Conference on Numerical Methods in Finance . Mercurio, F. & Morini, M. (2009). A note on hedging with local and stochastic volatility models. Available at SSRN: http://ssrn.com/abstract=1294284 . Mikhailov, S. & Nogel, U. (2003). Hestons stochastic volatility model. Implementation, calibration and some extensions. Wilmott Magazine , 7494. Nalholm, M. & Poulsen, R. (2006). Static Hedging and Model Risk for Barrier Options. Journal of Futures Markets , 26, 449463. Poulsen, R., Schenk-Hoppe, K. & Ewald, C. (2007). Risk minimization in stochastic volatility models: Model risk and empirical performance. Available at SSRN: http://ssrn.com/abstract=964739 .
71
BIBLIOGRAPHY
Price, K., Storn, R. & Lampinen, J. (2005). Dierential evolution. A practical approach to global optimization . Springer. Raible, S. (2000). Levy processes in nance: theory, numerics, and empirical facts . Ph.D. thesis, University of Freiburg. Schoutens, W., Simons, E. & Tistaert, J. (2004). A perfect calibration! Now what? Wilmott Magazine , 6678. Schwab, C., Hilber, N., Reich, N. & Winter, C. (2007). Numerical derivative pricing in non-BS markets. Presentation. Workshop PDE in Finance. Marne-laVallee . Smith, R.D. (2007). An almost exact simulation method for the Heston model. Journal of Computational Finance , 11(1), 115125. Tompkins, R. (2001). Stock index futures markets: stochastic volatility models and smiles. Journal of Futures Markets , 21(1), 4378. van Haastrecht, A. & Pelsser, A. (2008). Ecient, almost exact
simulation of the Heston stochastic volatility model. Available at SSRN: http://ssrn.com/abstract=1131137 . Wilmott, P. (2002). Cliquet options and volatility models. Wilmott Magazine , 7883. Zhu, J. (2000). Modular pricing of options: An application of Fourier analysis . Springer. Zhu, J. (2008). A simple and exact simulation approach to Heston model. Available at SSRN: http://ssrn.com/abstract=1153950 .
72
Curriculum vitae
Fiodar Kilin holds a diploma in applied mathematics from the Belarus State University and a Ph.D. in quantitative nance from the Frankfurt School of Finance and Management. He worked ve years at Quanteam AG as a senior consultant. He currently consults nancial institutions in the area of quantitative nance. His main research interests are stochastic volatility models and exotic equity derivatives.
73
Declaration
I herewith declare that I have produced this paper without the prohibited assistance of third parties and without making use of aids other than those specied; notions taken over directly or indirectly from other sources have been identied as such. This paper has not previously been presented in identical or similar form to any other German or foreign examination board. The thesis work was conducted from March 1, 2006 to June 14, 2009 under the supervision of Professor Uwe Wystup at the Frankfurt School of Finance & Management.
Frankfurt am Main, June 14, 2009

KILIN ForwardVolatilitySensitiveDerivatives

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

KILIN ForwardVolatilitySensitiveDerivatives

Uploaded by

Copyright:

Available Formats

Numerical advances in pricing forward volatility sensitive equity derivatives

6.5 6.6 6.7 6.8 6.9

6.10 Comparison of theoretical cliquet values. Scenario 5. . . . . . . . . . . .

BNS (Barndor-Nielsen&Shephard) latent state) Heston 3 5

2.3 Barndor-Nielsen&Shephard model with the Gamma-Ornstein-Uhlenbeck latent state

Barndor-Nielsen&Shephard model with the GammaOrnstein-Uhlenbeck latent state

2 d(ln St ) = (r k [] t /2)dt + t dWt + dJt ,

with the latent state following the process

Levy models with stochastic time

where Xt is a Levy process, Yt is a business time Yt = ys ds, (2.11)

2.5 Bergomi model

are used as are

an input to the Bergomi model. Initial forward variances

Ti (t0 ), i = 0, ..., N is modeled as

2 Bt = e2k1 (T t) E Xt + 2 e2k2 (T t) E Yt2 + 2e(k1 +k2 )(T t) E [Xt Yt ] ,

2.5 Bergomi model

Pricing exotic options in stochastic volatility models

3. PRICING EXOTIC OPTIONS IN STOCHASTIC VOLATILITY MODELS

3.2 Calibration and pricing of vanilla options

Calibration and pricing of vanilla options

A locally risk-minimizing hedge is described in El Karoui et al.(1997) and Bakshi et al.(1997)

3. PRICING EXOTIC OPTIONS IN STOCHASTIC VOLATILITY MODELS

Paper Gatheral(2005) Cont& Tankov(2003)

Instruments barrier options, cliquets, general analysis

Recommended models Heston Bates

barrier options, forward-start options general analysis cliquets

Mercurio& Morini(2009) Bergomi(2004)

3.2 Calibration and pricing of vanilla options

eiu L [v ] (R + iu) (iR u) du,

ezx v (x) dx,

( ) is the characteristic function of XT ( ) = EQ (eiXT ), where XT = ln (3.5)

3. PRICING EXOTIC OPTIONS IN STOCHASTIC VOLATILITY MODELS

3.3 Monte-Carlo pricing of exotic options

Monte-Carlo pricing of exotic options

V (u) dWV (u) =

3. PRICING EXOTIC OPTIONS IN STOCHASTIC VOLATILITY MODELS

(3) Draw a sample of log-spot conditional on V (t + ) and ln S (t + ) = ln S (t) + (V (t + ) V (t) ) t+ 1 + V (u) du + 1 2 2 t

(u) du: (3.10)

V (u) dWV (u).

Analytical pricing of exotic options

2 Rt =u 2 i ,tj ti+1 <tn <tj

Stn ln Stn1 St ln j Sti

3.4 Analytical pricing of exotic options

3. PRICING EXOTIC OPTIONS IN STOCHASTIC VOLATILITY MODELS

Accelerating the calibration of stochastic volatility models

4. ACCELERATING THE CALIBRATION OF STOCHASTIC VOLATILITY MODELS

2 3 = J iT + T ((1 + J )i exp(J (i/2)(i 1)) 1),

The variable XT is dened by (3.6).

4.1 Characteristic functions

f1 = iu 1 (u2 + iu)(1 eT )/2, f2 = iu 1 (u2 + iu)/2.

4. ACCELERATING THE CALIBRATION OF STOCHASTIC VOLATILITY MODELS

The characteristic exponent of the Variance Gamma process is c( ) =

l(K ) = ln and K denotes the strike price.

4.2 Pricing methods

Carr & Madan(1999) suggest a transformation of the

eikuj (uj )wj ,

4. ACCELERATING THE CALIBRATION OF STOCHASTIC VOLATILITY MODELS

eikuj (uj )wj =

(uj )wj , j = 0, . . . , NF F T 1}.

4.2 Pricing methods

Dk (h) denotes the FFT sum

denotes element-by-element vector multiplication.

4. ACCELERATING THE CALIBRATION OF STOCHASTIC VOLATILITY MODELS