On the pricing and hedging

of volatility derivatives
SAM HOWISON
Mathematical Institute, University of Oxford, 24 29 St. Giles,
OX1 3LB Oxford, UK
Email: howison@maths.ox.ac.uk
AVRAAM RAFAILIDIS
Department of Mathematics, Kings College London,
Strand, London WC2R 2LS, UK
Email: avraam@mth.kcl.ac.uk
HENRIK RASMUSSEN
Mathematical Institute, University of Oxford, 24 29 St. Giles,
OX1 3LB Oxford, UK
Email: henrik.rasmussen@orange.net
Abstract
We consider the pricing of a range of volatility derivatives, including volatility
and variance swaps and swaptions. Under risk-neutral valuation we provide
closed-form formulae for volatility-average and variance swaps for a variety
of diusion and jump-diusion models for volatility. We describe a general
partial dierential equation framework for derivatives that have an extra de-
pendence on an average of the volatility. We give approximate solutions of
this equation for volatility products written on assets for which the volatility
process uctuates on a time-scale that is fast compared with the lifetime of
the contracts, analysing both the outer region and, by matched asymptotic
expansions, the inner boundary layer near expiry.
1 Introduction
In this paper we discuss derivative products that provide exposure to the realised
volatilities or variances of asset returns, while avoiding direct exposure to the under-
lying assets themselves. These products are attractive to investors who either wish
to hedge volatility risk or who wish to take a view on future realised volatilities.
Indeed, much of the investor interest in volatility products seems to have been pro-
vided by the LTCM collapse in 1998, which was accompanied by a dramatic increase
in volatilities. As a result, a number of recent papers [4, 8, 9, 12, 15] address the
evaluation of volatility products; see also Chapter 13 of [17].
Like several of these authors, we take a stochastic volatility model as our starting
point; we also provide formulae for the case that the volatility follows a jump-
diusion process of the type described in [18]. The fact that stochastic volatility
models are able to t skews and smiles, while simultaneously providing sensible
Greeks, have made these models a popular choice in the pricing of exotic options.
Under this framework, we present a number of formulae for the fair delivery
1
price for volatility and variance swaps, and show how other related contracts can be
priced. In addition to providing formulae for a range of volatility and variance swaps,
we consider an asymptotic analysis under which we derive approximate solutions
for volatility and variance swaptions (options on the realised average volatility or
variance). The main motivation for this approximation is the empirical evidence
that volatility is mean-reverting over a time-scale which is fast compared with the
typical lifetime of options and other contracts. That is, when considering the time-
scale of months, stock and index volatility is observed to uctuate rapidly, see for
example the discussion in [10] or [19] and references therein. In this way, we show
how to calculate accurate approximations to the price both at O(1) times before
expiry and in a temporal boundary layer near expiry.
In what follows, the asset S is assumed to follow the usual log-normal process
dS
t
S
t
=
t
(t, ) dt +
t
(t, ) dW
t
, (1)
where W
t
is Brownian motion. For the rest of the paper we x notation as follows:
the conditional expectation at time t is denoted by E
t
= E[[T
t
] where T
t
is the
ltration up to time t and E
0
is thus the initial value of the expectation. All expec-
tations are considered with respect to the risk-neutral probability measure. Within
a stochastic volatility framework, the market is typically incomplete, admitting an
innite number of equivalent martingale measures. This is to say, the market price
of volatility risk is not unique, and it is an open question how one chooses in an
optimal way the appropriate measure to price derivatives. The view we take in the
present investigation, which is the standard machinery for many practical purposes,
is that the risk-neutral probability measure is chosen by the market and this has
a number of immediate implications for calibration issues, mainly that parameter
estimation is not possible from stock data but may be possible using derivatives.
Except from a brief discussion in 4.6 we do not further address these issues here.
The rest of the paper is organized as follows: we begin section 2 by briey
discussing the contracts, while the stochastic volatility framework is introduced in
3, and the general pricing equation is given in 3.3. In 4 we present in depth
the asymptotic analysis which concludes with second order approximations for the
volatility derivatives of interest. In 4.5 we concentrate on pure volatility products,
in which case the analysis greatly simplies. Then we give examples in 5. We
conclude in 6.
2 Variance and volatility swaps
The variance swap is a forward contract in which the investor who is long pays a xed
amount K
var
/$1 nominal value at expiry and receives the oating amount v
R
/$1
nominal value, where K
var
is the strike and v
R
= (
2
)
R
, where is the volatility, is
the realized variance. The entering price must be zero, that is, it costs nothing to
enter the contract; we use this condition to nd the fair value K
var
. The measure
of realized variance to be used is dened at the beginning of the contract; a typical
2
formula for it is
1
T
M

i=1
_
S
i
S
i1
S
i1
_
2
,
which in continuous time we approximate by
v
R
=
1
T
_
T
0

2
t
dt.
The corresponding payo is then
v
R
K
var
. (2)
It is also possible to construct contracts on the realised volatility. One measure of
this is
_
1
T
M

i=1
_
S
i
S
i1
S
i1
_
2
_1
2
derived from the standard deviation of the asset price random walk, and the corre-
sponding continuous-time payo for a volatility swap is
(v
R
)
1/2
K
s/d
=
_
1
T
_
T
0

2
t
dt
_
1
2
K
s/d
; (3)
we term this contract a standard deviation swap. However, this is not the only
possible measure of realised volatility. As discussed in [1], a more robust measure is
_

2MT
M

i=1

S
i
S
i1
S
i1

, (4)
and the associated continuous-time contract has payo

R
K
vol-ave
=
1
T
_
T
0

t
dt K
vol-ave
, (5)
which involves the average of realised volatility, rather than the square root of the
average realised variance as in (3). We term this contract a volatility-average swap.
In addition, we shall consider products based on an average of a suitable implied
volatility, for example the implied volatility
i
t
of the at-the-money call options
with the same expiry as the volatility derivative; this implied volatility swap has
continuous-time payo

i
R
K
i-vol
=
1
T
_
T
0

i
t
dt K
i-vol
. (6)
We could also use a single option throughout the life of the contract, for example
the option that is initially at-the-money; we could further construct implied variance
swaps, and so on.
3
Generalizing further, we consider variance and volatility swaptions [17], a typical
payo being
max(
R
K, 0), max(K
R
, 0)
for volatility call and put swaptions, or
max(v
R
K, 0), max(K v
R
, 0)
for variance swaptions. We can also contemplate contracts whose payo depends on
both a realised volatility or variance and the asset; for example, the payo
max(S
T
e
(
0

R
)

T
K, 0)
is a call option which pays more if the asset rises steadily without much volatility
than if it rises in a volatile way; here
0
is a reference volatility and T is the contracts
lifetime.
3 Risk-neutral pricing techniques
Working within the standard risk-neutral pricing framework, we now outline three
approaches to the valuation of volatility derivatives. In the case of volatility or
variance swaps, this shows that we need to choose
K
var
= E[v
R
] = E
0
_
1
T
_
T
0

2
t
dt
_
,
and for a standard-deviation swap we need K
s/d
= E[v
1/2
R
] (which may be less easy
to calculate explicitly; an approximation in terms of higher moments of v
t
is given
in [4, 15]). For the volatility-average swap, we have
K
volave
= E[
R
] = E
0
_
1
T
_
T
0

t
dt
_
and similarly, for the more complicated implied volatility average swap. Volatility
swaptions and so forth are priced in the usual manner; for example, the price of a
volatility swaption is
E
0
_
max
_
1
T
_
T
0

t
dt K, 0
__
.
There are, however, various approaches to the calculation of these prices, which we
now consider.
3.1 Pricing independently of the volatility model
As observed by [8], it is possible to derive risk-neutral prices for average-variance
products without making any assumption on the evolution of the volatility, although
the asset is still considered to follow the risk-neutral geometric Brownian motion
dS
t
S
t
= r dt +
t
dW
t
.
4
Since
d(log S
t
) = (r
1
2

2
t
) dt +
t
dW
t
,
we have that
log S
T
log S
0
=
_
T
0
(r
1
2

2
t
) dt +
_
T
0

t
dW
t
.
Taking expectations to get risk-neutral prices, we have
1
2
E
0
__
T
0

2
t
dt
_
= rT + log S
0
E
0
[log S
T
] ,
and the nal term on the right-hand side is the value of a log-contract, which can
be decomposed into a continuously parametrised strip of call and put options in a
standard way. Hence variance swaps can be easily priced in terms of vanilla options.
However, this method does not allow us to compute the Greeks, nor does it give
straightforward explicit formulae.
3.2 Pricing by expectations in a stochastic volatility frame-
work
It is possible to derive values for the quantities
E[v
R
] = E
0
_
1
T
_
T
0

2
t
dt
_
=
1
T
_
T
0
E
0
[
2
t
] dt
and
E[
R
] = E
0
_
1
T
_
T
0

t
dt
_
=
1
T
_
T
0
E
0
[
t
] dt
when either the volatility
t
or the variance v
t
follows a quite general random walk.
In this way, we can immediately give risk-neutral prices for variance and volatility-
average swaps, although standard-deviation swaps are less straightforward in this
framework.
In Appendix 1 we show how to calculate E[v
R
] and E[
R
] when
t
follows the
(risk-adjusted) process
d
t
= (a
1
+ a
2

t
)dt + (a
3
+ a
4

t
)dW
t
+ (a
5
+ a
6

t
)dN
t
, (7)
where N
t
is a standard compound Poisson process with constant intensity and
zero correlation with W
t
, and a
1
, , a
6
are constants.
1
We also calculate E
0
[v
R
]
(but not E
0
[
R
]) when v
t
, instead of
t
, follows a similar process; lastly we calculate
E
0
[v
R
] and E
0
[
R
] for the process
d
t
= (b
1
+ b
2

t
)dt + b
3

1/2
t
dW
t
(8)
1
We expect a
1
> 0, a
2
< 0, to model mean-reversion, and we note that for certain choices of
the parameters this model allows negative values of
t
. Models of volatility with jumps have been
considered by [18].
5
and, as above, E
0
[v
R
] if v
t
follows a similar process. Specic results for the mean-
reverting log-normal process
d
t
= (
t
)dt +
t
dW
t
, (9)
for constants , and , are reported in [13]. These, and the more general formulae
of the Appendix, serve to check the asymptotic approach developed in the following
sections.
This approach can also be used to calculate prices and hedge ratios at interme-
diate times. For example, if we have a volatility-average swap, with payo
1
T
_
T
0

2
s
ds K
volave
,
the value at earlier times t (0 t < T) is
e
r(Tt)
_
E
t
_
1
T
_
T
0

s
ds
_
K
volave
_
=
e
r(Tt)
T
__
t
0

s
ds +E
t
__
T
t

s
ds
_
TK
volave
_
since the contribution
_
t
0

s
ds to the nal average of the volatility is known at time
t. Using the formulae of Appendix 1, the conditional expectation
E
t
__
T
t

s
ds
_
= F(
t
, t),
say, is readily evaluated, and then the Vega of the contract is
1
T
e
r(Tt)
F

(10)
which is also readily evaluated once
t
is known.
3.2.1 Examples
Here we briey give some examples for the price of volatility products, the derivation
of which is based on the results of the Appendix. The mean-reverting log-normal
model
d
t
= (
t
)dt +
t
dW
t
,
is a special case of (7), and following the calculation of A1, under this model, we
have
K
volave
=
1 e
T
T
(
0
) + . (11)
Similarly, the value at any time t T is
V
t
=
e
r(Tt)
T
__
t
0
(
s
)ds +
1

(e
(Tt)
1)(
0

t
)
_
. (12)
6
In addition, for the fair strike of the variance swap we have
K
var
=
2
2
(2
2
)T
_
T
1 e
(2
2
)T
2
2
_
+
2(
0
)
(
2
)T
_
1 e
T

1 e
(2
2
)T
2
2
_
+

2
0
(2
2
)T
_
1 e
(2
2
)T
_
. (13)
The formula for the price of the contract at time t is
V
t
= e
r(Tt)
_
1
T
_
t
0

2
s
ds +
2
2
(2
2
)T
_
(T t)
1 e
(2
2
)(Tt)
2
2
_
+
2(
t
)
(
2
)T
_
1 e
(Tt)

1 e
(2
2
)(Tt)
2
2
_
+

2
t
2
2
_
1 e
(2
2
)(Tt)
_
K
var
_
, (14)
with K
var
given in (13).
Note that if we were to consider the variance v
t
as the underlying process sat-
isfying the mean reverting log-normal model, then the relevant expressions for the
variance swap would be dierent. We can illustrate the contrast between volatility
and variance driven models by considering the two cases
d
t
= k
1
(
t

)dt + k
2

dW
t
,
and
dv
t
= k
3
(v
t
v

) dt + k
4

v
t
dW
t
.
Then, following the calculations in A2, we have for the volatility-average swap
K
volave
=
1
T
_

0
k
1
_
e
k
1
T
1
_
+

_
k
1
T e
k
1
T
+ 1
_
_
,
in the rst case, and we cannot price this contract explicitly in the second framework.
However, the price of the fair variance strike is
K
var
=
1
T
_

1
T
2b
2

2
b
2
2
_
e
b
2
T
1
_
+
1
2b
2
_

2
0
+

2
b
2
+

1
2b
2
_
_
e
2b
2
T
1
_
_
,
for the rst model and
K
var
=
1
T
_

2
0
k
3
_
e
k
3
T
1
_
+ v
t
_
k
3
T e
k
3
T
+ 1
_
_
for the second; the various constants are as dened in the Appendix. It is apparent
that the second model, for v
t
rather than
t
, leads to considerably simpler formulae.
7
3.3 Pricing via partial dierential equations
Explicit formulae are in general only available for pure volatility products, and un-
der the assumption that the coecients in the process for volatility are independent
of S
t
. For more general cases, we must use either Monte-Carlo methods or nu-
merical/asymptotic solutions of the pricing dierential equation. In this section we
consider the latter.
From now on we assume that it is
t
that drives the volatility; if instead the
underlying process is written in terms of v
t
=
2
t
the appropriate modications are
easily made. We assume that S
t
and
t
follow the process
dS
t
S
t
=
t
dt +
t
dW
t
,
d
t
= M
t
dt +
t
d

W
t
(15)
where M
t
and
t
may depend on S
t
and
t
, and where the correlation coecient of
W
t
and

W
t
is
t
. As with Asian options, we need to introduce a variable to measure
the average volatility to date. The payo of the derivatives under consideration
involves an average of the form
I
T
=
_
T
0
F(
s
)ds;
for example,
I
var
T
=
_
T
0

2
s
ds
for a variance swap, for which the payo (2) is
1
T
I
var
T
K
var
,
and the same average is sucient for the standard-deviation swap payo (3), namely
_
1
T
I
var
T
_
1/2
K
s/d
.
Likewise the volatility-average swap and the implied-volatility swap have payos (5)
and (6), namely
1
T
I
volave
T
K
volave
,
1
T
I
ivol
T
K
ivol
respectively, where I
volave
T
=
_
T
0

s
ds, I
ivol
t
=
_
T
0

i
s
ds, and so forth. For times
before expiry, we use the running average (denoted generally by I
t
regardless of F)
I
t
=
_
t
0
F(
s
) ds,
8
and then it is a standard combination of stochastic volatility and Asian options
analysis [10, 21] to show that the value V (S, , I, t) of a derivative satises the
partial dierential equation
V
t
+
1
2

2
S
2

2
V
S
2
+ S

2
V
S
+
1
2

2
V

2
+ F()
V
I
+rS
V
S
+ (M )
V

rV = 0, (16)
where is the market price of volatility risk. We write the general payo condition
in the form
V (S, , I, T) = P(S, I). (17)
It is straightforward to show that the formulae derived in 3 and the Appendix
satisfy this equation when = 0. For example for the mean-reverting log-normal
model (9) we have
M = (
t
), =
t
.
and expressions (12), (14) satisfy equation (16).
4 Asymptotic analysis for fast mean-reversion
We now present an asymptotic analysis for the case of fast mean-reversion for
t
,
in the spirit of the Fouque et al. (FPS) analysis [10] for equity and xed-income
derivatives. The novelty here is rst in the application to volatility products, and
secondly in that we provide a fairly complete description of the solution both at O(1)
times before expiry which FPS do, and in the short boundary layer immediately
before expiry, which they do not. We initially make the simplifying assumptions
that M
t
and
t
, the coecients in the process for
t
, are independent of S
t
, and
we let
t
= 0,
t
= 0. The assumption of zero correlation is not as dramatic as it
would be for equity derivatives, and indeed if we consider payos depending only on
volatility averages it is irrelevant. We indicate briey at the end of this section how
the analysis should be extended for nonzero correlation.
We introduce a small parameter , where 0 < 1, to measure the ratio of the
mean-reversion time-scale to the lifetime of an option, and we assume that M
t
and

t
have the forms
M
t
=
m
t

,
t
=

t

1/2
,
the relative sizes of these coecients being chosen to ensure that
t
has a nontrivial
invariant distribution
p

() = lim
t
p(, t[
0
, 0)
where p(, t[
0
, 0) is the transition density function for
t
starting from
0
at time
zero.
With these assumptions the pricing equation becomes
V
t
+
1
2

2
S
2

2
V
S
2
+
1
2

2

2
V

2
+ rS
V
S
+
m

+ F()
V
I
rV = 0,
9
which we write in the form
_
1

L
0
+L
1
_
V = 0
where
L
0
=
1
2

2

2

2
+ m

,
L
1
=

t
+
1
2

2
S
2

2
S
2
+ rS

S
+ F()

I
r.
We note immediately that p

() satises
L

0
p

2
_
1
2

2
p

(mp

) = 0
where L

0
is the adjoint of L
0
, and we assume that
2
, m are such that p

exists; it
is then proportional to e
2
R

m(s)/
2
(s) ds
/
2
(). We introduce the notation , ) for
the usual inner product over 0 < < , and note the identity
L
0
u, v) = u, L

0
v) (18)
for suitable functions u and v, which we will use repeatedly.
We now expand
V (S, , I, t) V
0
(S, , I, t) + V
1
(S, , I, t) +
2
V
2
(S, , I, t) + ,
so that
1

L
0
V
0
+ (L
1
V
0
+L
0
V
1
) + (L
1
V
1
+L
0
V
2
) + = 0.
At lowest order, O(1/), we have
L
0
V
0
= 0
and so
V
0
= V
0
(S, t, I)
since the operator L
0
consists of derivatives with respect to only (the particular
solutions that depend on are in general ruled out by the conditions at large and/or
small S); however V
0
is as yet undetermined.
At O(1) we nd
L
0
V
1
+L
1
V
0
= 0, (19)
a Poisson equation for V
1
considering V
0
as known. Multiplying by p

, integrating
and using (18), we nd that the solvability (Fredholm Alternative) condition for this
equation can be expressed as
L
1
V
0
, p

) = 0.
10
That is, L
1
V
0
is orthogonal to p

, which, being a solution of the stationary forward

Kolmogorov equation for , is an eigenfunction of L

0
. Thus,
V
0
t
+
1
2
S
2

2
V
0
S
2
_

0
p

()
2
d + rS
V
0
S
+
V
0
I
_

0
p

()F()d rV
0
= 0,
where we have used the result that V
0
is independent of . Denoting the integrals
by
2
and F = F() respectively to represent the fact that they are averages of
2
and F(), we write this as
L
1
V
0
=
V
0
t
+
1
2

2
S
2

2
V
0
S
2
+ rS
V
0
S
+ F
V
0
I
rV
0
= 0, (20)
(note that L
1
= L
1
, p

)). Making the transformation

I = I + (T t)F
and writing V
0
(S, t, I) = V
0
(S, t; I), reduces this further, to
V
0
t
+
1
2

2
S
2

2
V
0
S
2
+ rS
V
0
S
rV
0
= 0, (21)
which is the BlackScholes equation with volatility
_

2
_
1/2
. The dependence on I
is retained parametrically via the payo, which takes the form
V
0
(S, T; I) = V
0
(S, T, I) = P(S, I),
and so we have V
0
(S, t, I) = V
0
(S, t, I + (T t)F). Of course, this formula is
considerably simpler for pure volatility products, for which P/S = 0 so that
V/S 0.
The next step is to calculate V
1
. First, observe that L
1
V
0
can be written as
L
1
V
0
=
1
2
(
2

2
)S
2

2
V
0
S
2
+ (F() F)
V
0
I
.
Hence, (19) can be written as
L
0
V
1
=
1
2
(
2

2
)S
2

2
V
0
S
2
+ (F F())
V
0
I
.
We seek a solution of the form
V
1
(S, t, , I) = f
2
()S
2

2
V
0
S
2
+ f
1
()
V
0
I
+ H(S, t, I), (22)
where H is independent of . The functions f
1
and f
2
are then the appropriate
solutions of the equations
1
2

2
()
d
2
f
2
d
2
+ m()
df
2
d
=
1
2
(
2

2
),
1
2

2
()
d
2
f
1
d
2
+ m()
df
1
d
= F F(). (23)
11
and can readily be found in integral form; one of the complementary solutions
is a constant and can be absorbed into H, and the other is unbounded at innity
(because p

exists). However, the function H(S, t, I) can only be determined by

proceeding to next order and applying the solvability condition to the equation
L
0
V
2
+L
1
V
1
= 0. (24)
We further see that the solution (22), which depends on , cannot satisfy the pay-
o condition V
1
(S, , I, T) = 0. This discrepancy is resolved by a boundary layer
analysis in which T t = O() (if the payo has discontinuities, as for a volatility
option, further local analysis near these points is also necessary). We point out that
the lowest-order analysis is quite general, and not specic to the random walk (15).
Only at higher order do we need to know more about these details, and even then
only certain moments need be calculated.
Now, the solvability condition for (24) takes the form:
L
1
V
1
, p

) = 0
or
L
1
_
f
2
()S
2

2
V
0
S
2
_
, p

) +L
1
_
f
1
()
V
0
I
_
, p

) +L
1
H, p

) = 0,
that is,
L
1
_
f
2
()S
2

2
V
0
S
2
_
, p

) +L
1
_
f
1
()
V
0
I
_
, p

) +L
1
H = 0, (25)
where L
1
is as in (20). Before solving (25), we note that since L
1
V
0
= 0, we also
have
L
1
_
S
n

n
V
0
S
n
_
= 0, n 1, L
1
_
V
0
I
_
= 0. (26)
We can therefore write the rst term in (25) as
L
1
_
f
2
()S
2

2
V
0
S
2
_
, p

) =
1
2
S
2

2
S
2
_
S
2

2
V
0
S
2
_
_

2
f
2
()
2
f
2
()
_
+

I
_
S
2

2
V
0
S
2
_
_
F()f
2
() F() f
2
()
_
. (27)
The second term, after some algebra, takes the form
L
1
_
f
1
()
V
0
I
_
, p

) =
1
2
S
2

3
V
0
IS
2
_

2
f
1
()
2
f
1
()
_
+

2
V
0
I
2
_
F()f
1
() F() f
1
()
_
.
As a result we can rewrite equation (25) as
L
1
H = A
1
S
2

3
V
0
IS
2
+ A
2

2
V
0
I
2
+ A
3
S
2

2
S
2
_
S
2

2
V
0
S
2
_
= W(S, t, I), (28)
12
say, where we have dened
A
1
=
_
F() f
2
() F()f
2
()
_
+
1
2
_

2
f
1
()
2
f
1
()
_
,
A
2
= F() f
1
() F()f
1
(),
A
3
=
1
2
_

2
f
2
()
2
f
2
()
_
. (29)
However, from (26), L
1
W = 0, which implies that L
1
((T t)W) = W. This
allows us to nd a particular solution of the form (T t)W, and we can then
express the general solution of (28) as
H(S, t, I) = (T t)
_
A
1
S
2

3
V
0
IS
2
+ A
2

2
V
0
I
2
+A
3
S
2

2
S
2
_
S
2

2
V
0
S
2
__
+ H
1
(S, t, I), (30)
where H
1
solves the equation
L
1
H
1
= 0. (31)
The solution of this last equation can be obtained directly via the transformation
I = I + F(T t), which transforms (31) into the Black-Scholes equation with
volatility
_

2
_
1/2
.
4.1 Boundary Layer Analysis
We have already satised the payo condition with V
0
, and so we expect that
V
1
(S, T, , I) = 0.
But this is not possible, since V
1
is a function of whereas the payo is not. The
remedy is to introduce a boundary layer in t near t = T, of thickness of O(). We
dene the scaled inner variable via
t = T + , < 0.
Our goal is now to nd the expansion in the boundary layer (the inner solution),
which we subsequently match with the solution outside the boundary layer (the
outer solution) using Van Dykes matching principle [20]. We introduce the following
operators:

L
0
=

+L
0
,

L
1
=
1
2

2
S
2

2
S
2
+ rS

S
+ F()

I
r,
and we have
L =
1

L
0
+

L
1
,
13
but we note that

L
0
, unlike L
0
above, contains the time derivative /. We write

V (S, , , I) for the value of the derivative in the boundary layer, and expand

V

V
0
+

V
1
+ .
As a result equation (16) becomes
1

L
0

V
0
+
_

L
1

V
0
+

L
0

V
1
_
+ = 0. (32)
At lowest order,

L
0

V
0
= 0
with the condition

V
0
(S, 0, , I) = P(S, I).
The solution is easily seen to be

V
0
(S, , , I) = P(S, I). (33)
Note that this matches automatically with our outer solution V
0
(S, t, I) as t T,
.
At the next order we have

L
0

V
1
=

L
1

V
0
=

L
1
P,
which can be written as

V
1

+L
0

V
1
=
1
2
_

2
_
S
2

2
P
S
2
+
_
F() F()
_
P
I
L
1
P (34)
with the nal condition

V
1
(S, 0, , I) = 0
(note that P/ = 0). We now need the solution of (34) as in order to
match with the outer solution. A particular solution of (34) is

V
1

= f
2
()S
2

2
P
S
2
+ f
1
()
P
I
L
1
P +

H(S, I) (35)
where

H(S, I) is arbitrary, and we conjecture that this is the correct form for the
asymptotic behaviour of

V
1
(S, , , I) as . Now

L
0

V
1
, p

) =

V
1

+L
0

V
1
, p

) =

V
1

, p

)
since L
0

V
1
, p

) = 0. We further note that from (34),

V
1
, p

) =

V
1

, p

) = L
1
P, p

) = L
1
P
and therefore

V
1
, p

) = L
1
P, (36)
14
where we have used the condition

V
1
(S, 0, , I) = 0. Now from (35) we have

1
, p

) = L
1
P + f
2
()S
2

2
P
S
2
+ f
1
()
P
I
+

H(S, I). (37)
We now compare (36) with (37) and we immediately have

H(S, I) = S
2

2
P
S
2
f
2
()
P
I
f
1
().
Thus, as ,
V
1


V

1
= S
2
_
f
2
() f
2
()
_

2
P
S
2
+
_
f
1
() f
1
()
_
P
I
L
1
P, (38)
since subtracting of the particular solution leaves a complementary function

V
1

1
whose inner product with p

vanishes, which satises the homogeneous version

of the parabolic equation (34), and which therefore vanishes as .
4.2 Matching
We now consider the matching. Expanding the outer solution to O() in the inner
variable , we have
V
0
(S, t, I) =

V
0
_
S, T + , I + F()(T t)
_
=

V
0
_
S, T + , I F
_
V
0
(S, T, I) +
V
0
t
_
S, T, I
_
F
V
0
I
(S, t, I)
= P(S, T, I) +
V
0
t
(S, T, I) F
V
0
I
(S, t, I)
(note that for t = T we have I = I). Using (20) this reduces to
V
0
(S, t, I) P(S, I) L
1
P,
and so, combining with (22), the two-term inner expansion of V
0
+ V
1
is
P(S, I) +
_
LP + f
1
()
P
I
+ f
2
()S
2

2
P
S
2
+ H(S, T, I)
_
. (39)
Similarly, from (33) and (38), the two-term expansion of the inner solution in terms
of the outer variable t is
P(S, I) +
_
_
f
2
() f
2
()
_
S
2

2
P
S
2
+
_
f
1
() f
1
()
_
P
I
L
1
P
_
. (40)
Matching (40) and (39), we see that the nal condition for H(S, t, I) is
H(S, T, I) = f
2
()S
2

2
P
S
2
f
1
()
P
I
,
15
and therefore (30) becomes
H
1
(S, T, I) = f
2
()S
2

2
P
S
2
f
1
()
P
I
,
and the boundary layer analysis thus leads to the missing nal condition for H
1
.
As stated above, we can reduce LH
1
= 0 to the Black-Scholes equation using the
substitution I = I +(T t)F, and hence we nd H
1
(S, t; I) = H
1
(S, t, I) by solving
L
BS
H
1
= 0,
with
H
1
(S, T; I) = H
1
(S, T, I) = f
2
()S
2

2
P
S
2
f
1
()
P
I
.
Recalling that V
0
(S, t; I) satises L
BS
V
0
= 0 with V
0
(S, T; I) = P(S, I), and (26),
we see that for t T,
H
1
(S, t; I) = f
2
()S
2

2
V
0
S
2
f
1
()
V
0
I
.
4.3 Summary
In summary, the outer expansion takes the form
V (S, t, , I) V
0
(S, t, I) + V
1
(S, t, , I)
where
V
0
(S, t, I) = V
0
(S, t; I + (T t)F)
and V
0
(S, t; I) satises the Black-Scholes problem
L
BS
V
0
= 0, V
0
(S, T; I) = P(S, I)
with volatility (
2
)
1/2
. The correction V
1
(S, t, , I) takes the form
V
1
(S, t, , I) = S
2

2
V
0
S
2
_
f
2
() f
2
()
_
+
V
0
I
_
f
1
() f
1
()
_
(T t)
_
A
1
S
2

3
V
0
IS
2
+ A
2

2
V
0
I
2
+ A
3
S
2

2
S
2
_
S
2

2
V
0
S
2
__
,(41)
where the functions f
1
(), f
2
() are dened in (23) and the constants A
1
, A
2
, A
3
in
(29). In the boundary layer, the solution takes the form

V (S, , , I) P(S, I) +

V
1
,
where

V
1
satises the parabolic problem (34) with nal data

V
1
(S, 0, , I) = 0; this
solution cannot be found explicitly but its large-time behaviour is sucient to allow
matching with the outer solution to this order.
16
4.4 Extensions
4.4.1 Payo singularities
It is implicit in our analysis that the payo is smooth enough for the expansion to be
valid. For example, in (30) we have terms such as
2
V
0
/I
2
which is a delta function
at expiry if the contract is a volatility call swaption with payo max(I/T K, 0).
Roughly speaking, such a discontinuity will propagate along the line I +(T t)F =
KT, where K is the strike, and a separate analysis is necessary in a small region
around this line; the result is an interesting mathematical problem which we briey
discuss in 5.5. We also note that we could (but do not) also consider the implications
of discontinuities in the S-dependence of the payo, which (because of the second
S-derivative in the pricing equation) may be expected to induce small square root
of time to expiry region in which the expansion is not valid.
4.4.2 Correlation and the market price of risk
We also briey consider the eects of correlation and the market price of volatility
risk. If
t
or
t
is of O(
1/2
) then their eect is to modify the operator L
1
(we do
not consider this special case here), but if they are of O(1) then our pricing equation
takes the form
_
1

L
0
+
1

1/2
L1
2
+L
1
_
V = 0
where L
0
and L
1
are as before, and
L1
2
= S

2
S

.
The expansion for V is now in the form
V V
0
+
1/2
V1
2
+ V
1
+
3/2
V3
2
+ ,
and the rst four equations are, in order,
L
0
V
0
= 0,
L
0
V1
2
+L1
2
V
0
= 0,
L
0
V
1
+L1
2
V1
2
+L
1
V
0
= 0, (42)
L
0
V3
2
+L1
2
V
1
+L
1
V1
2
= 0. (43)
We still have that V
0
is a function of (S, I, t) alone, hence L1
2
V
0
= 0 and so V1
2
is
also a function of (S, I, t). Thus L1
2
V1
2
= 0 and the solvability condition for (42)
gives
L
1
V
0
, p

) = 0
17
just as before, so also as before V
0
= V
0
(S, t; I + (T t)F). We then have
V
1
= f
2
()S
2

2
V
0
S
2
+ f
1
()
V
0
I
+ H(S, t, I),
also as before, and we note that L1
2
V
1
,= 0 but L1
2
H = 0. Proceeding to the
solvability condition for (43), we have L1
2
V
1
+L
1
V1
2
, p

) = 0, which becomes
LV1
2
= f

2
S

S
_
S
2

2
V
0
S
2
_
f

1
S

2
V
0
SI
+ f

2
S
2

2
V
0
S
2
+ f

1
V
0
I
, (44)
where we denote by f

1
and f

2
the corresponding derivatives with respect to . The
solution for V1
2
that vanishes at expiry is readily found to be
V1
2
(S, I, t) = (T t)
_
f

2
_
S
3

3
V
0
S
3
+ 2S
2

2
V
0
S
2
_
+ f

1
S

2
V
0
SI
f

2
S
2

2
V
0
S
2
f

1
V
0
I
_
(45)
where V
0
is already known. The calculation of V
1
, from a solvability condition at
O(
2
), is straightforward but even more cumbersome and we do not give details here;
in the next section we give them for a pure volatility product. In the boundary layer,
we have
_
1

L
0
+
1

1/2
L1
2
+

L
1
_

V = 0,

V

V
0
+
1/2

V1
2
+

V
1
+
where

L
0
and

L
1
are as before. Again

V
0
= P(S, I), matching automatically with
V
0
, and the problem for

V1
2
is

L
0

V1
2
= L1
2

V
0
= 0;
the solution is simply

V1
2
= 0, and this is consistent with matching with the two-
term outer expansion V
0
+
1/2
V1
2
, since in inner variables the T t in V1
2
means
that this term only contributes O(
3/2
) to the inner expansion of the outer solution.
Again,

V
1
can be calculated and matched although we do not give details here.
In summary, with nonzero O(1) correlation, the outer solution is given by
V (S, t, , I) V
0
(S, t, I) +
1/2
V1
2
(S, t, I)
where V
0
is the solution of a Black-Scholes problem and V1
2
is given in (45).
4.5 Pure volatility products
The analysis above is greatly simplied for pure volatility products for which all S-
derivatives vanish. Again using
t
as the underlying volatility variable, the pricing
equation is simply
V
t
+
1
2

2
V

2
+
_
m

1/2
_
V

+ F()
V
I
rV = 0
18
with payo
V (, I, T) = P(I).
For certain models and products it is possible to write down explicit solutions not
only for swapsas in 3.2but also for swaptions, since corresponding results for
Asian options can simply be transferred. In particular, for a Hull-White type model
in which =
0
and M = M
0
, so that
t
follows Geometric Brownian
Motion, there is a series solution for a volatility-average swaption and, if one were
to construct a swaption based on the continuously sampled geometric mean of
t
,
that too would have an explicit solution because the running geometric average of
a Geometric Brownian Motion is log-normally distributed.
4.5.1 Asymptotics for pure volatility products
As mentioned above, the approximate analysis is rather simpler for a pure volatility
product, since then all S-derivatives vanish. We therefore give more details of the
analysis of 4.4, in Appendix 2; in summary, the outer expansion takes the form
V V
0
+
1/2
V
1/2
+ V
1
+
where
V
0
(I, t) = e
r(Tt)
P(I + (T t)F), (46)
V1
2
(I, t) = (T t) f

1
V
0
I
= (T t) e
r(Tt)
f

1
P

(I + (T t)F), (47)
and
V
1
(I, t) =
_
f
1
() f
1
() (T t) f

3
_
V
0
I
+
_
_
f

1
_
2 (T t)
2
2
A
2
(T t)
_

2
V
0
I
2
= e
r(Tt)
_
f
1
() f
1
() (T t) f

3
_
P

(I + (T t)F)
+e
r(Tt)
_
_
f

1
_
2 (T t)
2
2
A
2
(T t)
_
P

(I + (T t)F). (48)
The function f
3
() is dened in Appendix 2. The boundary-layer solution is as
before.
Notice that the leading order term in the outer solution has a natural nancial
interpretation. If, say, the contract is a call swaption, the outer solution is zero for
I + (T t) F < KT;
since I is the contribution to the payo already in the bank and (T t)F is the
approximate expected remaining contribution (as in a fast mean-reverting model the
uctuations in average out at this order), the option is unlikely to pay out if the
sum of these is an O(1) amount less than KT. The discontinuity at I +(T t)F =
KT must be resolved by an inner expansion which we do not deal with here.
19
4.6 Calibration
We note briey that, like the original FPS [10] model, a great deal of calibration of
the current framework can be accomplished without reference to a specic model.
The rst two terms in (45) can be calibrated to pure equitie option volatility smiles as
in the FPS scheme. It is more likely that calibration is neseccary for pure volatility
products, and gives, say, a series of volatility swaptions of dierent prices, (46)
and (47) can be used to calculate F and f

1
. The calibration to O() is more
complicated and we do not discuss it here, although it is in principle possible.
5 Examples
In this section we illustrate the theory with several dierent products. The rst three
are the pure volatility swaps described in 2 where explicit formulae are available (see
3.2.1); the asymptotic results can be shown to be correct. For the implied-volatility
swap, there is some S-dependence via
i
, although the payo is still independent of
I. Finally we look briey at volatility-average swaptions.
5.1 The variance swap
For this contract, we have F() =
2
and so F =
2
, the average variance to be used
in the Black-Scholes equation (21). The payo is I
var
/T K
var
= I
var
/T K
var
and the rst three terms in the expression for the variance swap value are
V (t, , I) e
r(Tt)
_
I
var
t
+
2
(T t)
T
K
var
_

1/2
e
r(Tt)
(T t) f

1
T
+e
r(Tt)
_
f
1
() f
1
() (T t)f

3
_
T
.
For the random walk (9) for which
2
= 2
2
/(2
2
) = 2a
2
/(2a b
2
), it is
easily conrmed that we recover the O(1) term of the exact result (14).
5.2 The standard-deviation swap
For the standard-deviation swap payo (3), we still have F() =
2
but now the
payo is (I
var
/T)
1/2
K
s/d
. Hence, the standard-deviation swap price to order O()
is
V (t, , I) e
r(Tt)
__
I
var
t
+
2
(T t)
T
_
1/2
K
s/d
_

1/2
e
r(Tt)
2T
(T t) f

1
_
I
var
t
+
2
(T t)
T
_
1/2
+
e
r(Tt)
4T
2
_
2T
_
f
1
() f
1
() (T t) f

3
_
_
I
var
t
+
2
(T t)
T
_
1/2
+(T t)
_
A
2

_
f

1
_
2 (T t)
2
_
_
I
var
t
+
2
(T t)
T
_
3/2 _
.
20
5.3 The volatility-average swap
For the payo (5) we have F() = and we nd that
V (t, , I) e
r(Tt)
_
I
volave
t
+ (T t)
T
K
vol-ave
_

1/2
e
r(Tt)
T
(T t) f

1
+e
r(Tt)
_
f
1
() f
1
() + (T t) f

3
_
T
Again here, we recover the O(1) term of the exact solution given by (11) and (12)
for large mean reversion coecient.
5.4 The implied volatility swap
In this case F() is the implied volatility of an at the money option, say a call, with
price C(S, t, ). We can of course apply the same procedure to pure equity options
(this is the FPS analysis) to give
C(S, t, ) C
BS
_
S, t,
_

2
_
1/2
_
+
1/2
(T t)
_
f

2
_
S
3

3
C
BS
S
3
+ 2S
2

2
C
BS
S
2
_
f

2
S
2

2
C
BS
S
2
_
+O(). (49)
The implied volatility of this option can also be expanded in the form

i

i
0
+
1/2

i
1
+
and since by denition
C(S, t, ) = C
BS
(S, t,
i
),
we have
C(S, t, ) C
BS
(S, t,
i
0
) +
1/2

i
1
C
BS

i
0
+O(). (50)
Comparing (49) with (50) we clearly have

i
0
=
_

2
_
1/2
and, setting the strike of C equal to S and substituting from the Black-Scholes
formulae,

i
1
=
f

2
_
1
2

2
r
_
f

2

2
_

2
_
3/2
.
21
Hence we see an additional complication in that the averaging function F() itself
has an expansion
F()
_

2
_
1/2
+
1/2

i
1
+O().
Thus the price operator takes the form
1

_
L
0
+
1/2
L1
2
+ L
1
+
3/2
L3
2
+
_
where
L3
2
=
i
1

I
.
This means in turn that the right-hand side of (44) has an extra term
L3
2
V
0
, p

) =
i
1
V
0
I
and so there is an extra term (T t)
i
1
V
0
/I on the right hand side of (45), that
is an extra O(
1/2
) correction to V ; it has the obvious nancial interpretation as
the vega with respect to the average I. The O() correction, however, is more
complicated because of the S-dependence in the implied volatility and we do not
give details here.
5.5 Volatility-average swaptions
For the volatility-average swaption we have F() = so F = , and
V (t, , I) e
r(Tt)
max
_
I + (T t)
T
K, 0
_

1/2
e
r(Tt)
T
(T t) f

1
H(I + (T t) TK)
+
e
r(Tt)
T
_
_
f
1
() f
1
() (T t) f

3
_
H(I + (T t) TK)
+
_
_
f

1
_
2 (T t)
2
2
A
2
(T t)
_
(I + (T t) TK)
_
,
where His the Heaviside function and (x) is the delta function. As discussed above,
the O() contribution is singular on the line I + (T t)F and the expansion is not
valid in this region.
The singularity near I = KT (T t) is resolved by the introduction of an
inner layer coordinate dened by
I = KT (T t) + ,
leading to the problem
( )
V

+L
0
V +O() = 0
where the O() term contains the remaining derivatives and undierentiated terms
(if the averaging is with respect to F(), the coecient of V/ is F() F()).
22
This is to be solved for all ; that is, the leading order inner value V
0
satises the
forward-backward parabolic equation
1
2

2
V
0

2
+ m
V
0

= ( )
V
0

for < < , 0 < < , with matching conditions

V () 0, , and V () , +
for a call swaption and these conditions interchanged for a put swaption. This rather
unconventional problem has been considered in [2] in the context of inner layers for
models of neutron transport. It is crucial that the coecient of the time derivative
V
0
/ satises , p

) = 0, and given this condition it can be shown that there

is a unique solution connecting the behaviour as to that as +, even
though the equation itself is not standard. Similar remarks apply to derivatives with
other payo singularities such as digital options.
6 Conclusion
We have described a range of approaches to the pricing and hedging problem for a
variety of products depending on realised volatility. Some of these, especially those
based on realised variance, are already traded; but as we point out in 2, from a
statistical point of view the realised rst variation (4) is a more robust estimator
and hence we have described the application of the theory to volatility-averaged
options as well as to implied volatility averages, which are the type measured by
the VIX index [7]. We have presented an asymptotic analysis which leads to a
description of the derivative price even in the time leading up to expiry, and this
should be an accurate approximation whenever volatility is fast mean-reverting, as
well as straightforward to calibrate from implied volatility smiles.
A Appendix 1: Derivation of certain expectations
A.1 The random walk (7)
Let y
t
satisfy the stochastic dierential equation
dy
t
= (a
1
+ a
2
y
t
)dt + (a
3
+ a
4
y
t
)dW
t
+ (a
5
+ a
6
y
t
)dN
t
, (51)
with zero correlation between W
t
and N
t
. Dene
E
1
= E[y

[y
0
], E
1
=
_

0
E
1t
dt,
and
E
2
= E[y
2

[y
0
], E
2
=
_

0
E
2t
dt.
23
Recalling that E[dN
t
] = dt, we have
dE
1t
= (a
1
+ a
2
E
1t
+ (a
5
+ a
6
E
1t
)) dt
= (
0
+
1
E
1t
)dt,
where

0
= a
1
+
5
,
1
= a
2
+ a
6
,
from which, since E
10
= y
0
,
E
1
= y
0
e

1
(1 e

), (52)
and
E
1
=
y
0

1
(e

1)

0

2
1
(
1
e

+ 1). (53)
For a mean-reverting process
1
< 0, so the unconditional expectation of y
t
is
y

= lim

E
1
=

1
,
and the long-term average of y
t
is the same:
lim

1
E
1
=

1
;
note the relatively slow (algebraic) decay of the contribution to E
1
from the initial
value y
0
.
Now dene z
t
= y
2
t
. We easily nd that
dz
t
=
_
2y
t
(a
1
+ a
2
y
t
) + (a
3
+ a
4
y
t
)
2
_
dt
+2y
t
(a
3
+ a
4
y
t
)dW
t
+ (a
5
+ a
6
y
t
) (a
5
+ (a
6
+ 2) y
t
) dN
t
. (54)
Thus E
2
= E[z

[z
0
= y
2
0
] satises
dE
2t
= (
0
+
1
E
1t
+
2
E
2t
) dt
where

0
= a
2
3
+ a
2
5
,

1
= 2(a
1
+ a
5
) + 2(a
5
a
6
+ a
3
a
4
) = 2(
0
+
1
),

2
= 2(a
2
+ a
6
) + a
2
6
+ a
2
4
= 2(
1
+
2
), (55)
with
0
,
1
as before, and

1
= a
5
a
6
+ a
3
a
4
,
2
=
a
2
6
+ a
2
4
2
.
Using (53), we nd that
E
2
=
_

2
__
1 e

1
(
1
y
0
+
0
)

1
(
1

2
)
_
e

_
+ y
2
0
e

,
24
and we note that E
2
grows exponentially in unless
2
< 0, a condition analogous
to
1
< 0 for E
1
. In this case, the unconditional expectation of E
2
, namely
z

= lim

E
2
,
is (
0

0
)/
1

2
.
Lastly, we calculate the averaged expectation of z
t
,
E
2
=
_

2
_

+
_

2
+

1
(
1
y
0
+
0
)

1
(
1

2
)
y
2
0
__
1 e

2
_

1
(
1
y
0
+
0
)

2
1
(
1

2
)
(1 e

) . (56)
A.1.1 Special cases
We note some special cases:
(i) The case
1
= 0, ,= 0 corresponds to the condition =
2
/
6
, and then we
have
E
1
= y
0
+
0
, E
1
= y
0
+
1
2

2
,
E
2
=

0
+
1
y
0
+
2
y
2
0

2
_
e

+ 1
_
+

2
y
2
0

2
,
and
E
2
=

0
+
1
y
0
+
2
y
2
0

2
2
(e

1)

2
(
0
+
1
y
0
)

1
2
2

2
.
(ii) The case
1
=
2
,= 0 corresponds to
1
+a
2
6
+a
2
4
= 0, and then we have E
1
,
E
1
as in the general case, whereas
E
2
=
_

2
1
_
(1 e

) +

1
(
1
y
0
+
0
)

1
e

+ y
2
0
e

,
and
E
2
=
_

2
1
_
+

1

2
1
y
2
0

3
1
(1 e

) .
(iii) The case
1
=
2
= 0 corresponds to a
4
= a
6
= 0, and we have
E
2
= y
2
0
+ (
0
+ 2
0
y
0
) +
2
0

2
,
E
2
= y
2
0
+
(
0
+ 2
0
y
0
)
2
2
+

2
0

3
3
.
25
(iv) Finally, consider the case
1
,= 0,
2
= 0. Then we have that 2a
2
+ 2a
6
+
a
2
6
+ a
2
4
= 0 and we have E
1
and E
1
as in the general case, and
E
2
=
(
1

1
)

1
(
1
y
0
+
0
)

2
1
(1 e

) + y
2
0
and
E
2
=
(
1

1
)
2
2
1

1
(
1
y
0
+
0
)

3
1
(
1
e

+ 1) + y
2
0
.
A.1.2 Derivatives pricing
Using the general expressions (52), (53), (56) and (56), or any of the particular cases
described above, either volatility-average or variance swaps can be priced by taking
y
t
=
t
(in which case z
t
=
2
t
= v
t
) or by taking y
t
= v
t
directly, depending on
whether the volatility model is for
t
or v
t
. For example, the strike for a volatility-
average swap is
K
volave
=
1
T
E
0
_
_
T
0

t
dt
_
=
1
T
E
1T
[
y
0
=
0
,
where it is understood that the process for y
t
is the required model for volatility. In
a similar way, the expectation needed to calculate the vega (10), namely,
E
t
_
_
T
t

s
ds
_
,
is equal to
E
1(Tt)
[
y
0
=
t
that is, y
0
is replaced by
t
and by T t.
A.2 The process (8)
Suppose now that
dy
t
= (b
1
+ b
2
y
t
)dt + b
3
y
1/2
t
dW
t
(57)
and dene E
1
, E
1
, E
2
, E
2
as before. Also dene z
t
= y
2
t
so that
dz
t
= (2y
t
(b
1
+ b
2
y
t
) + b
2
3
y
t
)dt + 2b
3
y
3/2
t
dW
t
= ((2b
1
+ b
2
3
)y
t
+ 2b
2
z
t
)dt + 2b
3
y
3/2
t
dW
t
.
Then, proceeding as above, we nd linear ordinary dierential equations rst for
E
1
, then E
1
, E
2
and E
2
, yielding
dE
1t
= (b
1
+ b
2
E
1t
)dt,
26
so that
E
1
= y
0
e
b
2

b
1
b
2
(1 e
b
2

),
E
1
=
y
0
b
2
(e
b
2

1)
b
1
b
2
2
(b
2
e
b
2

+ 1), (58)
and then
dE
2t
dt
= (2b
1
+ b
2
3
)E
1t
+ 2b
2
E
2t
= (2b
1
+ b
2
3
)
_
y
0
e
b
2
t

b
1
b
2
(1 e
b
2
t
)
_
+ 2b
2
E
2t
=
1
+
2
e
b
2
t
+ 2b
2
E
2t
,
where

1
=
b
1
(2b
1
+ b
2
3
)
b
2
,
2
= y
0
(2b
1
+ b
2
3
)
1
.
Integrating, we have
E
2
=

1
2b
2

2
b
2
e
b
2

+
_
y
2
0
+

2
b
2
+

1
2b
2
_
e
2b
2

and so
E
2
=

2b
2

2
b
2
2
(e
b
2

1) +
1
2b
2
_
y
2
0
+

2
b
2
+

1
2b
2
_
(e
2b
2

1).
A.3 Popular Stochastic Volatility Models
For completeness we give the results of these calculations for some popular stochastic
volatility models:
(i) The Hull-White model [14]. This is a geometric Brownian motion for the variance

2
t
:
d
2
t
=
2
t
dt +
2
t
dW
t
.
We use (51) with y
t
=
2
t
, so that

0
= 0,
1
= .
We obtain
E
1
=
2
0
e

, E
1
=

2
0

(e

1) . (59)
(ii) Analogous to (i), we consider a geometric random walk for the volatility,
d
t
=
t
dt +
t
dW
t
.
27
Then in (51) we have
0
= 0,
1
= ,
0
= 0,
1
= 0,
2
= 2 +
2
, and we have
E
1
=
0
e

, E
1
=

0

(e

1) ,
and
E
2
=
2
0
e
(2+
2
)
, E
2
=

2
0
2 +
2
_
1 e
(2+
2
)
_
.
(iii) The mean reverting-version of the Ornstein-Uhlenbeck model for the volatility
is
d
t
= (
t
) dt + dW
t
.
We take y
t
=
t
,
0
= ,
1
= ,
0
=
2
,
1
= 2 and
2
= 2 in (51). Then
it is straightforward to show that
E
1
= (
0
)e

+ , E
1
=

0

_
1 e

_
+ ,
and that
E
2
=
2
2
+
2
2
_
1 e
2
_
+ 2 (
0
)
_
e
2
e

_
+
2
0
e
2
,
E
2
=
_
2
2
+
2
2
_
+
_

2
0
2 (
0
)
2
2
+
2
2
__
1 e
2
2
_

2 (
0
)

_
1 e

_
.
(iv) The Heston model [11]. This is
d
2
t
=
_

2
t
_
dt +
t
dW
t
.
In this case we use (57) with
y
t
=
2
t
, b
1
= , b
2
= , and b
3
= .
Then from (58) we have:
E
1
=
2
0
e

+
_
1 e

_
, E
1
=

2
0

_
1 e

_
+ .
(v) Finally we consider the mean reverting log-normal model [13]
d
t
= (
t
)dt +
t

t
dW
t
.
We use (51) with y
t
=
t
,
0
= ,
1
= ,
0
= 0,
1
= 2,
2
= 2+
2
. As
expected, we obtain the following expressions:
E
1
=
0
e

+
_
1 e

_
,
E
1
=

0

_
1 e

_
+ ,
28
E
2
=
2
2
2
2
_
1 e
(2
2
)
_
+
2 (
0
)

2
_
e

e
(2
2
)
_
+
2
0
e
(2
2
)
,
E
2
=
2
2
2
2
+
_
2
2
2
2
+
2(
0
)

2

2
0
_
_
e
(2
2
)
1
2
2
_

2(
0
)

2
_
1 e

_
.
B Appendix 2: Analysis for pure volatility prod-
ucts
Following 4.4, we have L
0
V
0
= 0, so V
0
= V
0
(S, t); similarly L
0
V1
2
= L1
2
V
0
= 0,
so V1
2
= V1
2
(I, t). From the solvability condition for (42), we have
L
1
V
0
=
V
0
t
+ F
V
0
I
rV
0
= 0
whose solution with V
0
(I, T) = P(I) is
V
0
(I, t) = e
r(Tt)
P
_
I + (T t)F
_
.
Then, solving (42),
V
1
= f
1
()
V
0
I
+ H(I, t).
where H is as yet undetermined.
Now the solvability condition for (43) gives
LV1
2
= f

1
V
0
I
so that the solution satisfying V1
2
(I, T) = 0 is
V1
2
(I, t) = (T t) f

1
V
0
I
.
Thus far we have paralleled the analysis of 4.4. Now we continue by nding the
solution of (43): it is
L
0
V3
2
= L1
2
V
1
L
1
V1
2
=
_
f

1
f

1
_
V
0
I
(T t) f

1
_
F F()
_

2
V
0
I
2
29
so that
V3
2
= f
3
()
V
0
I
(T t) f

1
f
1
()

2
V
0
I
2
+ H3
2
(I, t)
where f
3
() satises
1
2

2
d
2
f
3
d
2
+ m
df
3
d
= f

1
() f

1
and H3
2
is arbitrary. Now at O(),
L
0
V
2
+L1
2
V3
2
+L
1
V
1
= 0;
our nal application of the solvability condition, in the form L1
2
V3
2
+L
1
V
1
, p

) = 0,
gives
L
1
_
f
1
()
V
0
I
+ H
_
, p

) = L1
2
V3
2
, p

),
that is,
L
1
H =
_
A
2
(T t)
_
f

1
_
2
_

2
V
0
I
2
+ f

3
V
0
I
,
where A
2
is dened in (29). The solution is
H(I, t) =
_
A
2
(T t) +
_
f

1
_
2 (T t)
2
2
_

2
V
0
I
2
f

3
(T t)
V
0
I
+ H
1
(I, t),
where L
1
H
1
= 0, so that H
1
= H
1
_
I + (T t)F
_
; this last unknown function is
determined by matching with the boundary layer.
In the boundary layer, we have
_
1

L
0
+
1

1/2
L1
2
+

L
1
_
_

V
0
+
1/2

V1
2
+

V
1
+
_
= 0,
as before, and it is easy to see that

V
0
(I, t) = P(I),

V1
2
(I, t) = 0, and that, as before,

V
1

+L
0

V
1
=
_
F() F()
_
dP
dI
L
1
P
so that, as above,

V
1
(I, )
_
f
1
() f
1
()
_
dP
dI
L
1
P
as . Hence the matching condition is unaected at this order by the market
price of risk term, and the required nal condition for H
1
(I, t) is
H
1
(I, T) = f
1
()
dP
dI
.
30
Hence H
1
(I, t) = e
r(Tt)
f
1
()P

(I + (T t)F), where

= d/dI.
Acknowledgments SDH wishes to thank the Universita degli Studi di Siena for
hospitality during the completion of this work; AR wishes to thank the participants
of the second world congress of the Bachelier Finance Society, Crete, June 2002, for
stimulating discussions. AR acknowledges nancial support from the School of Phys-
ical Sciences and Engineering, Kings College London. We are grateful to George
Papanicolaou for helpful comments and especially for drawing the reference [2] to
our attention.
References
[1] Barndor-Nielsen, O. E. and Shephard, N. Realised Power Variation and
Stochastic Volatility Models, MPS-RR 2002-06, MaPhySto, Aarhus (2001).
[2] Bensoussan, A., Lions, J.L. & Papanicolaou, G.C. Boundary Layers and Ho-
mogenisation of transport Processes, Publ. RIMS Kyoto Univ. 15, 53157
(1979).
[3] Bjork, T. Arbitrage Theory in Continuous Time. Oxford University Press, 1998.
[4] Brockhaus, O. and Long, D. Volatility Swaps Made Simple, Risk, 2(1) 92-95,
1999.
[5] Buraschi, A. and Jackwerth J. C. The Price of a Smile: Hedging and Spanning
in Option Markets, Review of Financial Studies, 14(2) 495-527, 2001.
[6] Chriss, N. and Moroko, W. Market Risk for Volatility and Variance Swaps.
Risk, July 1999.
[7] Chicago Board of Options Exchange website: www.cboe.com.
[8] Demeter, K., Derman, E., Kamal, M. and Zou, J. More than You ever Wanted
to Know about Volatility Swaps. Goldman Sachs Quantitative Strategies Re-
search Notes, 1999.
[9] Detemple, J. and Osake, C. The Valuation of Volatility Options, Working Paper,
Montreal, December 1999.
[10] Fouque J. P., Papanicolaou G. and Sircar, K. R. Derivatives in Financial
Markets with Stochastic Volatility, Cambridge University Press, 2000.
[11] Heston, S. L. A closed-form Solution for Options with Stochastic Volatility with
Applications to Bond and Currency Options. Review of Financial Studies, 6(2):
237-343, 1993.
[12] Heston, S. L. Derivatives on Volatility: Some Simple Solutions Based on Ob-
servables. Federal Reserve Bank of Atlanta, Working Paper Series, November
2000.
31
[13] Howison, S. D., Rafailidis, A. and Rasmussen, H. O. A Note on the Pricing
and Hedging of Volatility Derivatives. Bachelier Finance Society Second World
congress proceedings, Crete, June 2002.
[14] Hull, J. and White, A. The Pricing of Options on Assets with Stochastic Volatil-
ities. Journal of Finance, 42(2): 281-300, 1987.
[15] Javaheri, A., Wilmott, P. and Hong, E. G. Garch and Volatility Swaps,
www.wilmott.com, 2002.
[16] Jex, M., Henderson, R. and Wang, D. Pricing Exotics Under the Smile. Risk
November 1999 and JP Morgan Derivatives Research.
[17] Lipton, A. Mathematical Methods for Foreign Exchange. World Scientic, 2001.
[18] Naik, V. Option Valuation and Hedging Strategies with Jumps in the Volatility
of Asset Returns. Journal of Finance, 48, 1969-84, 1993.
[19] Rasmussen, H. O. and Wilmott, P. Asymptotic Analysis of Stochastic Volatility
Models. In New Directions in Mathematical Finance, Eds. P. Wilmott and H.
O. Rasmussen, Wiley, 2002.
[20] Van Dyke, M. Perturbation Methods in Fluid Dynamics, Parabolic Press, 1975.
[21] Wilmott P. Derivatives: The Theory and Practice of Financial Engineering.
John Wiley and Sons, 1997.
32