You are on page 1of 15

The smile in stochastic volatility models

Lorenzo Bergomi and Julien Guyon Socit Gnrale, Global Markets Quantitative Research July 2011, revised April 2012
Abstract We consider multi-factor stochastic volatility models and derive the volatility smile at order two in the volatility-ofvolatility. At this order, the smile is quadratic in log-moneyness and depends on three eective quantities within which the dynamics of the spot and forward variances of any particular model is condensed. We supply explicit expressions of these quantities for a family of Heston-like models as well as a 2-factor version of the Bergomi model. For this model, comparison with the exact smiles shows good agreement for volatility-of-volatility levels that are typical of equity underlyings. Finally we derive short term asymptotics and highlight the structural dependence of the level of ATM skew and curvature on the ATM volatility, and we link the decay of ATM skew and curvature for long maturities to the time decay of spot/variance and variance/variance covariance functions.

Consider the following dynamics for a multi-factor diusive stochastic volatility model: q 1 t t 1 dxt = dt + x0 = x t dZt ; 2 t u u u (t; u; t ) dZt ; d t = 0 =

(1)

xt stands for the logarithm of the asset value St ; t ( u ; t u) stands for the instantaneous forward variance curve t u at time t. t is the instantaneous forward variance for date u observed at time t. The volatility = ( 1 ; : : : ; d ) of forward instantaneous variances takes values in Rd , and Z = Z 1 ; : : : ; Z d is a d-dimensional Brownian motion. The rst component of the Brownian motion, Z 1 , drives the spot dynamics. Spot/variance covariance is thus modeled through 1 . We assume that the asset pays no dividends and for the sake of simplicity take zero rates and repos. Forward variances u are driftless. Their initial value can be calibrated on market prices of variance swap (VS) d contracts: u = du ^ 2 u , where ^ u is the implied VS volatility for maturity u. Alternatively, the u can be chosen so as 0 0 u to recover market prices of other instruments, such as at-the-money vanilla options for all maturities. Second-generation stochastic volatility models, such as the Bergomi model [3], in which the dynamics of the u is directly modelled, are naturally expressed in the form of equation (1). By contrast, rst-generation stochastic volatility models, such as the Heston [8] or the double-lognormal model [6], are built on an autonomous dynamics for the instantaneous variance Vt = t . t They impose structural constraints on the shape of the initial variance curve u , which is unfortunate as the u are basic 0 underliers whose initial values should not be restricted in the model. From a modeling perspective, forward variances as opposed to Vt are natural objects as they are driftless by construction and are allowed to take any positive initial value. Still, rst-generation models can be cast as forward variance models and our analysis applies to them as well see the example of the Heston model further below. For general stochastic volatility models, the pricing equation for European payos is not analytically solvable. While many dierent approximations have been proposed for specic rst-generation models, the general case of forward variance models has not been considered in the literature. In this article we derive an approximation of the smile produced by the generic model (1) at second order in the volatility-of-volatility. To this end, we introduce a scaling factor " for the volatilities of instantaneous forward variances: ! " . Expanding at order 2 in " is thus exactly equivalent to expanding at order 2 in . Our derivation relies on the property that in model (1), the volatility of St incorporates no local volatility component, and that does not depend on St : mixed local/stochastic volatility models lie outside our scope, and the stochastic volatility must be an autonomous process. While it is obvious that the smile produced by stochastic volatility models is generated by the covariance of forward variances with themselves and St , our aim in this article is to pinpoint precisely which functionals of these covariances determine the vanilla smile. It is important to ensure, while varying "; that implied volatilities of some specic payos are unchanged, so that the overall volatility level is not altered in the model. In our framework, VS volatilities are unchanged as " is varied. This stands in contrast with other types of approximations whose accuracy is compromised by the fact that the overall volatility level in the model changes as " is increased. 1

Expansion of the price at second order in volatility-of-volatility

Consider a vanilla option whose payo at maturity T is g(xT ). Its value at time t is P (t; x; ): P is a function of x and a functional of the variance curve . P solves the equation: (@t + Ht ) P = 0 with terminal condition P (t = T; x; ) = g(x), where Ht
0 Ht 0 = Ht + Wt1 + Wt2 t

(2)

= = =

Wt1 Wt2 with (t; u; ) (t; u; u ; )


0

2 Z T
t

2 @x

@x
u

2 du (t; u; ) @x T

1 2

du

du0 (t; u; u0 ; ) @ 2u

u0

= =

h dt E d ud t dt

u dSt St d t

u t

i 0

1 (t; u;

) ) i (t; u0 ; )

d X i=1

i (t; u;

(t; u; ) is the instantaneous spot/variance covariance function: it quanties how a relative variation of S is correlated at time t with a variation of the instantaneous forward variance u for a date u > t. Likewise, (t; u; u0 ; ) quanties the 0 instantaneous covariance at time t of two forward variances u and u . We assume that the volatility of the asset is an 0 autonomous process: and do not depend on x. Note that in the expressions of Ht ; Wt1 ; Wt2 , derivatives with respect to x are standard derivatives while derivatives with respect to u are functional derivatives. We now introduce a dimensionless scaling factor " for volatilities of variances , which we will set to 1 in numerical applications. The expression of Ht becomes:
0 Ht = Ht + "Wt1 + "2 Wt2

We now derive an expansion of P at order two in ": P = P0 + "P1 + "2 P2 + (3)

It is possible to insert the expansion (3) of P into the pricing PDE (2), and derive sequentially PDEs for the Pi by equating to zero the contribution of each order in ". Each Pi is solution to a Black-Scholes PDE with a source term, and P0 , P1 and P2 can be explicitly computed using Feynman-Kac representation and martingale arguments. We present here a more economical derivation based on the time-dependent perturbation technique well-known in quantum mechanics. Integrating ODE (2), P is given by P (t; ) = UtT g, where operator Ust , with 0 s t T , is dened by: Ust = lim (1 + tHt0 ) (1 + tHt1 )
n!1

1 + tHtn r Z s
t

t=

t n

ti = s + i t

(4)

Ust satises the semi-group property: Urt = Urs Ust for 0

t H d

T , hence the following expression for Ust : :

Ust = : exp

where : : denotes time ordering. 0 0 Let us now write Ht = Ht + Ht , where Ht is the unperturbed operator corresponding to the Black-Scholes model with deterministic time-dependent volatility, and Ht is a perturbation. We now compute the expansion of Ust at order 2 in H. From (4), we get: Z t Z t Z t 0 0 0 0 Ust = Ust + d Us H U t + d 1 d 2 Us 1 H 1 U 01 2 H 2 U 02 t + (5)
s s
1

In our case Ht = "Wt1 +"2 Wt2 . Using equation (5) and keeping terms up to order two in " yields P = P0 +"P1 +"2 P2 + with: P0 P1 P2 = = =
0 UtT g Z T d Ut0 W 1 U 0T g

(6) (7) Z
T

d Ut0 W 2 U 0T g +

d
1

Ut0 1 W 11 U 01 2 W 12 U 02 T g

(8)

0 The expression of the free propagator Ust is: 0 Ust = e 2


1

Rt
s

0d

2 (@x

@x )

(9)

0 where we have removed the time-ordering symbol as operators Ht for dierent values of t commute. P0 in (6) is the RT standard Black-Scholes price: P0 (t; x; ) PBS (t; S; b), with S = ex , b2 = T 1 t t 0 d . The computation of P1 and P2 is straightforward once one notices that:

0 @x and Ust commute as is clear from expression (9) this is equivalent to saying that, in the Black-Scholes model p with time-dependent volatility, @x P0 is a martingale for all integer p.

1 0 0 2 @ u UtT g = 2 @x @x UtT g if u 2 [t; T ]: this is equivalent to the gamma/vega relationship for European options 2 0 = in the Black-Scholes model: dP = S 2 d P T . If u 2 [t; T ] @ u UtT g = 0, since g is a function of x only. This is easily d dS 2 0 checked on expression (9) for U .

Using these two rules and the semigroup property, we get P1 = = = = Z


T

d Z Z
T

Ut0

1 2 1 2

T 2 du ( ; u; ) @x u U 0T g

d Ut0
T

x Ct ( ) 2 @x @x 2

T 2 du ( ; u; ) @x @x

@x U 0T g

2 du ( ; u; ) @x @x

0 @x UtT g

@x P0

and for the two pieces making up the order two contribution: Z = = = = =
T

d Ut0 W 2 U 0T g Z Z Z Z
T

1 2

d Ut0
T

1 4 1 8 1 8

Z Z Z

du
T

Z Z Z

du0 ( ; u; u0 ; ) @ 2u
T

u0

U 0T g @x U 0T g
2

d Ut0
T

du
T

du0 ( ; u; u0 ; ) @
T

2 @x

d
T

Ut0 Z
T

du Z
2 T

2 du0 ( ; u; u0 ; ) @x

@x
2

U 0T g

du @x

2 du0 ( ; u; u0 ; ) @x

@x

0 UtT g

Ct ( ) 2 @x 8

P0

and Z
T

d
T

= = = = =

Z Z
1

d
1

Ut0 1 W 11 U 01 2 W 12 U 02 T g Z
T

d Z Z
T

d
1

0 2 Ut

du ( 1 ; u; )
1 u

2 @x T

0
1 2

1 2 1 2

d
T

Ut0 1

Z Z
t

T
1

2 du ( 1 ; u; ) @x T 2 @x

Z
2

T 2 du0 ( 2 ; u0 ; ) @x
u0

U 02 T g @x U 02 T g

d 2 U 01
1

T 2 du0 ( 2 ; u0 ; ) @x @x
2

0 1 Ut

du ( 1 ; u; )
1

C 1( )

2 @x @x

@x U 01 T g

1 2 2 @ @ 2 x x 1 2 2 @ @ 4 x x 1 2 2 + @x @x 2

@x @x

Z
2

d
T

Ut0 1
1

d
T

@x

x Ct ( )2 2 2 @x @x 8

T
1

du ( 1 ; u; ) @ u C x1 ( ) U 01 T g
1

0 du ( 1 ; u; ) C x1 ( ) UtT g T 0 du ( 1 ; u; ) @ u C x1 ( ) UtT g
1

d
2

@x

P0 +

Ct ( ) 2 2 @x @x 2

@x P0

where we have introduced the following three dimensionless quantities:


x the (doubly) integrated spot/variance covariance function Ct ( ):

x Ct ( ) =

ds

du (s; u; ) =

ds

du

u dSs Ss d s

ds

the (triply) integrated variance/variance covariance function Ct ( ): Ct ( ) = Z


T

ds

du

du0 (s; u; u0 ; ) =

ds

du

du0

h E d

u0 u sd s

ds

x the double time-integral of the instantaneous spot/variance covariance function times the sensitivity of Ct ( ) with respect to instantaneous forward variances: Z T Z T x Ct ( ) = ds du (s; u; ) @ u Cs ( ) t s

The nal expression of P at time t = 0 and at order two in " is: " Cx ( ) 2 P = 1 + " 0 0 @x @x @x 2 + "2 C0 ( 0 ) 2 @x 8 @x
2

(10) @x
2

C x ( )2 2 2 + 0 0 @x @x 8

C ( ) 2 2 + 0 0 @x @x 2

@x

!#

P0

It simply involves derivatives of the Black-Scholes price with respect to ln S. The subscript 0 in 0 indicates that x C0 , C0 , C0 are evaluated in the unperturbed state, that is using the variance curve observed at time t = 0.

Expansion of the implied volatility

We now translate the perturbation of the price into a perturbation of the implied volatility. To ease notations we set x C x = C0 ( 0 ), C = C0 ( 0 ) and C = C0 ( 0 ). The calculation is performed in appendix A: we prove that, in the 4

generic stochastic volatility model (1), at second order in ", the implied volatility for maturity T and strike K is simply quadratic in log-moneyness: b" (T; K) = bATM + ST ln T bATM T = ^ VS 1 + T K S0 + CT ln2 K S0 + O("3 ) (11)

with

where v =

The at-the-money (ATM) implied volatility bATM is the VS volatility plus a spread, which, at rst order, is worth T " p C x . Typically, for equities, spot returns and forward variances are negatively correlated, so that C x < 0 and 4 v this spread is negative: the ATM volatility lies below the VS volatility. In the case when spot returns and forward "2 variances are uncorrelated, C x = C = 0 so that bATM = ^ VS 1 32v2 (v + 4) C . As C 0, the ATM T T volatility lies again below the VS volatility. The higher the vol of vol, the lower the ATM implied volatility.

RT
0

"2 " x 2 12 C x C + v (v + 4) C 3 4v 32v " x "2 2 ST = ^ VS C + 3 4C v 3 C x T 2 2v 8v "2 2 CT = ^ VS 4 4C v + C v 6 C x T 8v pv s VS 0 ds and ^ T = T , the VS implied volatility for maturity T .

+ 4v (v

4) C

(12) (13) (14)

The ATM skew ST is a pure product of spot/variance covariance. It vanishes when spot and forward variances are uncorrelated, both at rst and second order as it should since it is a well-known result that, in the uncorrelated case, the smile is symmetrical in log-moneyness. At order one in ", ST has the sign of C x : at this order the dierence between the ATM volatility and the VS volatility has the same sign as the ATM skew. Note that, at this order, we recover the well-known relationship linking the ATM and VS volatilities and the ATM skew, for smiles that are linear in log-moneyness, at rst order in the ATM skew: bATM = ^ VS + T T ^ VS T 2
2

T ST

At order one in ", ST is simply proportional to C x . Why is ST related to the double time-integral of the spot/variance covariance? We show in appendix B that this exact same double integral appears in the expression of the skewness of ln ST at order one in ". Thus, at order one in ", ST is simply proportional to the skewness of ln ST . This link was already derived in [1] using a Gram-Charlier expansion, at order one in the third order cumulant of the density of ln ST . The curvature CT is of order "2 . Not only does it involve variance/variance covariance: spot/variance covariance C (squared) contributes as well. If spot and variances are uncorrelated, CT = 8v5=2 pT "2 , hence positive. CT can be negative if spot and variances are very negatively correlated see (17) below.
K K The expansion of the implied volatility at order 2 has no term in ln3 S0 and ln4 S0 , because a special cancellation occurs see appendix A. This cancellation does not occur in the expansion of the price itself. This property was already noticed by Lewis (see [10], page 85) in the particular case of a rst-generation model, along with the concurrent observation that expansion of the implied volatility is numerically more accurate than the expansion of the price at same order in ".

Expansion (11) at order two in " is quadratic in log-moneyness does this generate arbitrages? Strikes near the money are not a concern. Are there arbitrages further out? The absence of arbitrage requires that asymptotically, b2 (T; K) grow at most linearly with ln(K=S0 ) for small/large strikes see [9]. Formula (11) is thus bound to generate arbitrage far away from the money while the reminder in (11) is of order 3 in " it is large for small/large K. This is conrmed by noticing that the density for ln ST implied by expression (10) for P is a Gaussian density multiplied by a 1 plus a linear combination of Hermite polynomials in x of order up to 6: the density will thus go negative for small/large spot values. In numerical tests below we show however that for ranges of strikes around the money that are practically relevant, approximate volatilities given by (11) are close to true impled volatilities, which are arbitrage-free by construction. 5

Short maturities

Short-maturity asymptotics allows us to shed light on the dependence of the ATM skew and curvature on the ATM volatility. To keep computations straightforward, assume that the dynamics of the instantaneous spot variance Vt = t t satises dVt = dt + "Vt' dBt (15) This holds for the family of Heston-like models that we study below it also holds for the (unsmiled) Bergomi model for short maturities.1 Let denote the correlation between the asset and the instantaneous variance. Then the short maturity version of (12)-(14), at order zero in T reads: bATM S C where
0

'

'

" 4

0 2' 2 0

(16) 7 48
2

'

"2

1 ' 12
0

1 24

4' 5 0

(17)

as a function of b

0 0.

ATM

At order zero in T , bATM is equal to : S0 C0 ' ' " 4 "2

and we get the following limits for the short skew and curvature

bATM

2' 2

(18) 7 48
2

1 ' 12

1 24

Equation (18) shows that, at rst order in vol of vol, the short ATM skew does not depend on the short term ATM volatility if ' = 1. Historically, short ATM skews of index equity smiles appear to be rather independent on the level of short ATM volatility. Parameterizing our model with ' = 1 would ensure that we price vanishing short ATM volatility/ATM skew cross-gamma. As for the short ATM curvature, (19) shows that C0 does not depend on the ATM volatility if ' = 5=4. This is also the critical value of ' above which the ATM curvature is positive, irrespective of the spot-variance correlation. In the Heston model, ' = 1=2: expressions (18)-(19) recover the well-known result that in this model the short ATM skew is inversely proportional to the short term ATM volatility and the short ATM curvature is inversely proportional to the cube of the short ATM volatility. For ' = 1, while the short ATM skew is independent on the level of the short ATM volatility, the curvature is inversely proportional to the ATM volatility. These scaling relationships can be derived many other ways see for example [2],[5]. The short maturity expansion at order one of the smile in a stochastic volatility model is derived in [7].

bATM

4' 5

(19)

Long maturities

We now focus on the term-structure of the ATM skew and curvature for long maturities. For simplicity we take a at initial term-structure of VSs: t . Let us assume that the underlying model is time-homogeneous: the spot/variance 0 x covariance function is a function of u t only: (t; u; ) ( ). One of the integrals in expression (??) for Ct can be done analytically at order one in " the ATM skew has the following simple expression, given in [5]: RT (T ) ( )d ST = 0 (20) 3=2 2 2 T Let us assume that for large u t, the instantaneous spot/variance covariance function exhibits a power-law decay with exponent > 0: (t; u; y) / (u t) . Using (20) we get the following behaviour for ST for long maturities, at order one in vol of vol: T if < 1 ST / T 1 if > 1 Exponential decay produces the same T 1 scaling of the long term ATM skew as the case > 1. Whenever < 1, ST decays as a power-law for long maturities, with an exponent that is exactly a signature of the long-time decay of the spot/variance covariance function.
1 Obviously

" in (15) is not dimensionless anymore.

For a fast-decaying spot/variance covariance function, the 1=T decay of the ATM skew is expected: it is due to the p fact that log returns of St become in eect independent, thus leading to a decay of the skewness of ln ST in 1= T . How fast is fast? The expression above for ST shows that as soon as ( ) decays faster than 1= , the long-maturity ATM skew behaves as that of a process for ln St with independent increments. In the Heston model as well as the Bergomi model, one indeed observes an exponential decay of the spot/variance covariance function for large values of u t, hence the long maturity ATM skew in these models will scale as T 1 . 1 Typically, equity index smiles display a power-law decay ST / T with 2 : in the context of a stochastic volatility model, this corresponds to a slowly decaying spot/variance covariance function. While, for (very) long maturities, the Bergomi model generates a T 1 decay, the presence of multiple time scales in the dynamics of variances makes it possible to produce a power-law-like decay over a wide range of maturities compatible with market skews of equity indicessee [5]. In [5] a connection is also established between the decay of ST with T and the skew-stickiness ratio for long maturities, which quanties how the ATM volatility moves when the spot moves, in units of the ATM skew. Turning now to the ATM curvature, assume that for large u t, the instantaneous variance/variance covariance function decays as a power law with exponent : (t; u; u0 ; y) / (u t) (u0 t) , with > 0. This corresponds to a situation where long-dated forward variances u have volatilities that scale like (u t) and are 100% correlated. Let t us make the additional assumption not justied for equities that spot and variances are uncorrelated. Then at lowest order in vol of vol, for long maturities, T 2 if < 1 CT / T 2 if > 1 Again, exponential decay generates the same T 2 scaling as the case > 1. In conclusion, a slow decay of the ATM skew (resp. curvature) can be traced, in the context of a stochastic volatility model, to the existence of long-range spot/variance (resp. variance/variance) correlations and the characteristic exponent of these long-range correlations can be read o the vanilla smile.

Example 1: a family of Heston-like models


p 1 Vt dt + Vt dWt1 ; 2

Consider a Heston-like model of the following type: dxt dVt = = x0 = x


'

(21)
2 dW 2 t

k (Vt

V1 ) dt + (Vt )

1 The Heston model corresponds to ' = 2 . In model (21), the instantaneous forward variance reads u t

p dWt1 + 1

V0 = V

= E [Vu jVt ] = V1 + (Vt


k(u t) t t t ' t

V1 ) e p 1

k(u t)

and its dynamics is: d


u t

= e

dWt1 +

2 dW 2 t k(u t)

The volatility of u is simply the volatility of t structure of instantaneous forward variances is

= Vt multiplied by the exponential factor e = V1 + (V V1 ) e


ku

. The initial term-

u 0

As is typical of rst-generation stochastic volatility models, this term-structure is determined by the model parameters and the initial value of the instantaneous variance. In model (21), the volatility (t; u; ) of instantaneous forward variances depends on the variance curve only through the instantaneous spot variance t : t
1 (t; u; 2 (t; u;

) )

= =

t '

As a consequence,

p 1

k(u t) t '

k(u t)

Cx C C

= = =

2 XZ i=1

ds ( s ) 0
T

1 '+ 2

k(T

s)

ds 1 2

Z k

du Z
T

i (s; u;

0)

!2 Z
s

=
T

k2

ds ( s ) 0
1 2

2'

1 1

k(T

s)

2 2

'+

ds ( s ) 0

1 '+ 2

du (

u ' 0)

k(u s)

k(T

u)

This coincides with Equations (3.7) to (3.10) in [10], where J (1) = C x , J (3) = C =2, and J (4) = C .

Example 2: the Bergomi model


1 t dt + 2 t
u t

In the Bergomi model, dxt d with dhW S ; W X it =


SX dt, u t

= =

! (1
SY

t S t dWt

)e

kX (u t)

dWtX + e
XY 2

kY (u t)

dWtY

dhW S ; W Y it =

dt and dhW X ; W Y it = ) +2
2 XY

dt. The normalizing factor


1=2

= (1

(1

)+

(22) is a mixing parameter for

is such that the very short-term variance t has (lognormal) volatility !. We pick kX > kY , t the short-term and long term factors W X and W Y . Introduce now independent Brownian motions W 1 , W 2 and W 3 : WS WX WY where = W1 = =
1 SX W + SY

W1 +

q 1 q

2 W2 SX

2 SY

W2 +

q (1
SY 2 SY

2 ) (1

2 SY

)W 3

SX , read:

SY

XY

indeed make up a correlation matrix only if


1 (t; u;

= p XY 1 ) )
SX e

2 SX

2 [ 1; 1]. The volatilities of variance +


SY

SX p 1

=(

1;

2;

3)

then

= = =

! ! !

(1 (1 q (1 !

kX (u t)

kY (u t)

2 (t; u; ) 3 (t; u;

q 1
2 ) (1

2 e kX (u t) SX 2 SY

)e

kY (u t)

2 SY

kY (u t)

The

all have the following common form:


i (t; u;

)=

wiX e

kX (u t)

+ wiY e

kY (u t)

where weights w depend on the correlations and the mixing parameter .

6.1

The covariance functions


Z
T

The covariance functions are given by: Cx = = du Z


u

dt

t 0 1 (t; u; 0 )

! (1
3 XZ i=1 2 T

) Z
T

SX T

du

u 0

dt !2

t kX (u t) 0e

SY

du

u 0

q dt t e 0

kY (u t)

ds
3 XZ i=1

du

i (s; u; 0 )

= Z

ds wiX du p

ds

du

u kX (u s) 0e

+ wiY dt

s 0 1 (s; u; 0 )

In the general case of a non-constant term-structure of VS volatilities, C x ; C and C are double or triple integrals that must be evaluated numerically. They are e ciently calculated by Guassian quadrature using few points in our experience, using four points per dimension is usually su cient for getting accurate values. In the special case of a initial VS term structure ( t at ), the integrals can be explicitly computed. Setting: 0 I( ) = we get Cx C with w0 wXX and C = with C1 C2 and
00 wX 00 wXX 2

1 p 2

u 0

du

u kY (u s) 0e

!2
1 u (r; u;

T 1 (u; t; 0 )

dr

r 0

)
=
0

J( )=

1+e
2

K( ) =

e
2

H( ) =

J( )

K( )

= =

! 3=2 T 2 (w1X J (kX T ) + w1Y J (kY T )) 2 2 2 3 ! T (w0 + wX I(kX T ) + wY I(kY T ) + wXX I(2kX T ) + wY Y I(2kY T ) + wXY I((kX + kY ) T ))
3 X i=1 2 3 X wiX k T i=1 X 3 X wiY k T i=1 Y

wiY wiX + kX T kY T

; wX =

wiY wiX + kX T kY T

; wY =

wiY wiX + kX T kY T

3 3 3 X w2 X wiX wiY X w2 iY iX ; wY Y = ; wXY = 2 k2 T 2 k2 T 2 k k T2 i=1 Y i=1 X Y i=1 X

! 2 2 T 3 (C1 + C2 )

1 2 1 2 J (kY T ) J (kX T ) w H(kX T ) + w1Y H(kY T ) w1X w1Y 2 1X 2 (kY kX ) T 00 00 00 00 00 = wX J (kX T ) + wY J (kY T ) + wXX J (2kX T ) + wY Y J (2kY T ) + wXY J ((kX + kY ) T ) =
2 w1X w1X w1Y + ; kX T kY T 2 w1X 00 ; wY Y = kX T 2 w1Y w1X w1Y + kY T kX T 2 w1Y w1X w1Y 00 ; wXY = kY T kX T 00 wY =

= =

w1X w1Y kY T

In Figure 1, we plot C x , C and C as a function of maturity, for the model parameters given in Table 1 and ! = 400%. C x is proportional to !; C and C are proportional to ! 2 .

0.25

kX 8

kY 0.35

SX

SY

XY

-0.8

-0.48

-0.73

(0:2)2

Table 1: Parameters of the 2-factor Bergomi model.

6.2

Numerical experiments

We now check the numerical accuracy of expansion (11) in the example a two-factor Bergomi model, with a initial at term structure of VS variances we simply set " = 1 in expressions (12)-(14). We use the parameters in Table 1. Parameters kX ; kY ; ; XY specify the dynamics of forward variances: their values are such that they generate a power-law-like decay for the instantaneous volatility of VS volatilities as a function of maturity with an exponent around 0:4 see Figure 1 for ! = 400%. This power-law decay of volatilities of volatilities re ects the typical term-structure of historical volatilities of VS volatilities of equity indices. In our numerical experiments, !, the volatility of a very short-term variance, varies from 20% to 400%. Note that the overall level of volatilities of volatilities in Figure 1 is consistent with implied levels of volatilities of volatilities of equity indices, rather than historical levels, which are much lower. Picking ! = 400% corresponds to choosing a lognormal volatility of a very short-term volatility equal to 200%, and generates high levels of forward skew.

     
  
  !   
  
   
61 0 5 1 4 1 3 1 2 1 0 1 (1 2 (1 3 (1 4 (1 5 (0 61 

) )  0 6 2 7 3 8 4 9 5 "##% $$ "#%$ $#% "# &'

7
8
8  7
 7
7 7    89 9 
 7 

8 
8    7
31 62 31 52 31 42 31 12 02 1 62 1 52 1 42 1 1 2 1 3 4 3 5 4  6

7 6
7 7 8 9 6 5 0  !# $# ( " $% ' ) * !+ #$- . 0# 1 15 # (*,! *% ) 1 . 2 /

Figure 1: Left: Term-structure of C x , C and C . Right: Term-structure of instantaneous volatility of VS volatility. Both term-structures are computed in the 2-factor Bergomi model with parameters given in Table 1 and ! = 400%.

6.2.1

First order

We here consider the expansion at order one in ". At this order, in the case of a term structure of VS volatilities, the at expression of the ATM skew takes the following simple form, already given in [4]: ! kX T 1 e kX T kY T 1 e kY T ! ST = (1 ) SX + SY 2 2 2 kX T 2 kY T 2 where is dened in (22).The ATM volatility and the 95%-105% skew (dierence of the implied volatilities of the 95% and 105% strikes) are graphed in Figure 2, as a function of maturity, both at order one in " and computed in a Monte Carlo simulation, for dierent levels of volatility of volatility !. As is apparent, the order one approximation of the ATM skew is acceptable, even for ! = 400%. The values of SX ; SY ; we have chosen are such that the ATMF skew decays 1 approximately like pT . At order two not shown expression (13) for the ATM skew provides excellent accuracy. By contrast, the ATM volatility is well captured by the expansion at order one only for small values of ! (say, up to 60%). True ATM implied volatilities lie below their order-one approximation.

10

9
      
33 5 24 18 0 24 17 0 24 16 0 24 15 0 24 13 0 24 0 24 88 0 24 87 0 24 86 0 24 85 3  ! #%5 4& '! 0 " 3 ( '  ! #%7 4& '! 0 " 3 ( '  ! #%5 3 ( ' " 3 4& '! 0 0 5   ! #%5 4& " 3 )*  ! #%7 4& " 3 )*  ! #%5 3 )* " 3 4& 6         7  8

1 5 1 4 1 3 1 2 1 0  1  1   0 2

678

  0 1          0 1       5 1          5 1       0       1    0      1   3       1    3      1

3 4 !"$# " # %$ 

Figure 2: Term-structures of ATM volatility (left) and the 95%-105% skew (right) in the 2-factor Bergomi model with parameters given in Table 1, at order one in volatility-of-volatility.

02 5 02 4 02 0 02 3 02 1 32  32  1 2 2 1

6 9   78 7


3     3 3   4     3 4        3   

1 52 3 42 1 42 3 02 1 02 2 3

6 9   78 7
 

        4     4          

31 12 31 2 6!"  #
$ % &8 89 # 8
6 !

%8 6

01 12

01 2

2 1

1 2

32 1

1 12 1 32 6 8
 "
# $
%8  6 ! 6
 $889 " 

01 12

01 32

Figure 3: Smile of implied volatility for maturities 1 year, 3 years, and 8 years, in the 2-factor Bergomi model with parameters given in Table 1, at order one in volatility-of-volatility.

11

Figure 3 shows that the global shape of the smile is well captured by the rst order expansion, i.e., the true implied volatility for strike K is indeed approximately a ne in ln(K=S0 ), but the overall level of the smile is well captured only for small values of !: we need to go beyond the order one expansion to get acceptable accuracy. We now consider the expansion at order two (12)-(14). 6.2.2 Second order

The case of uncorrelated spot and variances Consider rst the case of uncorrelated spot and variances. In this case, the ATM skew vanishes, and so does its expansion at second order in ". We choose the parameters in Table 2. 0.25 kX 8 kY 0.35
SX SY XY

(0:2)2

Table 2: Parameters of the 2-factor Bergomi model when spot and variances are uncorrelated. In Figure 4, we plot the ATM implied volatility as a function of maturity, as well as the full smile for dierent maturities. The ATM implied volatility is extremely well estimated by the second order expansion, even for ! = 400% and for long maturities. The second order expansion of the implied volatility is excellent around the money, but deteriorates farther away. This is not surprising: while no arbitrage requires that for very small and very large strikes, the implied volatility squared b2 (T; K) grows at most linearly with ln(K=S0 ) (see [9]), from (11), the second order estimate for b2 (T; K) grows like ln4 (K=S0 ). This means that in (11) the remainder O(! 3 ) = R(!; T; K) is large for small or large K and nite !. The general case of correlated spot and variances The accuracy of the order two expansion in the case of correlated spot and variances is examined in Figures 5 and 6. We use again the parameters in Table 1. Going to second order in " indeed improves accuracy with respect to the order one approximation.

Conclusion

The general expansion at order two in volatility of volatility presented above for stochastic volatility models is accurate enough that it can be used practically for strikes not too far from the money and for levels of volatility-of-volatility that are adequate (say up to 400% for short-dated variances). For instance, our approximation can be used in a Monte Carlo pricer whenever implied volatilities of vanilla options are needed either as observables or as regressors in a Longsta-Schwartz algorithm. It can also be used to calibrate a atoschastic volatility model to a term structure of implied volatilities, be they ATM or not. The fact that our expansion exactly preserves the implied volatilities of specic payouts in our case VSs is instrumental in determining its accuracy, and sets it apart from other approaches. At order two in volatility of volatility, the vanilla smile re ects the dynamics of the model at hand through three dimensionless quantities only C x ; C ; C which compactly summarize the model dynamics. At order one in s volatility of volatility we establish a simple link between the ATM skew and the skewness of ln ST . The ATM skew is very well captured at order one and only depends through C x on the spot/variance covariance function (t; u; 0 ) evaluated on the initial variance curve. Conversely, for short maturities, the instantaneous covariance of the spot and the instantaneous variance can be read o the market smile in model-independent fashion see [5]. This is not the case for the volatility of volatility itself: expression (14) for the ATM curvature involves C , which depends on how C x varies as the variance curve changes: a snapshot of the smile is not su cient for backing out the volatility of volatility model-independently, even for short maturities. Turning to the long-maturity case, we highlight how, in a time-homogeneous model, the characteristic exponent of the decay of the spot/variance covariance function for long time separations is directly manifested in the decay of the ATM skew. Can the static vanilla smile produced by a stochastic volatility model be used to further characterize the dynamics of the underlying model? We leave this to future work.

Appendix A: expansion of the implied volatility

Let us denote by P (0; x; ; K) the price at time 0 of the call option with strike price K. The implied integrated variance V " (T; K) is dened by P (0; x; ; K) = PBS (x; V " (T; K); K) 12

7 24 66 0 24 53 0 24 56 0 24 13 0 24 16 0 24 $3 6 0

8
   89 99
  9 

7 01 6 01 5 01 4 01

8   9
9  
  0      2  0   2   7  0   7  

7
  89 8  
02 6 02 5 02 4 02 3 1 02 12 6
1  0   1   3  0   3     0     

  0 6   7   7 4     7 6   7   6 4     # 6   7   6 4  

  0 6 !"   7 4   7 6 !"   6 4   # 6 !"   6 4

3 01 2 01 0 01 01  01 ! 1

% # 3 8'(9  ( 9   8)

&

! 1

41 !

! !1

! 41

0! !1

0! 41

12 5

 2

42 

1 2

1 42

0 2

0 42

8#$  % & '" (


" 8 "
8 #  '

%"

7!"  # $ % &9 9
# 9 7 !  %9 7

Figure 4: Term-structure of ATM volatility and smile of implied volatility for maturities 1 year, 3 years, and 8 years, in the 2-factor Bergomi model with parameters given in Table 2, at order two in volatility-of-volatility.

9
         
5 24 35 5 24 33 0 24 18 0 24 17 0 24 16 0 24 15 13 0 24 0 24 88 0 24 87 0 24 86 3 0 5 % 6 & ()  )   * + 7 ' 8  !0   !5  #$

9
         
8 24 33 0 24 75 0 24 73 0 24 65 0 24 63 0 24 15 0 24 13 0 24 %5 0 24 %3 0 24 55 0 24 53

 !0   !8  #$

&

' 5 ()  )   * +

Figure 5: Term-structure of ATM volatility in the 2-factor Bergomi model with parameters given in Table 1.

0 35 89
12 714 4
                      0!" 9 #4$ %9 7&2 23 #7 24 0 !4 84 %29 0                            

02 1 2  2  2  2  2  2  2 1 2 2 1

3 68   45
47   7
        0     0          

1 12 1 2 3"# $7% &! 7


'5 ! 3 !57 3 "7 &556 $
!

1 12

1 2

Figure 6: Smile of implied volatility for maturities 1 year, 3 years, and 8 years, in the 2-factor Bergomi model with parameters given in Table 1, at order two in volatility-of-volatility. 13

where PBS (x; v; K) = E ex+


p vG
1 2v

N (0; 1)

(23)

= ex N (d+ (x; v; K))

KN (d (x; v; K))

is the well-known Black-Scholes price of the call option, with the usual notation x Z y 2 ln e e z =2 p dz; d (x; v; K) = pK N (y) = v 2 1
X Recall that C X = C0 ( 0 ), C

1p v: 2

= C0 ( 0 ) and C = C0 ( 0 ). Assuming

V " (T; K) = V0 (T; K) + "V1 (T; K) + "2 V2 (T; K) + O "3 we get V0 (T; K) V1 (T; K) V2 (T; K) = = = Z
T r

dr ln e K V0 (T )
x

1 2

V0 (T ) !

CX V12 (T; K) +
2 2 1 @x @v PBS CX 2 @v PBS 2

2 1 @v PBS C 2 @v PBS

2 @x @v PBS C @v PBS

where all the Black-Scholes greeks involved in the last formula are evaluated at (x; V0 (T ); K). The three ratios of greeks K K involved in this formula are polynomials in L = ln ex = ln S0 :
2 @v PBS @v PBS 2 @x @v PBS @v PBS 2 2 @x @v PBS @v PBS

= = =

v 1 L2 1 2v v 4 v 1 L2 +L+ 1 v v 4 L3 L2 L4 1 + 3 2 2 v 2v 2v v

3 v + 2 8

3 2

v2 32

(24)

A derivation may for instance be found in [10], page 85. The implied volatility is then simply r V " (T; K) " b (T; K) = = b0 (T ) + "b1 (T; K) + "2 b2 (T; K) + O "3 T with b0 (T ) = = = r V0 (T ) T 1 L + 2 V0 (T )

b1 (T; K) b2 (T; K) a0 (v) a1 (v) a2 (v) = = = 1 32v 5=2 p

2 X i=0

ai (V0 (T ))Li

CX p 2 T V0 (T )

with L = ln

K S0

and

12 C X

C v (v + 4) + 4C v (v

4)

1 p 4C v 3 C X 2 8v 5=2 T 1 p 4C v + C v 6 C X 7=2 T 8v

Hence the formula (11) for the smile. 14

Appendix B: linking ST and the skewness of ln ST


M2

We derive here a general relationship linking the ATM skew ST and the skewness sT of the distribution of the log-spot xT = ln (ST ), at rst order in ". ST is given by equation (13). Let us now compute the skewness sT = M3 where 3=2 Mn = E[(xT We have: xT E[xT ] = 1 2 Z
T t t t 0

E[xT ]) ] Z
T

dt +

RT When " = 0, we are back to the Black-Scholes model, where the density of ln ST is gaussian: M2 = 0 t dt, M3 = 0. 0 For the sake of computing sT at order one in ", we thus only need M2 at order zero in " and M3 at order one in ". We RT then take M2 = 0 t dt: 0 The expression of M3 at order one in " is given in [5]: M3 = 3" Z
T

t 1 t dZt

dt

ds

At rst order in the vol of vol, the skewness of (the distribution of) ln (ST =S0 ) is thus given by: sT = 3"C x RT
0 t 0 dt 3=2

s 0 1 (s; t; 0 )

= 3"C x

and, from (13), the ATM skew ST simply reads

sT ST = p + O("2 ) 6 T

References
[1] Backus, D., Foresi, S., and Wu, L., Accounting for Biases in Black-Scholes, 2004. Available at SSRN: http://ssrn.com/abstract=585623. [2] Balland. P., Forward smile, Global Derivatives Conference, Paris, 2006. [3] Bergomi L., Smile Dynamics 2, Risk Magazine, pages 67-73, October 2005. [4] Bergomi L., Smile Dynamics 3, Risk Magazine, pages 90-96, October 2008. [5] Bergomi L., Smile Dynamics 4, Risk Magazine, pages 94-100, December 2009. [6] Gatheral J., Consistent modeling of SPX and VIX options, Global Derivatives conference, Paris, 2008. [7] Henry-Labordre P., Analysis, Geometry, and Modeling in Finance: Advanced Methods in Option Pricing, Chapman & Hall/CRC Financial Mathematics Series, 2008. [8] Heston S., A Closed-Form Solution for Options with Stochastic Volatility with Applications to Bond and Currency Options, The Review of Financial Studies, 6(2), 327 343, 1993. [9] Lee R., The moment formula for implied volatility at extreme strikes, Mathematical Finance, 14(3), 469-480, 2004. [10] Lewis A., Option valuation under stochastic volatility, Finance Press, 2000.

15

You might also like