Вы находитесь на странице: 1из 30

An Asymptotic Expansion with Push-Down of Malliavin Weights

Akihiko Takahashi and Toshihiro Yamada October 27, 2011

Abstract This paper derives asymptotic expansion formulas for option prices and implied volatilities as well as the density of the underlying asset price in multi-dimensional stochastic volatility models. In particular, the integration-byparts formula in Malliavin calculus and the push-down of Malliavin weights are eectively applied. We provide an expansion formula for generalized Wiener functionals and closed-form approximation formulas in stochastic volatility environment. In addition, we present applications of the general formula to expansions of option prices for the shifted log-normal model with stochastic volatility. Moreover, with some results of Malliavin calculus in jump-type models, we derive an approximation formula for the jump-diusion model in stochastic volatility environment. Some numerical examples are also shown. Keywords: Malliavin calculus, Asymptotic expansion, Stochastic volatility, Implied volatility, Shifted log-normal model, Jump-diusion model, Integration-by-parts, Malliavin weight, Push-down, Malliavin calculus for Poisson processes

Introduction

This paper develops an asymptotic expansion method for generalized Wiener functionals by applying Malliavin weight (divergence given by the integration-by-parts formula) and push-down (the conditional expectation in Malliavin (1997) and Malliavin-Thalmaier (2006)). As applications, we propose a concrete approximation formula of option prices as well as the density of the underlying asset price in multi-dimensional stochastic volatility models, and then derives a new Taylor expansion formula of the implied volatilities. Moreover, we present applications of the general formula to expansions of option prices for the shifted log-normal model in stochastic volatility environment. Also, combining some results of Malliavin calculus in jump-type models by Bavouzet-Messaoud (2006) with our method, we derive an approximation formula for option prices in the jump-diusion model with stochastic volatility. To the best of our knowledge, it is the rst study with push-down of Malliavin weights for deriving analytical approximation formulas for option prices and implied volatilities in those models. A companion paper, Shiraya-Takahashi-Yamada (2009) applies the method to deriving a concrete approximation formula for valuation of barrier options with discrete monitoring under stochastic volatility models. Various stochastic volatility models has been proposed for calibration to market prices of options with so called volatility skews and smiles. However, closed-form solutions for option prices in the stochastic volatility environment are rarely found, and hence a large number of studies have been made in order to obtain analytical approximations and ecient numerical schemes for option prices and implied volatilities with stochastic volatility models. Takahashi (1995, 1999) proposed approximation formulas based on the asymptotic expansion method in Watanabe theory (Watanabe (1983, 1984, 1987), Yoshida (1992 a, b)) for valuation of various derivatives including options in stochastic volatility environment. Fourni et al. (1997) provided an expansion result of the second-order e partial dierential equation which satises the uniform ellipticity condition and showed its application to Monte Carlo simulations. Fouque et al. (2000) derived a closed form approximation formula for the fast mean reverting stochastic volatility model using a singular perturbation method, and then discussed the calibration problem. Hagan et al. (2002) introduced the SABR stochastic volatility model and obtained an approximation implied volatility formula. Labordre (2008) generalized SABR model to -SABR model and derived an approximation of implied e volatilities by applying the heat kernel expansion. Moreover, Gatheral et al. (2009) and Ben Arous-Laurence (2009) presented novel results for this direction. Recently, Antonelli-Scarlatti (2009) gave a Taylor series expansions of option prices with respect to a correlation parameter in a stochastic volatility model. Also, we will provide more detail comments on some related works in Section 3.3.
Forthcoming in SIAM Journal on Financial Mathematics. All the contents expressed in this research are solely those of the authors and do not represent the view of MTEC Co., Ltd. or any other institutions. The authors are not responsible or liable in any manner for any losses and/or damages caused by the use of any contents in this research. Graduate Mitsubishi

School of Economics, the University of Tokyo UFJ Trust Investment Technology Institute Co.,Ltd. (MTEC).

The organization of the paper is as follows: the next section derives an asymptotic expansion formula for generalized Wiener functionals after a brief summary of Malliavin calculus necessary for the remaining of the paper. Section 3 applies the general formula to pricing options in stochastic volatility environment and then obtains implied volatilities expansions. Section 4 presents numerical examples for expansions including calibrations of implied volatilities. As simple applications of our method, Section 5 provides approximation formulas for option prices (or/and implied volatilities) in the shifted log-normal and jump-diusion models under stochastic volatilities. Section 6 concludes. Appendix gives explicit calculations of push-down of Malliavin weights (coecients in the expansions) as well as the proof of Proposition 5.2.

2
2.1

Asymptotic Expansion
Malliavin calculus

This subsection summarizes basic facts on the Malliavin calculus which are necessary for the following discussion. We use the notations and denitions given below. Let (W, ) be the Wiener space, i.e. W = W d = C0 ([0, T ] : Rd ) = {w : [0, T ] Rd ; continuous, w(0) = 0} and is the Wiener measure. Next, let H be a Hilbert space such that

{
H= h W; hi (t)(i = 1, , d) is absolute continuous with respect to t and

d i=1 0

dhi (t) dt

}
dt <

i (t) i with an inner product h, h H = i=1 0 dhdt dt. Then, H is called the Cameron-Martin space. dt p Dene L (W) as L (W) = p<+ L (W) and a distance on L (W) as

dh (t)

dL (W) (F1 , F2 ) =

j=1

2j (min{F1 F2 Lj , 1}),

where Lp denotes the Lp -norm in (W, ). Given a separable Hilbert space G, let Lp (W : G) denote the space of measurable maps from W to G such that f G Lp (W). The same denition is made for L (W : G). Then, consider the space Dp (W : G) 1

{
=

F Lp (W : G) : there exists DF Lp (W : H G) such that for h H, lim

1 [F (w + h) F (w)] = (DF (w))(h) .

Here, DF is called the (Malliavin) derivative of F . Note also that the tensor product H G of two separable Hilbert spaces H and G is a Hilbert space formed of all linear operators A : H G of Hilbert-Schmidt type endowed with the Hilbert-Schmidt norm

(
AHS =

)1/2
Aei , e 2 j G ,

i,j=1

for some ONB(orthonormal basis)s {ei } in H and ONBs {e } in G, and j A2 = A2 HS HG . Then, the random variable DF takes values in H G. Due to a canonical identication of the Hilbert space L2 (W : H) and L2 ([0, T ] W), the Malliavin derivative DF may be considered as a stochastic process {Dt F = (Dt,1 F, , Dt,d F ) : t [0, T ]} such that (DF (w))(h) = DF, hH =
d i=1 0 T

(
(Dt,i F )

dhi (t) dt

)
dt.

A norm Dp (W : G) is given by F Dp (W:G) = F Lp (W:G) + DF Lp (W:HG) . Also, D (W : G) is dened 1 1 by D (W : G) := 1p<+ Dp (W : G), and a distance on D (W : G) is given by 1 1 1 dD (W:G) (F1 , F2 ) =
1 1

j=1

2j (min{F1 F2 Dj (W:G) , 1}).


1

For r 2 (r N), we introduce the spaces: Dp (W : G) = {F Dp (W : G) : DF Dp (W : H G)} r r1 r1 with F Dp (W:G) = F Dp r G). If G = Rn , Then we write Dp (W) for Dp (W : G). r r Some properties of these spaces are the following; Dp (W) Dp (W), r r, and p p. The dual space of r r p q q q (Dr (W)), (Dr (W)) is given by (Dr (W)) =Dr (W), with p1 + q 1 = 1, r 0. Furthermore, dene the space D (W) = p1,r0 Dp (W). Then, D (W) is a complete metric space under a r metric, dD (W) (F1 , F2 ) =

r1

(W:G)

+ Dr1 F Dp (H (r1) G) . We also dene Dp (W : G) as Dp (W : G) = Lp (W : 0 0


1

2pr (min{F1 F2 Dp , 1}). r

p,r=1

We call F D (W) the smooth functional in the sense of Malliavin. p Dp Given Z = (Z1 (w), , Zd (w)) 1 (W : H), there exists Di (Zi ) L (W), i = 1, , d such that d T E[ 0 Dt,i F (w)Zi (w)dt] = E[F (w)Di (Zi (w))] for all F D (W). Then, dene D (Z) := i=1 Di (Zi (w)). So, 1 there exists Cp > 0 such that D (Z)Lp Cp ZDp (W:H) . We call D (Z) the divergence of Z.
1

We also introduce the notation DZ F such that DZ F := E[F D (Z)]. Thus,

T
0

i=1

Dt,i F (w)Zi (w)dt. Then, we have E[DZ F ] =

D (F Z) = F D (Z) DZ F is obtained. (e.g. Proposition 1.16 in Malliavin-Thalmaier (2006)) Denition 2.1 Let F = (F1 , , Fn ) D (W : Rn ) be the n-dimensional smooth functional, we call F a nonij degenerate in the sense of Malliavin if the Malliavin covariance matrix {F }1i,jn
ij F

:= DFi , DFj H =

d k=1 0

(Dt,k Fi (w))(Dt,k Fj (w))dt

(1)

is invertible a.s. and (det F )1 L (W). Theorem 2.1 Let F D (W : Rn ) be a n-dimensional non-degenerate in the sense of Malliavin and G 1 D (W). Then, for Cb (Rn ), E[i (F )G] = E[(F )D (
n j=1 F where (ij )1i,jn is the inverse matrix of Malliavin covariance of F . F Gij DF j )]

(2)

(Proof ) See Lemma III.5.2. of Malliavin (1997). 2 Theorem 2.2 Let F D (W : Rn ) be a non-degenerate functional. F has a smooth density pF S(Rn ) where S(Rn ) denotes the space of all innitely dierentiable functions f : Rn R such that for any k 1, and for any multi-index {1, , n}j one has supxRn |x|k | f (x)| < . (i.e. S(Rn ) is the Schwartz space and S (Rn ) is its dual.) (Proof ) See Theorem III.5.1. of Malliavin (1997). 2 Denition 2.2 Consider the space D (W) = p1,r0 Dp (W), that is, the dual of D . We call F D (W) r a distribution on the Wiener space. We dene the duality form on D D , (F, G) F, GD D = E[F G] R. We call this duality form the generalized expectation.

2.2

Asymptotic Expansion for Expectation of Generalized Wiener Functionals

Let F D (W : Rn ) be a non-degenerate functional, and and pF be the law and the smooth density of F , respectively; that is, (dx) = F 1 (dx) = pF (x)dx. Also, we dene the range O as O := {x : pF (x) > 0} Rn . By Malliavin (1997) and Malliavin-Thalmaier (2006), the conditional expectation of g Lp (W, ) conditioned by a set {w : F (w) = x} in -eld (F ), E[g|F = x] gives a map, E F : Lp (W, ) g E[g|F = x] Lp (O, ). (3)

For multi-index (k) = (1 , , k ), we dene the iterated Malliavin weight. The Malliavin weight H(k) is recursively dened as follows: for G D , H(k) (F, G) = H(k ) (F, H(k1) (F, G)), (4)

where H(l) (F, G) = D

( n
i=1

)
F Gli DFi

(5)

F Here, F = {ij }1i,jn denotes the inverse matrix of the Malliavin covariance matrix of F .

Watanabe (1983, 1984) introduced the distribution on Wiener space as composition of a non-degenerate map F by a Schwartz distribution T . The next theorem restates the result of Watanabe (1984) in terms of Malliavin (1997) and Malliavin-Thalmaier (2006). Theorem 2.3 1. Let S be the Schwartz distributions. There exists a map (E F ) : S T T F D := s0 q1 Dq D . s (E F ) is called the lifting up of T . 2. The conditional expectation denes a map E F : D G E F [G] S(O), (7) (6)

where S(O) stands for the Schwartz space of the rapidly decreasing functions on O = {x : pF (x) > 0} Rn . We call this map the push down of G. 3. The following duality formula is obtained: (E F ) T, GD D = T, E F [G]pF (x)dx where , pF (x)dx is dened as follows: T, E F [G]pF (x)dx = S T, E F [G]pF S . (9) (8)

(Proof ) In this proof, we apply the discussions of Watanabe (1984), Malliavin (1997), Malliavin-Thalmaier (2006) and Nualart (2006). 1. Given T S , there exists Tn S such that Tn T in S , i.e. for m 1, Am Tn Am T 0, where f = supxRn |f (x)|, A = 1 + |x|2 , = formula, we can estimate as follows; for p Tn (F ) Tn (F )Dq
2m

n ,
2 . i=1 x2 i

(10)

n 1
2

By the Malliavin integration-by-parts

+q

= 1, |E[Tn (F )G] E[Tn (F )G]| E[(Am {Am Tn })(F )G] E[(Am {Am Tn })(F )G]
F F E[{Am Tn }(F )(2m) (G)] E[{Am Tn }(F )(2m) (G)] F Am Tn Am Tn (2m) (G)L1

=
p

sup
GD2m ,GDp 1
2m

= = =

sup
p GD2m ,GDp 1 2m

sup
p GD2m ,GDp 1 2m

sup
p GD2m ,GDp 1 2m

CAm Tn Am Tn 0,
2m

(11)
1 F (2m) (G)L1 < . Here, as in p.379 of
2m

F as n, n , where (2m) (G) D and C := supGDp

,GDp

Ikeda-Watanabe (1989) or Theorem 3.2.1 with the equations (3.2.1) and (3.2.2) of Sakamoto-Yoshida (1994), we have used the relation:
F E[(Am {Am Tn })(F )G] = E[{Am Tn }(F )(2m) (G)], F where (2m) (G) is recursively obtained by n i=1 F F F (2(m+1)) (G) = (2) ((2m) (G)), m 1.

(12)

F (2) (G)

= (1 + |F | )G 1/2
2

H(i,i) (F, G), (13)

Then (Tn (F ))nN is a Cauchy sequence in D and thus there exists (E F ) T = T (F ) D : a composite functional T (F ) is uniquely determined.

2. Given G D . For any multi-index s = (s1 , , sk ), For any Cb (Rn ), by push-down and then the integration by parts formula on Rn or the integration by parts formula on W and then push-down, we obtain
|s| E[s (F )G]

|s|

= (1)

|s|

|s| (x)s {E F =x [G]pF (x)}dx

=
Rn

F (x)E F =x [s ]pF (x)dx,

(14)

Rn F where s = Hs (F, G) D . It implies that

|s| F (1)|s| s {E F =x [G]pF (x)} = E F =x [s ]pF (x),

(15)

where
F E F =x [s ] Lp (O, ).

We dene O as O = {x Rn : pF (x) > }. Therefore,


|s| (1)|s| s {E F =x [G]pF (x)} Lp (O , dx),

(16)

for all s, which implies E F =x [G]pF (x) C (O). As pF (x) C (O) where C (O) stands for the set of real-valued C -functions on O, (pF (x))1 {E F =x [G]pF (x)} = E F =x [G] C (O). Note also that the conditional expectation has the following expression: E F =x [G]pF (x) = E[1{F >x} H(1,,n) (F, G)]. Thus, for all k N and for all j = 1, , n, if xj > 0, sup
xRd ,xj >0 |s| x2k s {E F =x [G]pF (x)} j

(17)

(18)

= <

sup
xRd ,xj >0

x2k |E[1F >x Hs (F, H(1,,n) (F, G))| j

E[|Fj |2k |Hs (F, H(1,,n) (F, G))|] , (19)

if xj < 0, we can derive a similar estimate. These facts imply E F =x [G] S(O). 3. Hence, because there exists Tn S, n N for T S such that Tn T in S , we have Tn , E F [G]pF (x)dx = E[Tn (F )G] (E F ) T, GD D = T, E F [G]pF (x)dx , as n . 2 Hereafter, we use the notation (20)

T (x)p(x)dx for T S (Rn ) and p S(Rn ) meaning that

S T, pS .

The next theorem presents an asymptotic expansion formula for the expectation of generalized Wiener functionals.
Theorem 2.4 Consider a family of smooth Wiener functionals F = (F1 , , Fn ) D (W : Rn ) such that F has an asymptotic expansion in D and satises the uniformly non-degenerate condition:

lim sup (det F )1 Lp < , f or all p < .


0

(21)

Then, for a Schwartz distribution T S (Rn ), we have an asymptotic expansion in R:

{
E[T (F )]
Rn

T (x)p

F0

(x)dx +

N j=1

j Rn

T (x)E

[ (j)
k

H(k) (F ,

k l=1

]
0, Fl l )|F 0

}
F0

=x p

(x)dx

= O(N +1 ), (22)

where pF is the density of F 0 , and Fi0,k := (k) = (1 , , k ) and


(j) k

1 dk F | , k! dk i =0

k N, i = 1, , n. Also, (k) denotes a multi-index,

k=1 1 ++k =j,i 1 (k) {1,,n}k

1 . k!

Moreover, Malliavin weight H(k) is recursively dened as follows: H(k) (F, G) = H(k ) (F, H(k1) (F, G)), where H(l) (F, G) = D

(23)

( n
i=1

)
F Gli DFi

(24)

F Here, F = {ij }1i,jn denotes the inverse matrix of the Malliavin covariance matrix of F .

(Proof) We use as an abbreviation of (k) in the proof. Under the uniformly non-degenerate condition of F D (W : Rn ), the lifting up of T S (Rn ), (E F ) T , has the asymptotic expansion in distributions on the Wiener space D , i.e. for N N, there exists s N s.t. (E F ) T {T F 0 +

N j=1

(j) k

k ( T ) F 0

k l=1

0, Fl l }
q Ds

= O(N +1 ),

(0, 1], q <

(25)

Then, there exists an asymptotic expansion of (E F ) T, 1D D . The push-down of the Malliavin weights are computed as follows:

k T (F 0 ),

k l=1

0, Fl l D D

= T (F ), H (F ,
0 0

k l=1

0, Fl l )

= T, E F [H (F 0 ,
0

k l=1

0, Fl l )]

D D

=
Rn

pF (x)dx k 0, Fl l )|F 0 = x]pF (x)dx.


0

T (x)E[H (F 0 ,

l=1

(26)

2 Corollary 2.1 The density pF (y) is expressed as following asymptotic expansion with the push-down of Malliavin weights: p
0

(y) = p

F0

(y) +

N j=1

[ (j)
k

H(k) (F ,

k l=1

]
0, Fl l )|F 0

= y pF (y) + O(N +1 ),

(27)

where pF (y) is the density of F 0 . (Proof) Take a delta function y S in the theorem above. 2

3 Asymptotic Expansion in Multi-Dimensional Stochastic Volatility Model


This section applies the general formula in the previous section to pricing options in a stochastic volatility model and then obtains a new implied volatilitys expansion formula.

3.1

Asymptotic Expansion of Option Prices

This subsection proves a basic result on an asymptotic expansion for a stochastic volatility model. Let (, F , (Ft )t[0,T ] , P ) be a ltered probability space and W = {(W1,t , , Wd,t ) : 0 t T } be a ddimensional Brownian motion with respect to (Ft )t[0,T ] . We consider the following stochastic volatility model (n-dimensional volatility factor). dSt dt
()

= rSt dt + V (t )St dW1,t , = A0 (t )dt + A(t )dWt ,


(0) () ()

()

()

()

()

S0 = S0

()

= s > 0, 0 = 0

()

(0)

= Rn , +

(28)

where V Cb (Rn R), A0 Cb (Rn Rn ), A Cb (Rn Rnd ), r > 0 and [0, 1]. Note that is (0) the volatility of volatility parameter. (t )t[0,T ] is a deterministic process and satises an ordinary dierential equation,

dt Assumption 3.1 For some s [0, T ], V We dene the logarithmic process of


(0) (s )

(0)

= A0 (t )dt.
() ST

(0)

(29) at = 0.

Next, we impose the following condition which gives the non-degeneracy of the Malliavin covariance of = 0. as = log
() (St )t[0,T ]

() Xt

St s

()

)
.

Let pSV (y) be the density of the underlying asset of the stochastic volatility model and C SV (T, K) and P SV (T, K) be the call and the put option prices under the stochastic volatility with maturity T and strike price K. Also let pBS (y) is the log-normal density of the Black-Scholes model, i.e.

p
BS

BS

(y)

:= y

T
0

1 V (t )2 dt
(0)

exp

T 2 0

y log (0) 2 s V (t ) dt

( )

)2

C (T, K, ) and P (T, K, ) denote the Black-Scholes formula of the call and the put options with maturity T and strike price K, i.e. C BS (T, K, ) P BS (T, K, ) where n(x) N (x) d1 d2 with :=
()

BS

:= :=

sN (d1 ) KerT N (d2 ), KerT N (d2 ) sN (d1 ),


x2 1 e 2 , 2

(30)

:= :=

n(y)dy,

:= :=

+ rT 1 K + T, 2 T (s) log K + rT 1 T, 2 T log

(s)

(31)

1 T

)1/2
V
(0) (t )2 dt

(32)

Let S (0) be the Malliavin (co)variance of ST at = 0, i.e.


T

S (0) = D1 ST 2 = (ST )2 H
T

T (0) V (s )2 ds,

(0)

(0)

(33)

We introduce the expressions; Z(t) St


(k)

:= = =

Dt,1 ST

(0)

(0) D1 ST 2 H

ST V (t ) (ST )2
(0)

(0)

(0)

T
0

V (u )2 du

(0)

1 k () S |=0 , k! k t
k i=1

1 ,,k (t) where l 1 satisfy

St

(i )

(34)

k l=1

l = j,

j N,

1 k j.

Given Z above, for D , D (Z) DZ : D D is expressed as (D (Z) DZ ) = Then, we have the following result.

Z(t)dW1,t
0 0

Dt,1 Z(t)dt.

(35)

Theorem 3.1 Under the stochastic volatility model (28), we have asymptotic expansions of the density and the option prices as follows: pSV (y) = pBS (y) +
N j=1

j E[j |ST = y]pBS (y) + O(N +1 ),


N j=1

(0)

j
R+

C SV (T, K)

C BS (T, K, ) +

erT (y K)+ E[j |ST = y]pBS (y)dy + O(N +1 ),


(0)

P SV (T, K)

P BS (T, K, ) +

N j=1

j
R+

erT (K y)+ E[j |ST = y]pBS (y)dy + O(N +1 ),


(0)

(36)

where j is the j th -order Malliavin weight, i.e. j with


(0) H1 (ST , 1 ,,k (T ))

k=1 1 ++k =j,i 1

1 (0) Hk (ST , 1 ,,k (T )) D , k!

= =
1 ,,k

Z(s)dW1,s
0

(T )
0

Ds,1 1 ,,k (T )Z(s)ds, (37)

Hk (ST , 1 ,,k (T ))

(0)

(D (Z) DZ )k 1 ,,k .

(Proof) Let (Yt )t be the solution of the following stochastic dierential equation; dYji (t) Yji (0) where ai (s) k,l bi (s) k
i j

= =

d n l=1 k=1 i j ,

ai (s)Yjk (s)dWtl + k,l

n k=1

bi (s)Yjk (s)ds, k

(38) (39)

= =

() k Ai (s ), l () k Ai (s ), 0

(40) (41)

and is the Kroneckers delta. Let Ds,k , k = 1, , d be the Malliavin derivative acting on the Brownian motion () Wk,t . Then, the Malliavin covariance of ST is given by; S ()
T

d i=1 () 0

(Ds,i ST )2 ds
d i=1 0 T

()

= where
() Ds,1 XT

(ST )2

(Ds,i XT )2 ds,

()

(42)

() (s )

n i=1 s

T () () V (u )i V (u )Ds,1 i,u du ()

+
() Ds,k XT

n i=1 n s

() i V (u )Ds,1 i,u dW1,u , T () () V (u )i V (u )Ds,k i,u du s T () i V (u )Ds,k i,u dW1,u . () ()

()

i=1

+
() V (x) | () xi x=u

n i=1 s

(43)

Here, i V (u ) =

and for s u, Ds,k i,u =


() n l,j=1 () Yli (u)Y 1 (s)l Aj (s ). j k

(44)

Assumption 3.1 yields the nondegeneracy of ST ; 1 Lp < , f or all p < . (0)


ST

(0)

(45)

Then, the similar argument to Takahashi-Yoshida (2004), we have an asymptotic expansion in D ; y (ST ) erT (ST K)+
() ()

y (ST ) +

(0)

N i=1

k=1 1 ++k =j,i 1 N i=1

1 k (0) y (ST )1 ,,k (T ) + O(N +1 ), k!

(46)

erT (ST K)+ +


(0)

k=1 1 ++k =j,i 1

1 rT k (0) e (ST K)+ 1 ,,k (T ) + O(N +1 ). k! (47)

By the integration by parts formula,

k y (ST ), 1 ,,k (T )

(0)

D D

y (ST ), Hk (ST , 1 ,,k (T )) y , E


(0) ST

(0)

(0)


D D

= =

[Hk (ST , 1 ,,k (T ))]

(0)

pBS (x)dx

(0) (0) E[Hk (ST , 1 ,,k (T ))|ST

= y]pBS (y),

(48)

where k y (ST ) = Similarly, we have


(0)

k y (x)|x=S (0) . xk T

erT k (ST K)+ , 1 ,,k (T )


(0)

D D

=
R+

erT (y K)+ E[Hk (ST , 1 ,,k (T ))|ST = y]pBS (y)dy, (49)


(0) (0)

where

k (x K)+ |x=S (0) . xk T Then, we obtain asymptotic expansion formulas of the density and the option prices; k ST K
(0)

)+

SV

(y)

BS

(y) +

N j=1

j E[j |ST = y]pBS (y) + O(N +1 ),


N j=1

(0)

(50)

j
R+

C SV (T, K)

C BS (T, K, ) +

erT (y K)+ E[j |ST = y]pBS (y)dy + O(N +1 ),


(0)

(51)

P SV (T, K)

P BS (T, K, ) +

N j=1

j
R+

erT (K y)+ E[j |ST = y]pBS (y)dy + O(N +1 ).


(0)

(52)

The Malliavin weights are computed by the iterated Skorohod integrals.

(
H1

(0) ST , 1 ,,k

)
= = = D

)
1 ,,k (0) (0) DST S
T

(D (Z) DZ ) 1 ,,k

1 ,,k

(T )
0

Z(s)dW1,s
(0)

Ds,1 1 ,,k (T )Z(s)ds,

(53)

Hk (ST , 1 ,,k (T )) 2

(0)

= =

H1 (ST , Hk1 (ST , 1 ,,k (T ))) (D (Z) DZ )k 1 ,,k . (54)

(0)

3.2

Implied Volatility Expansion

This subsection derives an asymptotic expansion formula for the implied volatility in the stochastic volatility model considered in the previous subsection, where we obtained an approximation formula of a call option: C SV (T, K) = C BS (T, K, ) + C1 + 2 C2 + 3 C3 + O(4 ), (55)

where Ci =

e
R

rT

se

y+rT 1 2

)+

[
E i |

]
V
(0) (t )dW1,t

=y

2 1 1 e 2 y dy, 2

i = 1, 2, 3.
(0)

(56)

We obtain an asymptotic expansion formula of the implied volatility around =

( T 1
T 0

V (t )2 dt

)1/2
.

Theorem 3.2 Under the stochastic volatility model (28), an asymptotic expansion of the implied volatility is given by
IV

C1 + 2 (T, K) = + BS C ( )

C3 BS C ( )

C1 BS C ( )2

){

2 C2 1 C1 BS C ( ) BS BS 2 C ( )3 C ( )

}
3 1 C1 BS C ( ) BS 3! C ( )4

2 C2 1 C1 BS C ( ) BS BS 2 C ( )3 C ( )

}
BS C ( )

}
+ O(4 ), (57)

BS BS BS where C ( ), C ( ), C ( ) are derivatives of the Black-Scholes formula with respect to the volatility, i.e. BS C ( ) BS C ( ) BS C ( )

:= := :=

BS C |= = s T n(d1 ), 2 BS s T C |= = n(d1 )d1 d2 , 2 3 s T C BS |= = n(d1 ){d2 d2 d1 d2 d2 d2 }. 1 2 1 2 3 2

(58)

(Proof) Suppose that an implied volatility is expanded as; IV (T, K) = + 1 + 2 2 + 3 3 + O(4 ). Then we have C BS (T, K, IV (T, K)) = + (59)

BS BS C BS (T, K, ) + C ( )1 + 2 C ( )2 + BS BS 3 3 C ( ) + 1 2 C ( ) +

1 3 BS 1 C ( ) + O(4 ). 3!

1 BS C ( ) (1 )2 2

}
(60)

By the denition of the implied volatility in the stochastic volatility, i.e. C SV (T, K) = C BS (T, K, IV (T, K)), the approximation terms of the implied volatility are given by 1 2 3 2 = = = C1 , BS C ( )

2 C2 1 C1 BS C ( ) BS BS 2 C ( )3 C ( )

}
,

C3 BS C ( )

C1 BS C ( )2

){

2 C1 C2 1 BS C ( ) BS BS 2 C ( )3 C ( )

}
BS C ( )

3 C1 1 BS C ( ). BS 3! C ( )4

(61)

3.3

Comments on Related Works

Fourni et al. (1999) applied Malliavin calculus to ecient Monte Carlo estimators for computing Greeks of options e in the Black-Scholes framework, e.g. for dSt = rSt dt + () St dWt and () = + , (0) () E[erT (ST K)+ ] = E[erT (ST K)+ ]|=0 = E[erT (ST K)+ ]. (62)

The estimator is the so-called Malliavin weight. Subsequently, a number of papers extended the method. In particular, related to our work, Siopacha and Teichmann (2010) developed strong and weak Taylor methods for a system of stochastic dierential equations (SDEs) with perturbations. Especially, the weak Taylor method is based on the integration by parts on the Wiener space, which is a powerful tool for ecient Monte Carlo simulations and is general enough to be applied to multi-dimensional SDEs. As an example, they applied the method to swaptions under a market model of interest rates with one-dimensional Heston-type stochastic volatility and obtained an approximation of the option prices including the expectation of the divergence on the Wiener space. In the last step, they used Monte Carlo simulations for computing the option prices and demonstrated the eciency of their method comparing with well-known existing Monte Carlo schemes.

10

Lewis (2000) developed a new method for approximations of European option prices and implied volatilities. In particular, he proposed an expansion with respect to a stochastic volatility parameter, vol-of-vol ( below) under a general one-dimensional time-homogeneous diusion process of the stochastic volatility, which is considered as a special case of our setup. However, he took a dierent approach from ours to obtain an approximation: for example, he considered the following generalized Heston model : dSt = rSt dt + vt St dW1t , dvt = b(vt )dt + (vt )((v)dW1t +

1 (v)2 dW2t ),

(63)

where r and are constants. Also, W = (W1t , W2t ) is a two-dimensional Brownian motion and (v) stands for the instantaneous correlation between S and v. For valuation of call options, he used the following formula: C(T, K) = S0 KerT 2

i + 2 i 2

eikx

H(k, v, T ) dk, k2 ik

(64)

where x = log S/KerT , and H( called the fundamental transform) is the solution to H 1 2H H = 2 2 (v) 2 + (b(v) ik(v)(v) v) {(k2 ik)/2}vH, T 2 v v (65)

with the initial condition H(k, V, 0) = 1. Then, he gave an approximation for a call price showing a scheme for the expansion of H with respect to : H(k, v, T ) = H (0) (k, v, T ) + H (1) (k, v, T ) + 2 H (2) (k, v, T ) + . He also presented an explicit result of the expansion for the following model.
dvt = ( vt )dt + vt (dW1t +

(66)

1 2 dW2t ),

where , , and are some appropriate constants. This paper applies an asymptotic expansion approach to stochastic volatility models and derive the approximation terms including the Malliavin weights. Then, these weights are transformed into a nite-dimensional integration through the duality formula of Malliavin (1997); that is, for T S , E[T (ST )] = (E ST ) T, D D = T, E ST []p(x)dx =

T (x)E[|ST = x]p(x)dx,
R

(67)

where p(x) is the density of ST D . This formula suggests that an element of the distributions on the Wiener space D is the adjoint operator of the conditional expectation. It also shows that the push-down of the Malliavin weights (the conditional expectation of the divergence on the Wiener space) is an element of the Schwartz space S. Thus, applying both the integration by parts and the duality formula, we can obtain analytical approximations for density functions, option prices and implied volatilities. () () () As a simple example, for dSt = rSt dt + t St dW1t and dt = A0 (t )dt + A1 ( () )(dW1t + 1 2 dW2t ) for some real-valued functions A0 (x) and A1 (x), we can obtain an approximation for a call option as follows: C(T, K) = E[erT (ST K)+ ]
()

= = =

() E[erT (ST K)+ ]|=0 + O(2 ) (0) (0) E[erT (ST K)+ ] + E[erT (ST K)+ ] + O(2 ) E[erT (ST K)+ ] +
(0)

C BS (T, K) +

(
R

sex 2 K

)+

(x)

2 1 1 e 2 x dx + O(2 ), 2

(68) where C BS (T, K) represents a Black-Scholes call price and

(x) = E[|
0

t dW1,t = x] =

(0)

1 3 1 1 1 x 2 x2 2 3x + . 3

Here, and are some constants. (See Corollary A.1. in Appendix A for the detail.) Further, the integral in the last line of (68) can be explicitly given. Consequently, our method is a natural extension of Fourni et al. (1999) and Siopacha and Teichmann (2010) e in a sense that we obtain analytical approximations that do not rely on Monte Carlo simulations by making use of push-down of Malliavin weights. Although Lewis (2000) took a dierent approach, our method can be regarded as its extension because his models are included in our framework: in fact, Theorem 3.1 and 3.2 show that our method and formulas can be applied to multi-dimensional models including time-inhomogeneous ones in a unied way. That is, we can readily deal with multi-dimensional stochastic volatility models by the same procedure as in one-dimensional ones. A concrete example will be shown in 3.4.2 Double Heston model. In this regard, it seems not an easy task for other analytical approximations such as Lewis (2000), Hagan et al.(2002), Labordere (2008) and Antonelli-Scarlatti (2009) to apply the methods to higher dimensional models.

11

4
4.1

Numerical Examples
One-dimensional Stochastic Volatility Models
dSt
()

This subsection shows numerical examples for the following stochastic volatility model: = = (t ) St dW1,t , ( t )dt + (t ) (dW1,t +
() () () ()

dt

()

1 2 dW2,t ),

(69)

where all the parameters will be specied later. We set the risk-free interest rate to be zero ( i.e. r = 0 ). Remark In Section 4, we deal with various stochastic volatility models including the Heston-type square-root volatility. (i.e. (69) with = 1/2.) Although the square-root volatility model does not satisfy the regularity conditions stated in Section 3 since the coecient function of the Heston-type model A() = has an unbounded derivative at = 0, our method is formally applicable. A rigorous treatment for the application of our method to this case could be made in the similar manner as in Takahashi-Yoshida (2005) that utilizes a smooth modication technique: In fact, we take a modied process (t )t[0,T ] of square-root process (t )t[0,T ] : dSt
() ()

= = = =

(t ) St dW1,t , ( t )dt + (t ) St dW1,t ,


() () () ()

()

()

t (dW1,t +
() ()

dt dSt

1 2 dW2,t ),

()

( t )dt + g(t )(dW1,t + 1 2 dW2,t ). Here, g(x) is a smooth modication of x such that g(x) = x when x 1 for some small 1 > 0, and g(x) = 0 when x 2 for some 2 (0, 1 ). Specically, we can set g(x) as follows: g(x) = h(x) x (x 2 ) h(x) = , 0 < 2 < 1 (x 2 ) + (1 x) dt (x) = e1/x for x > 0, (x) = 0 for x 0.

()

() () () () Let XT = log(ST /s) and XT = log(ST /s). Then, ()

XT

= =

1 2 1 2

(t )2 dt +
()

(t ) dW1,t

()

(t )2 dt +
()

0 T

() XT

(t ) dW1,t
0

()

Suppose that for a R-valued[ function f ] which for example stands for an option payo; f (x) = (ex K)+ , (K ex )+ , [ ] E f (XT )
() 2

< and E

f (XT )

()

< . Then, we have

[
E

() f (XT )

() f (XT ) 1{() =() }

( [
E

() |f (XT )|2

]1 2

() + E |f (XT )|2

]1 ) 2

P ({ () = () }) 2 .

It also holds that


P ({ = }) = P ({t 1 for some t [0, T ]}) 0 P ({ sup |t t | > a}) 0tT 0 +P ({t 1 for some t [0, T ]} { sup |t t | a}). 0tT

We can easily see that the second term after the last inequality is 0. The rst term is smaller than any n for n = 1, 2, by the following lemma of a large deviation inequality: Lemma
Suppose that Zt , t [0, T ] follows a SDE: dZt = a(Zt )dt + b(Zt )dWt .

12

where a(z) satises Lipschitz and linear growth conditions, and b(z) satises the linear growth condition. We assume that the unique strong solution exists. Then, there exists positive constants c1 and c2 independent of such that 0 P ({ sup |Zs Zs | > c}) c1 exp(c2 2 ) (70)
0sT

for all c > 0. The lemma can be proved by slight modication of lemma 7.1 in Kunitomo and Takahashi (2003). Note also that and satisfy the conditions in the lemma above. Hence, [ ] () () E f (XT ) f (XT ) = o(n ), n = 1, 2, . (71) Therefore, the dierence between f (XT ) and f (XT ) is negligible in the small disturbance asymptotic theory. Then, the mathematical conditions in Section 3 are met and our expansion method is rigorously applicable. Hereafter, we use the following notations; n(x) N (x) := :=
T
x2 1 e 2 , 2

()

()

(72)

n(y)dy,

:= :=

(0) (0) (t )(1) et (t )

t (0) es (s )+ dsdt,

0 T

(t )

(0)

)2
dt,

[
() S |=0 T

(x)

:= =

T
0

Dt,1 ST

(0)

dW1,t
0

(Ds,1 ST )2 ds

(0)

(
d1 d2 C BS C1 vega := := := := :=

1 1 1 1 3 x 2 x2 2 3x + , 3

Dt,1 ST () dt| Ds,1 ST |=0 T (0) (Ds,1 S )2 ds


0 T

(0)

T (0) (t ) dW1,t

]
=x

1 T

(t )

(0)

)2

)1/2

dt

(
R

log 1 + T, 2 T (s) log K 1 T, 2 T sN (d1 ) KN (d2 ),


sex 2
1 2

( )
0 s K

)+

(x)

1 2 2 T

exp

:=

s T n(d1 ).

1 x2 dx, 2 2 T

By using Corollary A.1 in Appendix A, the rst order approximations of the asymptotic expansions of the option price and the implied volatility are given by Malliavin AE and Malliavin IV := + C1 . vega := C BS + C1 ,

First, we give the numerical results on the accuracy of our approximation formula for call option prices. T = 1, = 1/2, = 1/2, s = 100, (0 ) 2 = 0.3, = 2.0, = 0.09, = 0.1, = 0.5 Strike price 70 80 90 100 110 120 Benchmark 31.5478 23.6382 17.0487 11.8647 7.9947 5.2356 Malliavin AE 31.5496 23.6471 17.0631 11.8816 8.0105 5.2474 Relative error 0.00% 0.04% 0.08% 0.14% 0.20% 0.23%
1

13

Benchmark : Hestons Fourier transform solutions. T = 5, = 1, = 1, s = 100, 0 = 0.4, = 0.1, = 0.4, = 0.2, = 0.5 Strike price 70 80 90 100 110 120 Benchmark 46.4585 41.5476 37.1770 33.2892 29.8338 26.7639 Malliavin AE 46.3591 41.4427 37.0591 33.1521 29.6697 26.5645 Relative error -0.21% -0.25% -0.31% -0.41% -0.55% -0.74%

Benchmark : Monte Carlo simulation (1,000,000 trials, 500 time steps). T = 10, = 1, = 1, s = 100, 0 = 0.4, = 0.1, = 0.4, = 0.2, = 0.5 Strike price 70 80 90 100 110 120 Benchmark 55.3842 51.4141 47.8183 44.5521 41.5785 38.8648 Malliavin AE 55.2118 51.2058 47.5671 44.2513 41.2207 38.4432 Relative error -0.31% -0.10% -0.53% -0.68% -0.86% -1.10%

Benchmark : Monte Carlo simulation (3,000,000 trials, 1000 time steps). Next, we show the numerical results for implied volatilities. T = 0.25, = 1/2, = 1/2, s = 100, (0 ) 2 = 0.15, = 4.0, = 0.0225, = 0.1, = 0.5 Strike price 70 80 90 100 110 120 Exact IV 17.25 16.39 15.62 14.95 14.39 13.96
1 1

Malliavin IV 17.17 16.35 15.63 14.98 14.40 13.86

Relative error -0.46% -0.24% 0.06% 0.20% 0.07% -0.72%

T = 0.5, = 1/2, = 1/2, s = 100, (0 ) 2 = 0.4, = 2.5, = 0.16, = 0.1, = 0.5 Strike price 70 80 90 100 110 120 Exact IV 40.63 40.30 40.02 39.78 39.58 39.41 Malliavin IV 40.34 40.20 40.07 39.96 39.85 39.76 Relative error -0.71% -0.26% 0.13% 0.45% 0.69% 0.89%

T = 1, = 1/2, = 1/2, s = 100, (0 )1/2 = 0.2, = 2.0, = 0.04, = 0.1, = 0.25 Strike price 70 80 90 100 110 120 Exact IV 20.68 20.36 20.10 19.90 19.75 19.63 Malliavin IV 20.60 20.36 20.15 19.96 19.80 19.64 Relative error -0.40% 0.00% 0.26% 0.32% 0.23% 0.06%

4.2

Double Heston Model

dSt
()

Our method is general enough to be applied to multi-dimensional models. For instance, Gatheral (2008) introduced the following double stochastic volatility model; = = = rSt dt +
() () ()

vt St dW1,t ,

()

()

(73)

dvt dt v

()

(t vt )dt + v ( vt )dt + 2
()

vt (1 dW1,t +
()

()

1 2 dW2,t ), 1

(74) (75)

()

vt (2 dW1,t +

1 2 dW3,t ). 2

14

with 1 , 2 [1, 1]. We derive an approximation formula for the double Heston model and show some numerical examples. Let C Double.SV and IV (T, K) be the call option and the implied volatility under the double Heston model; C Double.SV (T, K) C BS (T, K, IV (T, K)) = = erT E[(ST K)+ ],
()

C Double.SV (T, K).

We have the following approximation formulas for C Double.SV and IV (T, K). Proposition 4.1 C Double.SV (T, K) IV (T, K) where 1 = , 2 = 2 , C1i 1 2 vega with = . (Proof ) By the asymptotic expansion with push-down of the Malliavin weights, we have C Double.SV (T, K) = = = where erT E[(ST K)+ ]
() (0)

= =

C BS (T, K, ) + 1 C11 + 2 C12 + O(2 ), 1 1 C11 + 2 C12 + O(2 ), + 1 vega vega

(76) (77)

= = = = =

erT 1 1 2 1 2 2

(
R t

sex+rT 2 K

)+
i

es

1 3 1 1 1 x 2 x2 2 3x + 3

2 1 1 e 2 x dx (i = 1, 2), 2

s (0) eu vu duds,

0 t

es
(0)

s eu

(0) u

s e( )v dv

vu

vu duds,

(0)

+ (v0 ) s T n(d1 ),

(1 eT ) + (0 ) v T

(1 e T ) (1 eT ) T T

))1/2
,

erT E[(ST K)+ ] + erT E (ST K)+ H ST


(0)

(0,0)

C BS (T, K, ) + erT

() S |=0 t

)]

+ O(2 )

1 1 1 (sex+rT 2 K)+ (x) e 2 dx + O(2 ), 2 R

[ (

(x) = E H ST ,
() S |=0 t

(0)

() S |=0 | t

)
0

]
(0)

vs dW1,s = x .

(78)

is given by () (0,0) S |=0 = St Xt , t

where Xt =

(
0

1 vs
(0)

() 1 vs |=0 dW1,s 2

)
() vs |=0 ds .

Then, , the push-down of Malliavin weight is computed as follows;

[
(x) = = = E (D (Z) DZ )

x x

() S |=0 | t

]
(0)

vs dW1,s = x

E XT |
0

] )

vs dW1,s = x (79)

(0)

1 3 1 1 1 (1 + 2 ) x 2 x2 2 3x + , 3

15

where Z 1 2 Therefore, we have C Double.SV (T, K) = C


BS

= = =

D1 ST 1 1 2

(0)

D1 ST 2 H

(0) t

es
t

s (0) eu vu duds,

1 2 2 2

es

s eu

(0) u

s e( )v dv

vu

vu duds.

(0)

(T, K, ) + e

rT

(
R

sex+rT 2 K

)+
(1 + 2 )

1 1 1 1 3 x 2 x2 2 3x + 3

2 1 1 e 2 x dx 2

+O(2 ). By the similar argument as in 3.2, the implied volatility is expanded as follows: IV (T, K) = + 1 rT e vega

(80)

(
R

sex+rT 2 K

)+
(1 + 2 )

1 3 1 1 1 x 2 x2 2 3x + 3

2 1 1 e 2 x dx 2

+O(2 ). 2

(81)

Next, we show the numerical results on the accuracy of our approximation formula for the call option prices under the double Heston model. Also, in the tables below we dene Malliavin AE and Malliavin IV as follows: Malliavin AE Malliavin IV T = 1, s = 100, 0 = v0
1/2

:= :=

C BS + 1 C11 + 2 C12 , C12 C11 + 2 . + 1 vega vega

= 0.4, = 0.1, 1 = 0.1, 1 = 0.5, = 1.0, = 0.16, 2 = 0.1, 2 = 0.25 Benchmark 33.6198 26.5663 20.6084 15.7241 11.8225 8.7741 Malliavin AE 33.5304 26.4797 20.5439 15.6937 11.8301 8.8198 Relative error 0.27% 0.33% 0.31% 0.19% -0.06% -0.52%

Strike price 70 80 90 100 110 120

Benchmark : Monte Carlo simulation (1,000,000 trials, 500 time steps). T = 2.5, s = 100, 0 = v0
1/2

= 0.4, = 0.1, 1 = 0.1, 1 = 0.5, = 1.0, = 0.16, 2 = 0.1, 2 = 0.25 Benchmark 39.3315 33.5871 28.6053 24.3161 20.6430 17.5111 Malliavin AE 39.4345 33.6975 28.7170 24.4228 20.7397 17.5926 Relative error 0.26% 0.33% 0.39% 0.44% 0.47% 0.47%

Strike price 70 80 90 100 110 120

Benchmark : Monte Carlo simulation (1,000,000 trials, 500 time steps).

4.3

Parameter Identication

It is useful to see if the model parameter can be identied by our implied volatility expansion. For this purpose, we rst compute the implied volatility for a stochastic volatility model with some xed parameters by a numerical method such as Monte Carlo simulation. Next, we estimate the model parameters using our implied volatility expansion. As an example, we use the Heston model; vt (82) dXt = dt + vt dW1,t , 2 dvt = ( vt )dt + vt (dW1,t + 1 2 dW2,t ),

16

where [1, 1]. Here, and are known to be the speed and the level of the mean-reversion parameters, respectively. Also, the parameter = is regarded as a skewness parameter. We compute the implied volatility with S0 = eX0 = 100, v0 = 0.04, = 1.5, = 0.09 and = 0.05. Then, we obtain data set { M C (Ti , Kij )}i,j of the implied volatilities ( Exact IV ). Using our implied volatility expansion and the data set { M C (Ti , Kij )}i,j , we solve the following minimization problem to estimate model parameters as if they were unknown: min
v0 ,,, n m i=1 j=1

| M C (Ti , Kij ) Malliavin IV(Ti , Kij , v0 , , , )|2 ,

(83)

where Malliavin IV(T, K, v0 , , , ) with C1 1 = = = = + C1 , vega (84)

(
+ erT 1 2
(0) (v0

(
R t

(1 eT ) ) T
1

)1/2
, (85)

sex 2 K

)+

es

1 3 1 1 1 x 2 x2 2 3x + 3

2 1 1 e 2 x dx, 2

(86) (87)

s (0) eu vu duds.

As a result, we obtain the optimal parameters v0 = 0.0404, = 1.5022, = 0.0863 and = 0.0336, which are close to the true ones given above. Hence, the model parameters seems well identied by our implied volatility expansion for this case. The tting results are shown in the following tables, where Malliavin IV are computed by these optimal parameters (v0 , , , ) = (0.0404, 1.5022, 0.0863, 0.0336).

Monte Carlo parameter versus optimal parameter Parameters v0 T = 0.4: Strike price 70 80 90 100 110 120 T = 0.5: Strike price 70 80 90 100 110 120 Exact IV 24.71 24.13 23.64 23.18 22.83 22.48 Malliavin IV 24.74 24.17 23.66 23.20 22.79 22.41 Relative error 0.13% 0.14% 0.06% 0.10% -0.20% -0.33% Exact IV 24.31 23.68 23.17 22.67 22.30 21.92 Malliavin IV 24.23 23.66 23.17 22.72 22.32 21.96 Relative error -0.32% -0.05% -0.01% 0.23% 0.09% 0.17% Exact IV 0.0400 1.5000 0.0900 -0.0500 Malliavin IV 0.0404 1.5022 0.0863 -0.0336

4.4

Calibration to Market Data

In this subsection, we show calibration examples using an implied volatility data of Nikkei-225 option as of January 31, 2011, where the strikes are the closest to ATM with maturities 0.20, 0.28, 0.37, 0.62, 0.87,1.37 and 1.87 years. Moreover, the data for ten dierent strikes other than ATM with maturity 0.87 year are also used since this kind of data is most available for the maturity. We compare the tting result of our expansion in the single Heston model with that in the double Heston model. The single Heston model is expressed as vt (88) dXt = dt + vt dW1,t , 2 dvt = ( vt )dt + 1 vt (1 dW1,t + 1 2 dW2,t ). 1

17

The double Heston model is expressed as dXt dvt dt v = = = vt dt + vt dW1,t , 2 1 (t vt )dt + 1 vt (1 dW1,t + 1 2 dW2,t ), v 1 2 ( vt )dt + 2 vt (2 dW1,t + 1 2 dW3,t ), 2 (89)

with W = (W1 , W2 , W3 ) is a 3-dimensional Brownian motion. We dene the skewness parameters 1 and 2 as 1 = 1 1 and 2 = 2 2 . We also dene Malliavin IV 1 and Malliavin IV 2 as Malliavin IV 1 (T, K, v0 , 1 , , 1 ) and Malliavin IV 2 (T, K, v0 , v0 , 1 , 2 , , 1 , 2 ) where C1i (i = 1, 2) are previously given, and := 2 + 1 C11 C12 + 2 , vega vega := 1 + 1 C11 , vega

(
1 2 = = + (v0 )
(0)

(
+
(0) (v0

(1 eT ) T

)1/2
,

(1 eT ) ) + (0 ) v T

(1 e T ) (1 eT ) T T

))1/2
.

To reduce the degree of freedom in Malliavin IV 2, we give the condition v0 = v0 = . For the Nikkei-225 data { Data (Ti , Kij )}i,j , we solve the following minimization problems using our implied volatility expansions. Optimization for Malliavin IV 1 (Single Heston model) Error 1 = min
v0 ,1 ,,1 n m i=1 j=1

| Data (Ti , Kij ) Malliavin IV 1 (Ti , Kij , v0 , 1 , , 1 )|2 .

(90)

Optimization for Malliavin IV 2 (Double Heston model) Error 2 = min


v0 ,1 ,2 ,,1 ,2 n m i=1 j=1

| Data (Ti , Kij ) Malliavin IV 2 (Ti , Kij , v0 , 1 , 2 , , 1 , 2 )|2 .

(91)

Then, we obtain the optimal parameters for the single and double Heston models, as shown in the following two tables. Parameters Malliavin IV 1 v0 0.024 1 2.805 0.0444 1 -0.136 Parameters v0 1 1 2 2 Malliavin IV 2 0.038 9.355 0.038 -0.195 9.070 0.580

The total dierences between the calculated volatilities and the actual data are given as Error 1= 2.956% for Malliavin IV 1 and Error 2=2.663% for Malliavin IV 2. Hence, the tting result of the expansion of the double Heston is better than that of the single Heston. Also, the gures of the result for the double Heston model are shown in Figure 1 and 2.

Applications

This section provides approximation formulas for option prices under the shifted log-normal and jump-diusion models with stochastic volatilities; an expansion of the implied volatility is also given for the jump-diusion model. Also, for simplicity we set the risk-free interest rate r = 0 in this section.

18

Figure 1: Term Structure

Figure 2: Skew

5.1

Shifted Log-normal Model

This subsection derives an approximation formula of the option price in the shifted log-normal model with stochastic volatility: dSt
()

= =

V (t )(St
()

()

()

)dW1,t ; S0 = s > 0,
()

()

dt

()

A0 (t )dt + A1 (t )(dW1,t +

1 2 dW2,t ),

(92)

where is a constant such that s > . At = 0, the option price is given by

C BS ()

=
R

((s )ex 2 (K ))+

x2 1 e 2 dx. 2

(93)

We dene the following deterministic process:

{
t Then, the following proposition is obtained. := exp
0

}
(0) A (u )du 0

Proposition 5.1 An asymptotic expansion formula for the shifted log-normal model (92) is given as follows; C SV () where C1 () with (x) =
0

C BS () + C1 () + O(2 ),

(94)

x2 1 1 e 2 dx, ((s )ex 2 (K ))+ (x) 2 R

(95)

V (t )2 dt,

(0)

= =

1 3 1 1 1 x 2 x2 2 3x + , 3
T

t 0

V (t )t V (t )
0

(0)

(0)

1 (0) (0) s A1 (s )V (s )dsdt.

(96)

19

(Proof) We rst note that dSt


() Ft () St ()

= dFt

()

= V (t )Ft dW1,t ; F0
+ () (FT

()

()

()

= s > 0,
+

(97)

where = . Note also that K) = (K )) . Hence, in stead of the original problem we can consider the option pricing problem with the underlying asset price process F () and strike K . Then, the same argument as in Theorem 3.1 and Corollary ] A.1 can be applied to computation of the push-down of the [ Malliavin weight; (x) := E 1 |

() (ST

T
0

V (t )dW1,t = x , where 1 1 (T ) = = H1 (FT , 1 (T )) () F |=0 . T


(0)

(0)

(98)

Thus, the result is obtained. 2

5.2

Jump-Diusion with Stochastic Volatility Model

This subsection applies the Malliavin calculus to a jump-diusion stochastic volatility (SVJ) model. Let (, F , P ) be a probability space on which we dene a Brownian motion (Wt )t , a Poisson process (Nt )t with intensity and i.i.d random variables (j )jN such that j N (0, 1). We will assume that the -algebras generated by (Wt )t , (Nt )t , (j )j are independent. First, we introduce the perturbed stochastic dierential equation : For [0, 1], dSt
()

= V (t )St dW1,t + St (dJt mdt),


() ()

()

()

()

S0 = s, dt
()

()

= A0 (t )dt + A1 (t )(dW1,t +

1 2 dW2,t ), (99)

0 = , where Jt is dened by
N (t)

()

Jt = with Yj = a + bj and m = E[eYj 1]. The solution of (99) is given by

j=1

(eYj 1),

(100)

{
ST = s exp

V (t )dW1,t

()

{
= s exp

0 T

1 2

T ()

V (t )2 dt

} N (T )
j=1

eYj

V
0

() (t )dW1,t

1 2

N (T )

() (t )2 dt

j=1

}
Yj . (101)

We dene the following notations: Xt


()

:= := :=

log (St /s),

()

BS n n r dn
BS Cn (T, K, n )

( (
:= := := := := :=

1 T

T (0) 0

)1/2
,

(102)

V (t )2 dt
T (0)

) )1/2 ( )1/2

V (t )2 dt + b2 n ,
0

1 T

V (t )2 dt +

(0)

b2 n T

= ( BS )2 + (b2 n/T )

mT + (a + b2 /2)n , [ T( ) ( ) ] 1 s 2 log + r+ n T , K 2 n T T r sN (dn ) e KN (dn n T ),


((1 + m)T )n e(1+m)T n=0

C M (T, K, BS ) C
SV J

n!
()

BS Cn (T, K, n ),

(T, K)

E[(ST K)+ ].

20

Moreover, let us call SV J.IV (T, K) an implied volatility, which satises C SV J (T, K) = C M (T, K, SV J.IV (T, K)). Then, the following proposition is obtained. Proposition 5.2 Under the jump-diusion stochastic volatility (SVJ) model (99), the call option price C SV J (T, K) and its implied volatility SV J.IV (T, K) are expanded up to the -order as follows: C SV J (T, K) SV J.IV (T, K) where
M C1 (T )n eT ( n=0

(103)

= =

M C M (T, K, BS ) + C1 + O(2 ), 1 M BS + C1 + O(2 ), vegaM

(104) (105)

n!

se

xmT 1 2

T
0

V (t

(0) 2

)+
K n (x)

) dt+na

1 1 x2 e 2n dx, 2n (106)

vegaM with

((m + 1)T )n e(m+1)T BS n=0

n!

s T n(dn ),

(
n (x) t = = =

1 3 1 1 1 x 2 x2 2 3x + n 3 n n n
T

)
,

V (t )V (t )t
0

(0)

(0)

1 (0) (0) s A1 (s )V (s )dsdt,

{
0

}
(0) A (u )du 0

exp

(107)

(Proof) See Appendix B.

Conclusion

This paper developed an asymptotic expansion method for generalized Wiener functionals based on the integrationby-parts formula in Malliavin calculus and the push-down of Malliavin weights. As an application, we derived asymptotic expansion formulas for option prices and implied volatilities as well as the density of the underlying asset price in stochastic volatility environment. we also presented numerical examples for expansions up to the rst order (-order). Moreover, we applied the general formula to expansions of option prices for the shifted log-normal model with stochastic volatility. Further, combining some existing results of Malliavin calculus in jump-type models with our method, this paper we presented an approximation formula for the jump-diusion model in stochastic volatility environment. More detailed numerical experiments and higher-order expansions are our next research topics.

Computation of Push-down of Malliavin Weights

This section derives the push-down of Malliavin weights for the rst and second order approximation terms of the asymptotic expansion of option prices. We consider the following stochastic volatility model; dSt dt
()

= V (t )St dW1,t , = A0 (t )dt + A1 (t )(dW1,t +


(0) () ()

()

()

()

1 2 dW2,t ),

S0 = S0

()

= s,

where V, A0 , A1 Cb (R), [1, 1] and [0, 1]. Also we use the following notations;

Bt t t t
()

= = = =

W1,t +

1 2 W2,t ,

(108) ,

exp
0

(0) A (u )du (1)

(0)

+ t

+ 2 t

(2)

(1)

() |=0 = t

+ O(3 ),

t 0

1 t s A1 (0s )dBs ,

21

t vt

(2)

= = = = =

1 2 () |=0 = 2 2 t V (t ), V (0t )t ,
(1) (0)

A1 (0s )
0

1 s u A1 (0u )dBu dBs ,

(0) (1) (2)

vt

vt

2V (0t )t

(2)

+ V (0t )(t )2 . 1 2

(1)

XT XT

(0)

V (0t )dW1,t
0

V (0t )2 dt,
0

(1)

= =

() X |=0 T
T

vt dW1,t
0

(1)

V (0t )vt dt,

(1)

(2) XT

= =

2 () X |=0 2 T
T

(vt )2 dt
(1) 0

vt dW1,t

(2)

V (0t )vt dt,

(2)

(
=

1 T

)1/2
.

(0) (t )2 dt

The closed-form approximation of the density and the call option price at t = 0 with strike K and maturity T are given by pSV (y) C
SV

= =

pBS (y) + E[1 |ST = x]pBS (y) + 2 E[2 |ST = x]pBS (y) + O(2 ), C
BS

(0)

(0)

(T, K)

(T, K, ) (y K)+ E[1 |ST = x]pBS (y)dy


(0)

+
R

+ =

2 R

(y K)+ E[2 |ST = x]pBS (y)dy + O(3 )

(0)

C BS (T, K, )

(
R

+ +2
R

se

1 x 2

T
0

V (0s )2 ds

)+
K K E[1 |


E[2 |
0

V (0t )dW1,t = x]n( : x)dx


T

se

1 x 2

T
0

V (0s )2 ds

)+

V (0t )dW1,t = x]n( : x)dx + O(3 ).


0

where 1 and 2 is the Malliavin weights: 1 2 = = H1 ST , () S |=0 , T ( ( ) ( )2 ) 2 1 1 () (0) () (0) H1 ST , 2 ST |=0 + H2 ST , S |=0 . 2 2 T
(0) (0) (0) (0) (0) (0) (0) (0) (0)

(109) (110)
(0)

Note rst that for t [0, T ], Dt,2 XT = Dt,2 ST = 0, Dt,1 XT = V (t ), Dt,1 ST = ST Dt,1 XT = ST V (t ) T (0) (0) (0) (0) and DST 2 = D1 ST 2 = (ST )2 where = 0 V (t )2 dt. Let Z be the function on L2 [0, T ] dened by H H Z(s) =
Ds,1 XT
(0) (0) D1 XT 2 H

V (s

(0)

. Recall also that D1 and D1 are the Malliavin derivative and its adjoint operator

(Skorohod integral) for the Brownian motion W1 . Then, the second-order Malliavin weight is computed as follows;

(
H1
(0) ST ,

2 () S |=0 2 T

)
=
D1

( (
= = =
D1

D1 ST ST { 2 XT |=0 + ( XT |=0 )2 }

(0)

(0)

()

()

) )

D1 ST 2 H D1 ST ST
(0) (0) 2 () XT |=0 2 (0) D1 ST 2 H (2)

(0)

(
D1

D1 ST ST ( XT |=0 )2

(0)

(0)

()

(
(2)

H1 XT , XT
T

(0)

+ H1 XT , (XT )2

(0)

(1)

D1 ST 2 H

(0)

XT

Z(s)dW1,s
0

Ds,1 XT Z(s)ds
0

(2)

(1) +(XT )2

Z(s)dW1,s

Ds,1 (XT )2 Z(s)ds,

(1)

(111)

22

(
H2
D1 (0) ST ,

(
=

() S |=0 T
(0) D1

)2 )
= H1 D1 ST (ST
(0)

(
(0) ST , H1

(
(0) ST ,

D1 ST

(
= = = = = =
D1

D1 ST 2 H D1 ST D1 ST
(0)

(0)

(
D1

D1 ST 2 H
(0)

(0)

D1

() (0) X | )2 T =0 (0) D1 ST 2 H

))

() S |=0 T

)2 ))

(0) (1) (ST XT )2 Z

))

{
ST (XT )2
(0) (1)


(0) (1) ST (XT )2 0

Z(s)dW1,s

T (0) (1)

})
Ds,1 (ST (XT )2 )Z(s)ds

(
D1

D1 ST 2 H
(0) D1 ST (0) D1 ST 2 H (0) D1 ST (0) D1 ST 2 H (0) D1 XT (0) D1 XT 2 H (0) (1)

(0)

{ {
ST (XT )2
(0) (1)

Z(s)dW1,s

0 T (0) (1) (D1,s ST )(XT )2 Z(s)ds 0

0 T

})
(0) (1) ST (Ds,1 (XT )2 )Z(s)ds

(
D1


(1)

0 T

(1) 0

(0) 0

T (1)

})
(Ds,1 (XT )2 )Z(s)ds

Z(s)dW1,s ST (XT )2
0

(0)

(0) V (s )Z(s)ds ST

(
D1

{
(XT )2

)
0

})
(1)

Z(s)dW1,s
0 (0) (1)

(Ds,1 (XT )2 )Z(s)ds (XT )2 (112)

(1)

H2 XT , (XT )2 H1 XT , (XT )2 .

Then, we have 2 = = 1 1 (0) (0) (2) (1) H1 XT , XT + H2 XT , (XT )2 2 2 1 H1 2 1 H2 2

)
(113)

XT ,

(0)

vt dW1,t

(2)

(
(0) XT ,

(
0

T (1) vt dW1,t

)2 )

1 H1 2

T (1)

)
(vt )2 dt

XT ,

(0)

1 H1 2

(
XT ,
(0)

T (2)

)
V (0t )vt dt

T (1) vt dW1,t

(
1 + H2 2
(0) XT ,

(
0

)2 )
V
(1) (0t )vt dt

H2

(0) XT ,

(1) (0t )vt dt

. (114)

In order to compute the expansion coecient including the push-down of the Malliavin weights, we give the following formula (115) which is modied version of Malliavin (1997) and Malliavin-Thalmaier (2006). Proposition A.1 For D (W : R) and Z = E[D (Z) DZ |D (h) = x] =
h h2 H

with h H, (115)

x E[|D (h) = x] E[|D (h) = x]. x h2 H

(Proof ) We follow Proposition (8.2) in p.82 of Malliavin (1997). Let g be a non-degenerate Wiener smooth map, g D (W : Rn ) and z = (z 1 , z 2 , , z n ) be a vector eld on n R . Suppose also that Z = (Z 1 , Z 2 , , Z d ) be a lift of z to the Wiener space ( a covering vector eld of z ); Z k (t) =

1s,ln

s,l (Dt,k g s )z l ,

where s,l is the (s, l)-element of the inverse matrix of the Malliavin covariance matrix of g. Then, Proposition (8.2) in p.82 of Malliavin (1997) says that for D (W : R) and k(x) = E g=x [] where E g=x denotes the conditional expectation under g = x, E g=x [D (Z) DZ ] = k(x)E g=x [D (Z)] z k(x), where z k(x) := z, (dk(x)/dx). In our case, set g = D (h) and then, g D (W : R) (i.e. n = 1). Also, let z 1 and hence, h Dg = . Z= Dg2 h2 H H Finally, note that E g=x [D (Z)] =
x h2 H

in our case, and thus the result (115) is obtained. 2

Using (115) above, we give the rst order approximation formula of the call option price under the stochastic volatility.

23

Corollary A.1 The rst order approximation of the call price is given by C SV (T, K) = C BS (T, K, ) +

(
R

sex 2 K

)+

(x)

2 1 1 e 2 x dx, 2

(116) where (x) with t where A (u ) = 0


(0) dA0 (x) |x=(0) . dx u

(
= =
0

1 3 1 1 1 x 2 x2 2 3x + , 3
(0) V (s )2 ds,

(117)

V (t )t V (t )
(0) (0)

(0) (0) 1 s A1 (s )V (s )dsdt,

(118)

{
= exp
0

}
(0) A (u )du 0

(119)

(Proof ) By (115) and the formula 1 ;

]
h1v dW1,v = x =

(
0

)(
h2u h1u duh3t h1t dt
(0)

E
0 0

h2u dW1,u h3t dW1,t |

x2 1 2

)
, (120)

for hi L2 [0, T ], i = 1, 2, 3, The push-down of the Malliavin weight H1 (XT ,

T
0

T (1) vt dW1,t

)
|

T 0

vt dW1,t ) is obtained as follows.

(1)

E H1

(0) XT , 0

V (0s )dW1,s = x

[
= = =

0 T 0

T (1) vt dW1,t |

]
V (0s )dW1,s = x

E (D (Z) DZ )

( T

x x
3

[
E ,

V (0t )

1 t s A1 (0s )dBs dW1,t |

]
V (0s )dW1,s = x (121)

3x x 2 3

) t
0

where = 0 V (0t )t V (0t ) By (115) and the formula; E

1 s A1 (0s )V (0s )dsdt.

[
0

h2t dW1,t |

]
h1v dW1,v = x =

(
0

)
h2t h1t dt

0 (0)

the push-down of the Malliavin weight H1 XT ,

T
0 T 0

V (0t )vt dt

(1)

) ]

x ,

(122)

is obtained as

[
E = = = =
(0) H1 (XT ,

V
(1) (0t )vt dt)|

V (0s )dW1,s = x

[ [


t T t 0 T

E (D (Z) DZ ) E (D (Z) DZ )
T 0 T

V (0t )V (0t )
1 s A1 (0s )

1 t s A1 (0s )dBs dt|

]
V (0s )dW1,s = x

V (0t )V (0t )t dtdBs | V (0t )V (0t )t dtdBs |

]
V (0s )dW1,s = x

x x
2

[
E
0

1 s A1 (0s )

0 T

V (0s )dW1,s = x (123)

1 x 2

)
.

Therefore, we obtain E[1 |

(
V (0t )dW1,t = x] =

1 3 1 1 1 x 2 x2 2 3x + . 3

(124)

2
1 For

the derivation and more general results, see Section 3 in Takahashi-Takehara-Toda (2009).

24

A.1

Second Order Approximation

This subsection lists up the results necessary for computing push-down of the Malliavin weights of 2 -order terms of (114). The detail of the computation will be given upon request.

A.1.1

H1 (XT ,

(0)

T
0

vt dW1,t )
(0) E[H1 (XT , T (2) vt dW1,t )| 0

(2)

)
0

V (0s )dW1,s = x]

(
= where b11 b12 b13
(0)

x4 6x2 3 3 + 2 4

(2b11 + b12 ) +

1 x2 2

)
b13 , (125)

= =
2 0 T 0 T 0

V (0t )V (0t )

A (0s )s V 1

(0s )
0

1 u A1 (0u )V (0u )dudsdt,

(
0

)2

1 V (0s )s A1 (0s )ds

2 V (0t )t V (0t )dt,

V (0t )V (0t )
0

)2
1 t t A1 (0s )

dsdt.

(126)

A.1.2

H1 (XT ,

T
0

(vt )2 dt)
T

(1)

(vt )2 dt)|
(1) 0

(
V (0s )dW1,s = x] =

E[H1 (XT ,
0

(0)

3x x3 2 3

)
b21 +

x b22 ,

(127)

where b21 b22


(0)

= =
0

(V (0t )t )2 (

1 s A1 (0s )V (0s )ds)2 dt,

0 T

2 0

(V (0t )t )

1 (s A1 (0s ))2 dsdt.

(128)

A.1.3

H1 (XT ,

T
0

V (0t )vt dt)


T

(2)

(
V (0s )dW1,s = x] = 2

E[H1 (XT ,
0

(0)

V (0t )vt dt)|


0

(2)

x3 3x 2 3

)
b31 +

x3 3x 2 3

)
b32 +

x b33 , (129)

where b31 b32 b33


(0)

= = =
0

V (0t )V (0t )

A (0s )s V (0s ) 1

1 u A1 (0u )V (0u )dudsdt,

0 T

0 t 0

2 (V (0t )t V (0t ))( 0 T 2 (V (0t )t V (0t ))

1 s A1 (0s )V (0s )ds)2 dt,

1 (s A1 (0s ))2 dsdt.

(130)

A.1.4

H2 (XT , (

T
0

vt dW1,t )2 )
T (1) vt dW1,t )2 )| 0

(1)

)
0

(0) E[H2 (XT , (

V (0s )dW1,s = x].

(
=

x6 15x4 45x2 15 + 3 6 5 4

b41 +

x4 6x2 3 3 + 4

)
(2b42 + 2b43 + b44 ) +

x2 1 2 2

)
b45 , (131)

25

where b41 b42 b43 b44 b45


(0)

(
=

T 0

V (0t )t V (0t ) V (0t )t V (0t )


T


0 0 t

)2
1 s A1 (0s )V

(0s )dsdt

=
0

V (0s )s V (0s )
t

1 (u A1 (0u ))2 dudsdt, s

= = =

2
T

V (0t )t V (0t )

0 t

V (0s )A1 (0s )

1 V (0u )u A1 (0u )dudsdt,

(
(V (0t )t )2

)2
dt, (132)

1 s A1 (0s )V (0s )ds

0 T

2 0

0 t 1 (s A1 (0s ))2 dsdt.

(V (0t )t )
(1)

A.1.5

H2 (XT ,

T
0

vt dWt

T
0

V (0t )vt dt)


T (1) vt dW1,t 0

(1)

)
0

V
(1) (0t )vt dt)|

(0) E[H2 (XT ,

V (0t )dW1,t = x]
0

(
= where b51 b52 b53 q1t q2t q3t q5t = =

x5 10x3 15x + 3 5 4

b51 +

x3 3x 2 3

(b52 + b53 ),

(133)

(
2

q3t q1t

) (
q2s q1s dsdt

(
q4u

T u

)
q5s ds q1u du

( (
= = = = =
0 2

0 T

0 T

0 t

(
s T

)
t

q3t q1t

q4s q2s

q5u du dsdt

(
t

)
0

)
q2s q1s dsdt ,

q4t q3t

q5u du

V (0t ),
1 q4t = t A1 (0t ),

V (0t )t , V (0t )V (0t )t . (134)

A.1.6

H2 (XT , (
T

(0)

T
0

V (0t )vt dt)2 )


T

(1)

[(
V (0s )dW1,s = x] = 2

E[H2 (XT , (
0

(0)

V (0t )vt dt)2 )|


0

(1)

x4 6x2 3 3 + 4

)
b61 +

x2 1 2

]
b62 , (135)

where b61 b62 q1t q2t q3t = =


0

2 0 T

(
q3t

) ( t (
q2u q1u du
t u

q3s ds q2u q1u du dt

= = =

0 t 2 q2u

q3t V (0t ),

q3s ds dudt,

1 t A1 (0t ),

V (0t )V (0t )t .

Proof of Proposition 5.2


M C1

M Let C1 denote the coecient of -order in the asymptotic expansion of C SV J around = 0:

() E (ST K)+ |=0

26

[
= = where x E x
n=0

{( [

(0) ST

K
(0)

)+ }

{(

E x

ST K

)+ }

() S |=0 T

] ]

() S |=0 1{N (T )=n} , T

(136)

{(

ST K

(0)

)+ }
=

(x K)+ |x=S (0) . x T

Next, by the integration by parts formula including the jump amplitudes (e.g. Bavouzet and Messaoud (2006)), the right hand side of the last equality in the above equation is expressed as
M C1

n=0

[(
E

(0) ST

)+

]
H1,n 1{N (T )=n} , (137)

where H1,n is the Malliavin weight on {N (T ) = n}, which will be given below: Let D0 be the Malliavin derivative with respect to Brownian motion for W1 and Di be the Malliavin derivative with respect to the jump amplitudes i , i = 1, , n. Then, on {N (T ) = n}, D0,t ST Di ST XT =
0 (0) (0) (0)

= =
T

ST D0t XT = ST V (t ), t [0, T ], ST Di XT = ST b, i = 1, , n,
(0) (0) (0) (0)

(0)

(0)

(0)

(0)

(0)

(138) Yj ,
(0) (0)

V (t )dW1,t

1 2

V (t )2 dt +

(0)

n j=1

where we use Di XT = b (e.g. see p.280 in Bavouzet-Messaoud (2006)) as well as D0,t XT = V (t ). (0) Then, the Malliavin covariance of ST , S (0) is given by
T

S (0)
T

(
T

=
0 (0)

D0,t ST (ST )2
0

(0)

)2
dt +
(0)

n ( i=1

Di ST
(0)

(0)

)2

= = where

V (t )2 dt + (ST )2 b2 n (139)

(ST )2 n ,

(0)

n :=
0

V (t )2 dt + (ST )2 b2 n,

(0)

(0)

and

n ( i=1

Di ST

(0)

)2

= (ST )2 (b2 n)

(0)

is the contribution by the jump part. (e.g. See p.282 in Bavouzet-Messaoud (2006).) Using the Skorohod integral for the Brownian motion and the jump amplitude, on {N (T ) = n} we can compute the Malliavin weight H1,n as follows; H1,n =
n i=0

(
Di

(0) () S | DS T =0 i T

)
=

n i=0

(
Di

ST XT ST Di XT (ST )2 n
(0)

(0)

(1)

(0)

(0)

S (0)
T

n ) ( 1 (0) (1) Di XT Di XT n

= +

1 n n

i=0

(0) (1) XT D0 (D0 XT )

0 (0)

)
(0) (1) (D0,t XT )(D0,t XT )dt

n ( 1 i=1

(XT Di (Di XT ) (Di XT )(Di XT ) , (1) (1) (0)

(140)

where each Di (i = 0, 1, , n) is the adjoint operator of Di (i = 0, 1, , n) and

XT

(1)

= =

() X |=0 T
T 0

V (0t )

1 t s A1 (0s )dBs dW1,t

V (0t )V (0t )

1 t s A1 (0s )dBs dt,

27

with Bt
(0)

W1,t +

1 2 W2,t .

Note that D0 (D0 XT ) is expressed by the usual It integral, o (0) D0 (D0 XT )

=
0

V (t )dW1,t .
(0)

(0)

Recall that i is Gaussian random variable N (0, 1). Hence, from p.280 in Bavouzet-Messaoud (2006), Di (Di XT ) for i = 1, , n is given by Di (Di XT ) = Di (b) = b (0)

log p(i ) = bi , i

(141)

where
x2 1 p(x) = e 2 . 2

Because XT t [0, T ],

(1)

is a stochastic integral driven by the Brownian motion which is independent of i , we have for

D0,t XT

(1)

V (t )
(0)

1 (0) 1 t s A1 (s )dBs + t A1 (t ) (0)

(0) V (u )u dW1,u

(0) 1 t A1 (t )

(0) (u )V

(0) (u )u du,

and for i = 1, , n, Di XT = 0. Hence, (140) is expressed as


(1)

(
H1,n Then,
M C1

1 n

(
0

(1) XT

(0) (t )dW1,t

n i=1

)
bi

T (1) D0,t XT V

)
(0) (t )dt

(142)

n=0

[(
E

ST K

(0)

)+
H1,n 1{N (T )=n}

] )+
H1,n

(T )n eT n=0

[(
E

] T
0

n!

ST K

(0)

(T )n eT (
n=0

n!

se

xmT 1 2

V (t

(0) 2

)+
K

) dt+na

1 1 x2 n (x) e 2n dx, 2n (143)

where n (x) is the push down of the Malliavin weight H1,n ;

[
n (x) = E H1,n |

(0) (t )dW1,t

n i=1

]
bi = x .

We easily evaluate n (x) as

(
n (x) Thus, we obtain (104): C SV J (T, K) Next, let us regard C M as a function of BS :
BS

1 1 1 1 3 x 2 x2 2 3x + n 3 n n n

)
. (144)

M C M (T, K, BS ) + C1 + O(2 ).

(145)

C (T, K,
M

BS

)=

((m + 1)T )n e(m+1)T n=0

n!

BS Cn

(( (

BS 2

+ (b n/T )

)1/2 )
.

28

Also, the vega of C M with respect to BS , vegaM is given by vegaM = = Suppose that SV J.IV satises C SV J (T, K) = C M (T, K, SV J.IV (T, K)), and that SV J.IV (T, K) is expanded around = 0: SV J.IV (T, K) = BS + 1 + O(2 ). Then, we have C M ( SV J.IV ) = C M (T, K, BS ) + vegaM 1 + O(2 ). (147) C M (T, K, BS ) BS n!

((m + 1)T )n e(m+1)T BS n=0

s T n(dn ).

(146)

Finally, comparing (104) with (147), we obtain 1 which leads to the result (105). 2 =
M C1 , vegaM

(148)

References
[1] Antonelli, F. and Scarlatti, S.: Pricing Options under Stochastic Volatility: a Power Series Approach, Finance Stoch. (2009), no. 13, 269-303. [2] Bavouzet, M-P. and Messaoud, M.: Computation of Greeks Using Malliavins Calculus in Jump Type Market Models, Electronic Journal of Probability. 11 (2006), no. 10, 276-300. [3] Ben Arous, G. and Laurence, P.: Second Order Expansion for Implied Volatility in Two Factor Local Stochastic Volatility Models and Applications to the Dynamic -SABR Model, Preprint (2009) [4] Fouque, J-P., Papanicolaou G. and Sircar, K.R.: Derivatives in Financial Markets with Stochastic Volatility, Cambridge University Press, Cambridge, (2000). [5] Fourni, E., Lasry, J.-M., Lebuchoux, J., Lions, P.-L. and Touzi, N.: Applications of Malliavin Calculus to e Monte-Carlo Methods in Finance, Finance Stoch. 3 (1999), no. 4, 391-412. e [6] Fourni, E., Lebuchoux, J. and Touzi, N.: Small Noise Expansion and Importance Sampling, Asymptotic Anal. (1997), 361-376. [7] Gatheral, J.: Further Developments in Volatility Derivatives Modeling, Global Derivatives Trading & Risk Management, Paris, (2008) [8] Gatheral, J., Hsu, E.P. and Laurence, P., Ouyang, C. and Wang, T-H.: Asymptotics of Implied Volatility in Local Volatility Models: forthcoming in Mathematical Finance (2009) [9] Hagan, P.S., Kumar, D., Lesniewskie, A.S. and Woodward, D.E.: Managing Smile Risk, Wilmott magazine, (2002). [10] Ikeda, N. and Watanabe, S.: Stochastic Dierential Equations and Diusion Processes, Second Edition, NorthHolland/Kodansha. (1989) [11] Kunitomo, N. and Takahashi, A.: On Validity of the Asymptotic Expansion Approach in Contingent Claim Analysis, Annals of Applied Probability, Vol.13, No.3, (2003). e [12] Labordre, P.H.: Analysis, Geometry and Modeling in Finance : Advanced Methods in Options Pricing, Chapman and Hall, (2008). [13] Lewis, A.L.: Option Valuation under Stochastic Volatility with Mathematica Code, Newport Beach, CA: Finance Press. (2000) [14] Malliavin, P. and Thalmaier, A.: Stochastic Calculus of Variations in Mathematical Finance, Springer. (2006) [15] Malliavin, P.: Stochastic Analysis, Springer. (1997) [16] Nualart, D.: The Malliavin Calculus and Related Topics, Springer. (2006) [17] Sakamoto, Y. and Yoshida, N.: Asymptotic Expansions of Mixture Type Statistics Based on Generalized Wiener Functionals, Cooperative Research Report 58, (1994) The Insitute of Statistical Mathematics, Tokyo, Japan.

29

[18] Shiraya, K., Takahashi, A. and Yamada, T.: Pricing Discrete Barrier Options under Stochastic Volatility, forthcoming in Asia-Pacic Financial Markets (2010) [19] Siopacha, M. and Teichmann, J.: Weak and Strong Taylor Methods for Numerical Solutions of Stochastic Dierential Equations, Quantitative Finance. 11(4), (2010) 517-528. [20] Takahashi, A.: Essays on the Valuation Problems of Contingent Claims, Ph.D. Dissertation, Haas School of Business. (1995) University of California, Berkeley. [21] Takahashi, A.: An Asymptotic Expansion Approach to Pricing Financial Contingent Claims, Asia-Pacic Financial Markets 6 (1999), 115-151 [22] Takahashi, A., Takehara, K. and Toda, M.: Computation in an Asymptotic Expansion Method, Preprint, CARF Working Paper Series CARF-F-149. (2009) [23] Takahashi, A. and Yoshida, N.: An Asymptotic Expansion Scheme for Optimal Investment Problems, Statistical Inference for Stochastic Processes 7 (2004), 153-188. [24] Takahashi, A. and Yoshida, N.: Monte Carlo Simulation with Asymptotic Method, The Journal of Japan Statistical Society, Vol. 35-2, (2005), 171-203. [25] Watanabe, S.: Malliavins Calculus in Terms of Generalized Wiener Functionals, Lecture Notes in Control and Inform. Sci., 49, Springer. (1983) [26] Watanabe, S.: Lectures on Stochastic Dierential Equations and Malliavin Calculus, Springer. (1984) [27] Watanabe, S.: Analysis of Wiener Functionals (Malliavin Calculus) and its Applications to Heat Kernels, Ann. Probab. 15 (1987), no. 1, 1-39. [28] Yoshida, N.: Asymptotic Expansion for Small Diusions via the Theory of Malliavin-Watanabe, Probability Theory and Related Fields, 92, (1992a), 275-311. [29] Yoshida, N.: Asymptotic Expansion for Statistics Related to Small Diusions, J.Japan Statist. Soc. 22 (1992b), no.2, 139-159.

30

Вам также может понравиться