Time Dependent Heston Model: Abstract

TIME DEPENDENT HESTON MODEL
E. BENHAMOU , E. GOBET , AND M. MIRI
Abstract. The use of the Heston model is still challenging because it has a closed formula only when
the parameters are constant [Hes93] or piecewise constant [MN03]. Hence, using a small volatility of
volatility expansion and Malliavin calculus techniques, we derive an accurate analytical formula for the
price of vanilla options for any time dependent Heston model (the accuracy is less than a few bps for
various strikes and maturities). In addition, we establish tight error estimates. The advantage of this
approach over Fourier based methods is its rapidity (gain by a factor 100 or more), while maintaining a
competitive accuracy. From the approximative formula, we also derive some corollaries related first to
equivalent Heston models (extending some work of Piterbarg on stochastic volatility models [Pit05b]) and
second, to the calibration procedure in terms of ill-posed problems.
Key words. asymptotic expansion, Malliavin calculus, small volatility of volatility, time dependent
Heston model
This version: March 24, 2009
hal-00370717, version 1 - 24 Mar 2009
AMS subject classifications. 60J75, 60HXX
1. Introduction. Stochastic volatility modeling has emerged in the late nineties

as a way to manage the smile. In this work, we focus on the Heston model which
is a lognormal model where the square of volatility follows a CIR1 process. The call
(and put) price has a closed formula in this model thanks to a Fourier inversion of
the characteristic function (see Heston [Hes93], Lewis [Lew00] and Lipton [Lip02]).
When the parameters are piecewise constant, one can still derive a recursive closed
formula using a PDE method (see Mikhailov and Nogel [MN03]) or a Markov argument in combination with affine models (see Elices [Eli08]), but formula evaluation
becomes increasingly time consuming. However, for general time dependent parameters there is no analytical formula and one usually has to perform Monte Carlo
simulations. This explains the interest of recent works for designing more efficient
Monte Carlo simulations: see Broadie and Kaya [BK06] for an exact simulation and
bias-free scheme based on Fourier integral inversion; see Andersen [And08] based
on a Gaussian moment matching method and a user friendly algorithm; see Smith
[Smi08] relying on an almost exact scheme; see Alfonsi [Alf08] using higher order
schemes and a recursive method for the CIR process. For numerical partial differential equations, we refer the reader to Kluges doctoral dissertation [Klu02].
Comparison with the literature. A more recent trend in the quantitative literature has been the use of the so called approximation method to derive analytical
formulae. This has led to an impressive number of papers, with many original ideas.
For instance, Alos
` et al. [ALV07] have been studying the short time behavior of implied volatility for stochastic volatility using an extension of Itos
formula. Another
trend has focused on analytical techniques to derive the asymptotic expansion of the
implied volatility near expiry (see for instance Berestycki et al. [BBF04], [Lab05],
Hagan et al. [HKLW02], Lewis [Lew07], Osajima [Osa07] or Forde [For08]). But in
Pricing Partners, 204 rue de Crimee 75019, Paris, FRANCE (eric.benhamou@pricingpartners.com).

Jean Kuntzmann, Universite de Grenoble and CNRS, BP 53, 38041 Grenoble cedex 9,
FRANCE (emmanuel.gobet@imag.fr). Corresponding author.
Pricing Partners, 204 rue de Crim
ee 75019, Paris, FRANCE (mohammed.miri@pricingpartners.com).
1 Nice properties for the CIR process are derived by Dufresne [Duf01], Going-Jaeschke
and Yor [GJY03],

Diop [Dio03], Alfonsi [Alf05] and Miri [Mir09].
Laboratoire
Electronic copy available at: http://ssrn.com/abstract=1367955
E. BENHAMOU, E. GOBET AND M. MIRI
hal-00370717, version 1 - 24 Mar 2009
these works the implied volatility near expiry does not have a closed formula because
the related geodesic distance is not explicit. It can, however, be approximated by a
series expansion [Lew07]. The drawback to these methods is their inability to handle
non-homogeneous (that is to say time dependent) parameters. For long maturities,
another approach has been the asymptotic expansion w.r.t. the mean reversion parameter of the volatility as shown in [FPS00]. In the case of zero correlation, averaging
techniques as exposed in [Pit05b] and [Pit05a] can be used. Antonelli and Scarlatti
take another view in [AS09] and have suggested price expansion w.r.t. correlation.
For all of these techniques, the domain of availability of the expansion is restricted to
either short or long maturities, to zero correlation, or to homogeneous parameters.
In our work, we aim to give an analytical formula which covers both short and long
maturities, that also handles time inhomogeneous parameters as well as non-null
correlations. As a difference with several previously quoted papers, our purpose
consists also of justifying mathematically our approximation.
The results closest to ours are probably those based on an expansion w.r.t. the
volatility of volatility by Lewis [Lew00]: it is based on formal analytical arguments
and is restricted to constant parameters. Our formula can be viewed as an extension
of Lewis formula in order to address a time dependent Heston model, using a direct
probabilistic approach. In addition, we prove an error estimate which shows that
our approximation formula for call/put is of order 2 w.r.t. the volatility of volatility.
The advantage of this current approximation is that the evaluation is about 100 to
1000 times quicker than a Fourier based method (see our numerical tests).
Comparison with our previous works [BGM09b] and [BGM09a]. Our approach
here consists of expanding the price w.r.t. the volatility of volatility, and of computing
the correction terms using Malliavin calculus. In these respects, the current approach
is similar to our previous works [BGM09b] and [BGM09a], however, the techniques
for estimating error are different. Indeed, we use the fact that the price of vanilla
options can be expressed as an expectation of a smooth price function for stochastic
volatility models. This is based on a conditioning argument as in [RT96]. Consequently, the smoothness hypotheses (H1 , H2 , H3 ) of our previous papers are no longer
required. Note also that the square root function arising in the martingale part of the
CIR process is not Lipschitz continuous. Hence, the Heston model does not fit the
smoothness framework previously used. Therefore, to overcome this difficulty, we
derive new technical results in order to prove the accuracy of the formula.
Contribution of the paper. We give an explicit analytical formula for the price
of vanilla options in a time dependent Heston model. Our approach is based on an
expansion w.r.t. a small volatility of volatility. This is practically justified by the fact
that this parameter is usually quite small (of order 1 or less, see [Lew00] or [BK06]
for instance). The resulting formula is the sum of two terms: the leading term is the
Black-Scholes price for the model without volatility of volatility while the correction
term is a combination of Greeks of the leading term with explicit weights depending
only on the model parameters. Proving the accuracy of the expansion is far from
straightforward, but with some technicalities and a relevant analysis of error, we
succeed in giving tight error estimates. Our expansion enables us to obtain averaged
parameters for the dynamic Heston model.
Formulation of the problem. We consider the solution of the stochastic differen-
Electronic copy available at: http://ssrn.com/abstract=1367955
tial equation (SDE):
vt
dt, X0 = x0 ,
2
dvt = (t vt )dt + t vt dBt , v0 ,
dXt =
vt dWt
(1.1)
(1.2)
dhW, Bit = t dt,
hal-00370717, version 1 - 24 Mar 2009
where (Bt , Wt )0tT is a two-dimensional correlated Brownian motion on a filtered

probability space (, F , (Ft )0tT , P) with the usual assumptions on filtration (Ft )0tT .
In our setting, (Xt )t is the log of the forward price and (vt )t is the square of the volatility
which follows a CIR process with an initial value v0 > 0, a positive mean reversion ,
a positive long-term level (t )t , a positive volatility of volatility (t )t and a correlation
(t )t . These time dependent parameters are assumed to be measurable and bounded
on [0, T].
To develop our approximation method, we will examine the following perturbed
process w.r.t. [0, 1]:
dXt =
vt dWt
vt
2
X0 = x0 ,
dt,
dvt = (t vt )dt + t
vt dBt ,
v0 = v0 ,
(1.3)
so that our perturbed process coincides with the initial one for = 1 : Xt1 = Xt , v1t = vt .
For the existence of the solution v , we refer to Chapter IX in [RY99] (moreover, the
process is non-negative for kt 0, see also the proof of Lemma 4.2). Our main
purpose is to give an accurate analytic approximation, in a certain sense, of the
expected payoff of a put option :
g() = e
RT
0
rt dt
RT
E[(K e 0
(rt qt )dt+XT
)+ ]
(1.4)
where r (resp. q) is the risk-free rate (resp. the dividend yield), T is the maturity and
= 1. Extensions to call options and other payoffs are discussed later.
Outline of the paper. In Section 2, we explain the methodology of the small
volatility of volatility expansion. An approximation formula is then derived in Theorem 2.3 and its accuracy stated in Theorem 2.4. This section ends by explicitly
expressing the formulas coefficients for general time dependent parameters (constant, smooth and piecewise constant). Our expansion allows us to give equivalent
constant parameters for the time dependent Heston model (see Subsection 2.6). As a
second corollary, the options calibration for Hestons model using only one maturity
becomes an ill-posed problem; we give numerical results to confirm this situation.
In section 3, we provide numerical tests to benchmark our formula with the closed
formula in the case of constant and piecewise constant parameters. In Section 4,
we prove the accuracy of the approximation stated in Theorem 2.4: this section is
the technical core of the paper. In Section 5, we establish lemmas used to make the
calculation of the correction terms explicit (those derived in Theorem 2.3). In Section
6, we conclude this work and give a few extensions. In the appendix, we recall
details about the closed formula (of Heston [Hes93] and Lewis [Lew00]) in the case
of constant (and piecewise constant) parameters.
2. Smart Taylor expansion.
2.1. Notations. N 2.1. Extremes of deterministic functions.

For a càdlàg function l : [0, T] R, we denote lIn f = inft[0,T] lt and lSup = supt[0,T] lt .
N 2.2. Differentiation.
(i) For a smooth function x 7 l(x), we denote by l(i) (x) its i-th derivative.
(ii) Given a fixed time t and for a function ft , we denote (if it has a meaning) the
ith derivative at = 0 by fi,t =
i ft
| .
i =0
2.2. Definitions. In order to make the approximation explicit, we introduce the

following family of operators indexed by maturity T.
(.,.)
D 2.1. Integral Operator. We define the integral operator .,T as follows:
(i) For any real number k and any integrable function l, we set
T
Z
(k,l)
t,T =
eku lu du, t [0, T].
hal-00370717, version 1 - 24 Mar 2009
(ii) For any real numbers (k1 , , kn ) and for any integrable functions (l1 , , ln ), the
n-times iteration is given by
(k ,l1 ), ,(kn ,ln )
t,T1
write
(k ,l ), ,(kn ,ln )
(k1 ,l1 .,T2 2
)
= t,T
, t [0, T].
(iii) When the functions (l1 , , ln ) are equal to the unity constant function 1, we simply
k , ,kn
1
t,T
(k ,1), ,(kn ,1)
= t,T1
, t [0, T].
2.3. About the CIR process. Assumptions. In order to bound the approximation
errors, we need a positivity assumption for the CIR process.
Assumption (P). The parameters of the CIR process (1.2) verify the following
conditions:
In f > 0,
2
)In f 1.
2
This assumption is crucial to ensure the positivity of the process on [0, T], which is
stated in detail in Lemma 4.2 (remember that v0 > 0). We have
P(t [0, T] : vt > 0) = 1.
When the functions and are constant, Assumption (P) coincides with the usual
Feller test condition 2
1 (see [KS88]).
2
Note that the above assumption ensures that the positivity property also holds
for the perturbed CIR process (1.3): for any [0, 1], we have
P(t [0, T] : vt > 0) = 1
(see Lemma 4.2). We also need a uniform bound of the correlation in order to preserve
the non degeneracy of the SDE (1.1) conditionally on (Bt )0tT .
Assumption (R). The correlation is bounded away from -1 and +1:
||Sup < 1.
2.4. Taylor Development. In this paragraph, we present the main steps leading
to our results. Complete proofs are given later.
If (FtB )t denotes the filtration generated by the Brownian motion B, the distribuRT p
tion of XT conditionally to FTB is a Gaussian distribution with mean x0 + 0 t vt dBt
R
RT
1 T
2
2 0 vt dt and variance 0 (1 t )vt dt ( [0, 1]). Therefore, the function (1.4) can be
expressed as follows:
Z
q
t
g() = E[PBS (x0 +

0
Z
vt dBt
2t
2
Z
vt dt,
(1 2t )vt dt)],
(2.1)
where the function (x, y) PBS (x, y) is the put function price in a Black-Scholes model
RT
with spot ex , strike K, total variance y, risk-free rate req =

RT
r(t)dt
,
T
dividend yield
hal-00370717, version 1 - 24 Mar 2009
q(t)dt
0
T
qeq =
and maturity T. For the sake of completeness, we recall that PBS (x, y)
has the following explicit expression
!
!
Kereq T
Kereq T
1
1
1
1
x qeq T
req T
y e e
N log( q T )
y .
Ke
N log( q T ) +
2
2
y
y
ex e eq
ex e eq
In the following, we expand PBS (., .) with respect to its two arguments. For this, we
note that PBS is a smooth function (for y > 0). In addition, there is a simple relation
between its partial derivatives:
PBS
PBS
1 2 PBS
(x, y)
(x, y) = (
(x, y)),
2
y
2 x
x
x R, y > 0,
(2.2)
which can be proved easily by a standard calculation left to the reader.

Under assumption (P), for any t, vt is C2 w.r.t at = 0 (differentiation in Lp -sense).
This result will be shown later. In addition, v does not vanish (for any [0, 1]).
Hence, by putting vi,t =

dv1,t
i vt
i
v1,t dt + t
, we get
vt dBt + t
v1,t
p dBt ,
2 vt
v1,0 = 0,
v1,t
[v1,t ]2
v2,t
dv2,t = v2,t dt + t p dBt + t p dBt t
dBt ,
4[vt ]3/2
vt
2 vt
From the definitions vi,t
i vt
| ,
i =0
v2,0 = 0.
we easily deduce
Z
t
v0,t = e
(v0 +
es s ds),
0
t
Z
v1,t = et
es s v0,s dBs ,
(2.3)
0
t
Z
t
v2,t = e
es s
v1,s
1
(v0,s ) 2
dBs .
(2.4)
Note that v0,t coincides also with the expected variance E(vt ) because of the linearity
of the drift coefficient of (vt )t . Now, to expand g(), we use the Taylor formula twice,
first applied to vt and
p
vt at = 1 using derivatives computed at = 0:
v1t = v0,t + v1,t +

q
v1t =
v2,t
+ ,
2
v1,t
v0,t +
v2,t
2(v0,t ) 2
4(v0,t ) 2
v21,t
3
8(v0,t ) 2
+ ,
secondly for the smooth function PBS at the second order w.r.t. the first and second
R T 2
RT
RT
variable around (x0 + 0 t v0,t dBt 0 2t v0,t dt, 0 (1 2t )v0,t dt). For convenience,
we simply write
T
Z
P BS =PBS (x0 +
0
i+j P BS i+j PBS

=
(x0 +
xi y j
xi y j
t v0,t dBt
Z
t v0,t dBt
2t
2
Z
0
Z
v0,t dt,
0
T
2t
2
(1 2t )v0,t dt),
T
Z
v0,t dt,
0
(2.5)
(1 2t )v0,t dt).
hal-00370717, version 1 - 24 Mar 2009
Then, one gets

g(1) =E[P BS ]
(2.6)
v21,t
T
T 2
v1,t
v2,t
v2,t
P BS
t
t (
(
+
(v1,t +
)dt)]
)dB
t
3
1
1
x
2
0
0 2
2(v0,t ) 2 4(v0,t ) 2 8(v0,t ) 2
Z
v2,t
P BS T
+E[
(1 2t )(v1,t +
)dt]
y 0
2
Z T
Z T 2
t
v1,t
1 2 P BS
t
(
+ E[
dBt
v1,t dt)2 ]
1
2
2
x
0
0 2
2(v0,t ) 2
Z T
1 2 P BS
(1 2t )v1,t dt)2 ]
(
+ E[
2
y2
0
Z T
Z T
Z T 2
t
v1,t
2 P BS
2
t
(1 t )v1,t dt)(
(
dBt
v1,t dt)]
+E[
1
xy
0
0 2
0
2(v0,t ) 2
+E[
+E
(2.7)
(2.8)
(2.9)
(2.10)
(2.11)
(2.12)
where E is the error in our Taylor expansion. In fact, we notice that:

E[P BS ] = E[E[e
RT
0
rt dt
T
Z
= PBS (x0 ,
q
RT
RT
v
x0 + 0 (rt qt 0,t
v0,t (t dBt + 12t dB
2 )dt+ 0
t )
(K e
)+ |FTB ]]
v0,t dt),
0
where B is a Brownian motion independent on FTB . Furthermore, the relation (2.2)

remains the same for P BS and this enables us to simplify the expansion above. This
gives:
P 2.2. The approximation (2.12) is equivalent to
T
Z
g(1) = PBS (x0 ,
0
P BS
v0,t dt) + E[
y
Z
0
1 2 P BS
(
(v1,t + v2,t )dt] + E[
2
y2
Z
0
v1,t dt)2 ] + E.
The details of the proof are given in Subsection 5.2. At first sight, the above formula
looks like a Taylor formula of PBS w.r.t. the cumulated variance. In fact, it is different,
note that the coefficient of v2,t is not 1/2 but 1. We do not have any direct interpretation
of this formula.
The next step consists of making explicit the correction terms as a combination
of Greeks of the BS price.
T 2.3. Under assumptions (P) and (R), the put2 price is approximated by
RT
0
rt dt
RT
E[(K e 0
(rt qt )dt+XT1
)+ ] =PBS (x0 , varT ) +
2
X
ai,T
i=1
1
X
b2i,T
2i+2 P
i=0
BS
x2i y2
i+1 PBS
(x0 , varT )
xi y
(x0 , varT ) + E,
(2.13)
where
T
hal-00370717, version 1 - 24 Mar 2009
(,v0,. ),(,1)
a1,T = 0,T
v0,t dt,
varT =
0
(2,2 v0,. ),(,1),(,1)
b0,T = 0,T
(,v0,. ),(0,),(,1)
a2,T = 0,T
,
b2,T =
a21,T
2
The proof is postponed to Subsection 5.3. Finally, we give an estimate regarding the
error E arising in the above theorem.
T 2.4. Under assumptions (P) and (R), the error in the approximation (2.13) is
estimated as follows:

E = O [Sup T]3 T .
In view of Theorem 2.4, we may refer to the formula (2.13) as a second order
approximation formula w.r.t. the volatility of volatility.
2.5. Computation of coefficients.
Constant parameters. The case of constant parameters (, , ) gives us the coefficients a and b explicitly. Indeed in this case, the operator is a simple iterated
integration of exponential functions. Using Mathematica, we derive the following
explicit expressions.
P 2.5. Explicit computations. For constant parameters, one has:
varT =m0 v0 + m1 ,
2
a2,T =() (q0 v0 + q1 ),
a1,T =(p0 v0 + p1),

b0,T =2 (r0 v0 + r1 ).
2 The approximation formula for the call price is obtained using the call/put parity relation: in (2.13),
it consists of replacing on the l.h.s. the put payoff by the call one, and on the r.h.s., the put price function
PBS by the similar call price function, while coefficients remain the same.
where
m0 =
p0 =
q0 =
hal-00370717, version 1 - 24 Mar 2009
r0 =

eT 1 + eT
,

eT T + eT 1

eT T + eT (T 2) + 2
m1 = T
,
2

eT T(T + 2) + 2eT 2
23

e2T 4eT T + 2e2T 2
43
p1 =
,
q1 =
r1 =

eT 1 + eT
,
2

eT 2eT (T 3) + T(T + 4) + 6
,
23

e2T 4eT (T + 1) + e2T(2T 5) + 1
43
R 2.1. In the case of constant parameters (, , ), we retrieve the usual Heston

model. In this particular case, our expansion coincides exactly with Lewis volatility of
volatility series expansion (see Equation (3.4), page 84 in [Lew00] for Lewis expansion
formula and page 93 in [Lew00] for the explicit calculation of the coefficients J(i) with = 12 ).
Using his notation, we have a1,T = J(1) , a2,T = J(4) and b0,T = J(3) .
Smooth parameters. In this case, we may use a Gauss-Legendre quadrature
formula for the computation of the terms a and b.
Piecewise constant parameters. The computation of the variance varT is straightforward. Thus, it remains to provide explicit expressions of a and b as a function of
the piecewise constant data. Let T0 = 0 T1 Tn = T such that , , are constant
on each interval ]Ti , Ti+1 [ and are equal respectively to Ti+1 , Ti+1 , Ti+1 . Before giving
(,v0,. )
the recursive relation, we need to introduce the following functions: 1,t = 0,t
(2,2 v0,. )
0,t
,
(2k,2 v ),(,1)
0,t 0,.
.
(k,v ),(0,)
0,t 0,.
,
2,t =
t =
t =
P 2.6. Recursive calculations. For piecewise constant coefficients, one
has:
1
a1,Ti+1 = a1,Ti +
1,Ti + Ti+1 Ti+1 f,v
(Ti+1 , Ti , Ti+1 ),
T ,T
0,T
i
i+1
2
a2,Ti+1 = a2,Ti +
0,
1,Ti + (Ti+1 Ti+1 )2 f,v
(Ti+1 , Ti , Ti+1 ),
T ,T Ti + Ti+1 Ti+1
T ,T
0,T
i
i+1
i+1
b0,Ti+1 = b0,Ti +
,
2,Ti + 2T
T ,T Ti +
T ,T
i
Ti+1 =
i+1
i+1
Ti + Ti+1 Ti+1 (Ti+1 Ti ) 1,Ti + 2T 2T g1,v0,T (Ti+1 , Ti , Ti+1 ),

i+1
i+1
i
Ti+1 = Ti +
2,Ti + 2T
T ,T
i
1,Ti+1 =
i+1
0
f,v
(Ti+1 , Ti , Ti+1 ),
0,T
i+1
i+1
g2,v0,T (Ti+1 , Ti , Ti+1 ),

i
1,Ti + Ti+1 Ti+1 h1,v0,T (Ti+1 , Ti , Ti+1 ),

i
2,Ti+1 = 2,Ti + 2T h2,v0,T (Ti+1 , Ti , Ti+1 ),

i+1
v0,Ti+1 = e(Ti+1 Ti ) (v0,Ti Ti+1 ) + Ti+1 ,

where
0
f,v
(, t, T) =
0
e2T (e2t (2v0 )+e2T ((2t+2T5)+2v0 )+4e(t+T) ((t+T+1)+(tT)v0 ))

,
43
T eT ((t+T2)+v )et ((tT2)tv +Tv +v )
e
(
)
0
0
0
0
1 (, t, T) =
,
f,v
0
2
(t+3T)
(t+3T)
2(t+T)
e
2e
(((Tt)3)+v
)+e
((((tT)4)(tT)+6)(((tT)2)(tT)+2)v
(
0
0 ))
2 (, t, T) =
f,v
,
0
23
T
t
2
2
2e +e ( (tT) v0 (((tT)2)(tT)+2))
g1,v0 (, t, T) =
,
22
g2,v0 (, t, T) =
h1,v0 (, t, T) =
h2,v0 (, t, T) =
eT (e2T e2t (2v0 )+2e(t+T) ((tT)(v0 )v0 ))

,
22
eT +et ((tT1)+(Tt)v0 )
,
(et eT )(et (2v0 )eT )

,
2
tu
eTu (tu+Tu1)+etu
Tu
, u,u
and ut (T) = e u+e , 0,u
t (T) =
t (T) =
u2
Proof. According to Theorem 2.3, one has :
Ti
Z
a1,Ti+1 =
Z
t
e
0
(,1)
t t v0,t t,T dt +
i+1
Ti
Z
=a1,Ti +
Ti
Z
(,1)
i+1
(,1)
(,1)
=a1,Ti + T ,T
i
Ti
2u2
et t t v0,t t,T
et t t v0,t T ,T dt +
i
Ti+1
(etu eTu )
i+1
et t t v0,t dt +
Ti+1
Ti
Z Ti+1
Ti
i+1
dt
(,1)
et t t v0,t t,T
i+1
dt
(,1)
et t t v0,t t,T
i+1
dt
1
=a1,Ti +
1,Ti + Ti+1 Ti+1 f,v
(i+1 , Ti , Ti+1 ),
T ,T
0,T
hal-00370717, version 1 - 24 Mar 2009
i+1
1 and
where the functions f,v
are calculated analytically using Mathematica. The
0
other terms are calculated analogously.
2.6. Corollaries of the approximation formula (2.13).

Averaging Hestons model parameters. We derive a first corollary of the approximation formula in terms of equivalent Heston models. As explained in [Pit05b],
this averaging principle may facilitate efficient calibration. Namely, we search for
,
for the Heston model3
equivalent constant parameters ,
,
vt
v t dWt dt, X 0 = x0 ,
2
dv t = (
t v t )dt + v t dBt , v 0 = v0 ,
dX t =
dhW, Bit = dt,
that equalize the price of call/put options maturing at T in the time dependent model
(equality up to the approximation error E). The following rules give the equivalent
parameters as a function of the variance varT and the coefficients a1,T , a2,T , b0,T that
are computed in the time dependent model. Results are expressed using
a=
a2,T m1
,
m1 q0 m0q1
b=
a1,T m1
,
m1 p0 m0 p1
c = varT (
p1
q1
),
m1 p0 m0 p1 m1 q0 m0q1
where m0 , m1 , p0 , p1 , q0 , q1 , r0 and r1 are given in Proposition 2.5.

Averaging rule in the case of zero correlation. If t 0, the equivalent constant
parameters (for maturity T) are
s
= ,
varT m0 v0
,
=
m1
b0,T
r0 v0 + r1
= 0.
3 In this approach, we leave the initial value v equal to v . Indeed, it is not natural to modify its value
0
0
since it is not a parameter, but rather an unobserved factor.
10
hal-00370717, version 1 - 24 Mar 2009
Proof. Two sets of prices coincide at maturity T if they have the same approximation formula (2.13). In this case a1,T = a2,T = b2,T = 0, thus the approximation formula
depends only on two quantities varT and b0,T . It is quite clear that there is not a
single choice of parameters to fit these two quantities. A simple solution results from
the choice of = and = 0: then, using Proposition 2.5, we obtain the announced
parameters and .
we exactly retrieve Piterbargs
R 2.2. In this case of zero correlation and = v0 = ,
results for the averaged volatility of volatility ( see [Pit05b]).
Averaging rule in the case of non zero correlation. We follow the same arguments as before. Now the approximation formula also depends on the four quantities
varT , a1,T , a2,T and b2,T . Thus, equalizing call/put prices at maturity T is equivalent
and .
to equalizing these four quantities in both models, by adjusting ,
,
Unfortunately, we have not found a closed expression for these equivalent parameters.
An alternative and simpler way of proceeding consists of modifying the unobserved
initial value v0 of the variance process while keeping = . For non-vanishing
correlation (t )t , it leads to two possibilities
(b
v0 =b
s
=
p1 varT
b2 4ac)
,
2a
m1 p0 m0 p1
b0,T
r0 v0 + r1
varT m0 v0
=
,
m1

(b
2a
.
b2 4ac)
In practice, only one solution gives realistic parameters. However, this rule is heuristic since there is a priori no guarantee that these averaged parameters satisfy the
assumption (P), which is the basis for the arguments correctness.
Proof. Using Proposition 2.5, one has to solve the following system of equations
varT =m0 v 0 + m1,

2 (q0 v0 + q1 ),
a2,T =()
The first equation gives =
v 0 =(
a1,T
()
varT m0 v0
.
m1
p1 varT
m1
)
,
m1 p0 m1 p1 m0
0 v 0 + p1),
a1,T =(p
b0,T =2 (r0 v0 + r1 ).
Replacing this identity in a1,T and a2,T gives
v 0 =(
a2,T
2
()
q1 varT
m1
)
.
m1 q0 m1 q1 m0
It readily leads to a quadratic equation ax2 + bx + c = 0 with x = 1 . By solving this

equation, we easily complete the proof of the result.
Collinearity effect in the Heston model. Another corollary of the approximation
formula (2.13) is that we can obtain the same vanilla prices at time T with different
sets of parameters. For instance, take on the one hand v0 = = 4%, 1 = 2 and 1 = 30%
(model M1 ) and on the other hand v0 = = 4%, 2 = 3 and 2 = 38.042% (model M2 ),
both models having zero correlation. The resulting error between implied volatilities
within the two models are presented in Table 2.1: they are so small that prices can
be considered as equal. Actually, this kind of example is easy to create even with
non-null correlation: as before, in view of the approximation formula (2.13), it is
sufficient to equalize the four quantities varT , a1,T , a2,T and b2,T .
As a consequence, calibrating a Heston model using options with a single maturity is an ill-posed problem, which is not a surprising fact.
11
T 2.1
Error in implied Black-Scholes volatilities (in bp) between the closed formulas (see appendix) of the two models
M1 and M2 expressed as relative strikes. Maturity is equal to one year.
strikes K
80%
90%
100%
110%
120%
model M1
model M2
errors (bp)
20.12%
20.11%
0.69
19.64%
19.65%
-0.35
19.50%
19.51%
-0.81
19.62%
19.62%
-0.42
19.92%
19.92%
0.34
hal-00370717, version 1 - 24 Mar 2009
3. Numerical accuracy of the approximation. We give numerical results of the

performance of our method. In what follows, the spot S0 , the risk-free rate r and
the dividend yield q are set respectively to 100, 0% and 0%. The initial value of the
variance process is set to v0 = 4% (initial volatility equal to 20%). Then we study
the numerical accuracy w.r.t. K, T, , , and by testing different values for these
parameters.
T 3.1
Set of maturities and strikes used for the numerical tests.
T/K
3M
6M
1Y
2Y
3Y
5Y
7Y
10Y
70
60
50
40
30
20
10
10
80
70
60
50
40
30
30
20
90
80
80
70
60
60
50
50
100
100
100
100
100
100
100
100
110
110
120
130
140
150
170
190
120
130
150
180
200
250
300
370
125
140
170
210
250
320
410
550
130
150
180
240
290
400
520
730
In order to present more interesting results for various relevant maturities and
strikes, we allow the range of strikes
to vary over the maturities. The strike values
evolve approximately as S0 exp(c T) for some real numbers c and = 6%. The
extreme values of c are chosen to be equal to 2.57, which represents the 1%-99%
quantile of the standard normal distribution. This corresponds to very out-of-themoney options or very deep-in-the-money options. The set of pairs (maturity, strike)
chosen for the tests is given in Table 3.1.
Constant parameters. In Table 3.2, we report the numerical results when = 6%,
= 3, = 30% and = 0%, giving the errors of implied Black-Scholes volatilities
between our approximation formula (see Equation (2.13)) and the price calculated
using the closed formula (see appendix), for the maturities and strikes of Table 3.1.
The table should be read as follows: for example, for one year maturity and strike
equal to 170, the implied volatility is equal to 24.14% using the closed formula and
24.20% with the approximation formula, giving an error of -6.33 bps. In Table 3.2, we
observe that the errors do not exceed 7 bps for a large range of strikes and maturities.
We notice that the errors are surprisingly higher for short maturities. At first sight,
it is counterintuitive as one would expect our perturbation method to work better
for short maturities and worse for long maturities, since the difference between our
proxy model (BS with volatility (v0,t )t ) and the original one is increasing w.r.t. time.
In fact, this intuition is true for prices but not for implied volatilities. When we
12
T 3.2
Implied Black-Scholes volatilities of the closed formula, of the approximation formula and related errors (in bp),
expressed as a function of maturities in fractions of years and relative strikes. Parameters: = 6%, = 3, = 30%
and = 0%.
3M
6M
1Y
2Y
3Y
hal-00370717, version 1 - 24 Mar 2009
5Y
7Y
10Y
23.24%
23.06%
18.01
24.32%
24.12%
19.69
24.85%
24.78%
7.72
24.86%
24.86%
-0.21
24.95%
24.94%
1.80
24.88%
24.86%
1.38
25.03%
24.97%
5.72
24.72%
24.71%
0.42
22.14%
22.19%
-4.86
23.29%
23.36%
-7.17
24.06%
24.12%
-6.49
24.36%
24.40%
-3.51
24.53%
24.55%
-2.12
24.56%
24.57%
-0.96
24.46%
24.46%
-0.43
24.51%
24.51%
-0.28
21.43%
21.42%
0.53
22.55%
22.57%
-1.89
23.14%
23.14%
0.26
23.82%
23.82%
-0.12
24.10%
24.10%
-0.33
24.20%
24.20%
0.03
24.30%
24.30%
-0.02
24.34%
24.34%
0.02
21.19%
21.19%
0.38
21.99%
21.98%
0.93
22.90%
22.89%
1.12
23.61%
23.61%
0.68
23.89%
23.89%
0.39
24.12%
24.12%
0.17
24.23%
24.22%
0.09
24.30%
24.30%
0.05
21.39%
21.38%
0.65
22.10%
22.09%
1.05
23.06%
23.06%
0.72
23.73%
23.72%
0.37
23.98%
23.98%
0.19
24.17%
24.17%
0.10
24.27%
24.27%
0.04
24.34%
24.34%
0.02
21.86%
21.88%
-2.68
22.75%
22.79%
-3.97
23.66%
23.71%
-4.54
24.16%
24.19%
-2.54
24.27%
24.28%
-1.27
24.38%
24.39%
-0.58
24.42%
24.42%
-0.33
24.44%
24.44%
-0.17
22.14%
22.19%
-4.86
23.17%
23.24%
-7.12
24.14%
24.20%
-6.33
24.46%
24.50%
-3.62
24.53%
24.55%
-2.12
24.53%
24.54%
-0.95
24.54%
24.55%
-0.54
24.54%
24.54%
-0.29
22.44%
22.49%
-4.71
23.60%
23.65%
-4.57
24.38%
24.42%
-4.27
24.76%
24.78%
-1.71
24.74%
24.75%
-1.26
24.69%
24.70%
-0.59
24.65%
24.66%
-0.35
24.62%
24.62%
-0.19
compare the price errors (in Price bp4 ) for the same data, we observe in Table 3.3
that the error terms are not any bigger for short maturities but vary slightly over
time with two observed effects. The error term first increases over time as the error
between the proxy and the original model increases over time, as forecasted. But for
long maturities, presumably because the volatility converges to its stationary regime,
errors decrease. When we convert these prices to implied Black-Scholes volatilities,
these error terms are dramatically amplified for short maturities due to very small
vega. Finally, note that for fixed maturity, price errors are quite uniform w.r.t. strike
K.
Impact of the correlation. Analogous results for correlations equal to 20%, 20%
and 50% are reported in Tables 3.4-3.5, 3.6-3.7 and 3.8-3.9. We notice that the errors
are smaller for a correlation close to zero and become larger when the absolute
value of the correlation increases. However, for realistic correlation values (-50% for
instance), the accuracy for the usual maturities and strikes remains excellent (error
smaller than 20 bps), except for very extreme strikes.
Impact of the volatility of volatility. In view of Theorem 2.4, the smaller the volatility
of volatility, the more accurate the approximation. In the following numerical tests,
we increase , while maintaining Assumption (P). Thus, the new Hestons parameters
are = 10, = 1 and = 50%, the other parameters remaining unchanged. The
comparative results on implied volatilties and prices are presented in Table 3.10 and
3.11. As expected, the approximation is less accurate than for = 30%, but still
accurate enough to be efficiently used for fast calibration. The results for prices are
4 Error
price bp=
Price ApproximationTrue Price

Spot
10000
13
T 3.3
Put prices of the closed formula, of the approximation formula and related errors (in bp), expressed as a function
of maturities in fractions of years and relative strikes. Parameters: = 6%, = 3, = 30% and = 0%.
3M
6M
1Y
2Y
3Y
5Y
hal-00370717, version 1 - 24 Mar 2009
7Y
10Y
30.00
30.00
0.03
40.01
40.01
0.05
50.01
50.01
0.04
60.03
60.03
0.00
70.02
70.02
0.01
80.01
80.01
0.01
90.00
90.00
0.00
90.01
90.01
0.00
20.08
20.08
-0.11
30.07
30.08
-0.16
40.11
40.11
-0.21
50.20
50.20
-0.18
60.15
60.15
-0.09
70.15
70.15
-0.04
70.42
70.42
-0.04
80.23
80.23
-0.02
10.87
10.87
0.06
20.52
20.52
-0.18
21.84
21.84
0.06
32.08
32.08
-0.03
41.70
41.70
-0.08
43.80
43.80
0.01
53.15
53.15
-0.01
55.22
55.22
0.01
4.22
4.22
0.08
6.20
6.19
0.26
9.12
9.11
0.44
13.26
13.26
0.38
16.39
16.39
0.27
21.26
21.26
0.15
25.14
25.14
0.09
29.92
29.92
0.06
1.14
1.14
0.09
2.72
2.71
0.26
3.08
3.07
0.23
4.71
4.71
0.17
5.73
5.73
0.11
8.50
8.50
0.08
9.32
9.32
0.04
11.49
11.49
0.03
0.24
0.24
-0.15
0.40
0.40
-0.34
0.51
0.52
-0.51
0.79
0.79
-0.43
1.21
1.21
-0.31
1.61
1.61
-0.19
1.97
1.97
-0.14
2.62
2.62
-0.09
0.10
0.10
-0.14
0.14
0.14
-0.29
0.15
0.16
-0.29
0.28
0.29
-0.29
0.36
0.37
-0.22
0.58
0.58
-0.15
0.66
0.67
-0.10
0.84
0.84
-0.07
0.04
0.04
-0.07
0.05
0.05
-0.08
0.09
0.09
-0.12
0.11
0.11
-0.06
0.15
0.15
-0.07
0.21
0.21
-0.04
0.26
0.26
-0.03
0.33
0.33
-0.02
more satisfactory than for implied volatilities. Once again, for short maturities, the
errors in implied volatilities may be quite significant, except for options not-far-fromthe-money.
Impact of the assumption (P). The assumption (P) is a technical assumption that we
use to establish error estimates for the approximation formula (2.13). In the test that
follows, we relax this assumption by taking new parameters = 3%, = 2, = 40%
and = 0% for which the ratio 2/2 = 0.75 < 1. Results are reported in Tables 3.12
and 3.13. We observe that the approximation formula still works (errors are smaller
than 20 bps) but it is less accurate (compare with Table 3.2 for which the ratio 2/2
is equal to 4 > 1). An extension of the validity of our formula by relaxing Assumption
(P) is presumably relevant. This investigation is left for further research.
Piecewise constant parameters. Hestons constant parameters have been set to:
v0 = 4%, = 3. In addition, the piecewise constant functions , and are equal
respectively at each interval of the form ] 4i , i+1
4 [ to 4% + i 0.05%, 30% + i 0.5% and
20% + i 0.35%.
In the same Tables 3.15 and 3.16, we report values using three different formulas. For
a given maturity, the first row is obtained using the closed formula with piecewise
constant parameters (see appendix), the second row uses our approximation formula
(2.13) and the third row uses the closed formula with constant parameters computed
by averaging (see Section 2.6). In order to give complete information on our tests,
we also report in Table 3.14 the values used for the averaging parameters (following
Section 2.6).
Of course, the quickest approach is the use of the approximation formula (2.13).
As before, its accuracy is very good, except for very extreme strikes. It is quite
interesting is to observe that the averaging rules that we propose are extremely
14
T 3.4
and = 20%.
3M
6M
1Y
2Y
3Y
hal-00370717, version 1 - 24 Mar 2009
5Y
7Y
10Y
24.50%
24.04%
45.76
25.68%
25.19%
49.49
26.20%
25.92%
28.04
26.03%
25.95%
7.83
26.06%
25.95%
11.21
25.83%
25.75%
8.29
26.02%
25.82%
20.23
25.43%
25.40%
3.46
23.07%
23.14%
-7.65
24.38%
24.45%
-7.75
25.14%
25.23%
-8.22
25.28%
25.35%
-6.41
25.40%
25.44%
-3.39
25.28%
25.30%
-1.76
24.97%
24.99%
-1.59
24.99%
25.00%
-0.94
21.92%
21.93%
-1.25
23.31%
23.38%
-7.32
23.65%
23.68%
-2.65
24.29%
24.32%
-2.54
24.57%
24.60%
-2.44
24.47%
24.47%
-0.65
24.56%
24.57%
-0.59
24.49%
24.49%
-0.20
21.16%
21.15%
0.38
21.94%
21.93%
0.99
22.82%
22.81%
1.32
23.51%
23.50%
0.93
23.78%
23.78%
0.61
24.01%
24.01%
0.32
24.11%
24.11%
0.21
24.19%
24.18%
0.14
20.84%
20.82%
2.35
21.65%
21.63%
2.22
22.47%
22.44%
3.45
23.18%
23.16%
2.37
23.47%
23.45%
1.65
23.75%
23.74%
0.84
23.86%
23.85%
0.60
23.97%
23.96%
0.38
20.91%
20.87%
3.68
21.68%
21.64%
4.10
22.51%
22.49%
2.08
23.09%
23.08%
1.57
23.34%
23.32%
1.71
23.57%
23.56%
1.01
23.70%
23.69%
0.69
23.81%
23.80%
0.48
21.04%
21.06%
-2.73
21.88%
21.96%
-8.10
22.72%
22.89%
-16.41
23.17%
23.25%
-8.04
23.36%
23.41%
-5.11
23.55%
23.56%
-1.92
23.65%
23.67%
-1.50
23.75%
23.76%
-0.95
21.21%
21.37%
-16.51
22.15%
22.47%
-32.52
22.86%
23.17%
-31.56
23.29%
23.56%
-26.37
23.42%
23.58%
-16.68
23.55%
23.65%
-9.38
23.64%
23.70%
-6.16
23.72%
23.76%
-3.98
accurate.
Computational time. Regarding the computational time, the approximation formula (2.13) yields essentially the same computational cost as the Black-Scholes formula, while the closed formula requires an additional space integration involving
many exponential and trigonometric functions for which evaluation costs are higher.
For instance, using a 2, 6 GHz Pentium PC, the computations of the 64 numerical
values in Table 3.2 (3.4, 3.6 or 3.8) take 4.71 ms using the approximation formula and
301ms using the closed formula. For the example with time dependent coefficients
(reported in Table 3.15), the computational time for the 64 prices is about 40.2 ms
using the approximation formula and 2574 ms using the closed formula. Roughly
speaking, the use of the approximation formula enables us to speed up the valuation
(and thus the calibration) by a factor 100 to 600.
4. Proof of Theorem 2.4. The proof is divided into several steps. In Subsection
4.1 we give the upper bounds for derivatives of the put function PBS , in Subsection 4.2
the conditions for positivity of the squared volatility process v, in Subsection 4.3 the
RT
upper bounds for the negative moments of the integrated squared volatility 0 vt dt,
in Subsection 4.4 the upper bounds for derivatives of functionals of the squared
volatility process v. Finally, in Subsection 4.5, we complete the proof of Theorem 2.4
using the previous Subsections.
Notations. In order to alleviate the proofs, we introduce some notations specific
to this section.
Differentiation. For every process Z , we write (if these derivatives have a meaning):
(i) Zi,t =
i Zt
| ,
i =0
15
T 3.5
3M
6M
1Y
2Y
3Y
5Y
hal-00370717, version 1 - 24 Mar 2009
7Y
10Y
30.01
30.00
0.10
40.01
40.01
0.19
50.02
50.02
0.23
60.05
60.05
0.12
70.03
70.03
0.12
80.02
80.02
0.06
90.00
90.00
0.02
90.01
90.01
0.02
20.10
20.11
-0.21
30.10
30.10
-0.22
40.15
40.15
-0.32
50.25
50.25
-0.39
60.19
60.19
-0.17
70.18
70.18
-0.09
70.47
70.47
-0.17
80.26
80.26
-0.06
10.93
10.93
-0.15
20.60
20.60
-0.74
21.95
21.96
-0.60
32.21
32.21
-0.69
41.82
41.83
-0.62
43.91
43.91
-0.28
53.25
53.26
-0.24
55.30
55.30
-0.11
4.22
4.22
0.08
6.18
6.18
0.28
9.08
9.08
0.52
13.21
13.20
0.52
16.32
16.31
0.41
21.16
21.16
0.28
25.02
25.02
0.21
29.78
29.78
0.16
(ii) the ith Taylor residual by RZi,t = Zt
1.07
1.06
0.32
2.61
2.60
0.54
2.89
2.88
1.08
4.46
4.44
1.09
5.44
5.43
0.94
8.17
8.16
0.66
8.94
8.93
0.55
11.07
11.06
0.43
0.19
0.19
0.18
0.31
0.31
0.30
0.39
0.39
0.20
0.62
0.62
0.23
0.99
0.99
0.38
1.35
1.35
0.30
1.68
1.68
0.26
2.29
2.29
0.24
0.07
0.07
-0.06
0.10
0.10
-0.26
0.10
0.10
-0.57
0.19
0.20
-0.50
0.26
0.26
-0.42
0.44
0.44
-0.26
0.51
0.51
-0.24
0.66
0.67
-0.20
0.03
0.03
-0.18
0.03
0.03
-0.41
0.05
0.06
-0.64
0.06
0.07
-0.69
0.09
0.10
-0.63
0.14
0.14
-0.51
0.18
0.18
-0.44
0.24
0.24
-0.38
j
j=0 j! Z j,t .
Pi
Generic constants. We keep the same notation C for all non-negative constants
(i) depending on universal constants, on a number p 1 arising in Lp estimates,
on In f , v0 and K;
(ii) depending in a non decreasing way on ,
1
,
1||2Sup
Sup , Sup ,
Sup
In f
and T.
We write A = O(B) when |A| CB for a generic constant.

Miscellaneous.
p
(i) We write t = vt for the volatility for the perturbed process.
(ii) if (Z)t[0,T] is a càdlàg process, we denote by Z its running extremum:
Zt = sup |Zs |, t [0, T].

st
(iii) The Lp norm of a random variable is denoted, as usual, by kZkp = E[|Z|p ]1/p.
4.1. Upper bounds for put derivatives.
L 4.1. For every (i, j) N2 , there exists a polynomial P with positive coefficients
such that:

P( y)

i+j PBS
.
(x,
y)
sup

(2 j+i1)+
i j
xR x y
y 2
Proof. Note that it is enough to prove the estimates for j = 0, owing to the relation
(2.2). We now take j = 0. For i = 0, the inequality holds because PBS is bounded. Thus
16
T 3.6
and = 20%.
3M
6M
1Y
2Y
3Y
hal-00370717, version 1 - 24 Mar 2009
5Y
7Y
10Y
21.81%
22.41%
-59.86
22.75%
23.41%
-66.39
23.31%
23.83%
-52.67
23.53%
23.77%
-23.90
23.70%
23.93%
-23.06
23.81%
23.96%
-14.87
23.92%
24.21%
-28.79
23.94%
23.99%
-5.60
21.10%
21.11%
-1.80
22.05%
22.16%
-10.95
22.83%
22.91%
-8.05
23.33%
23.34%
-1.04
23.56%
23.58%
-1.95
23.76%
23.77%
-0.62
23.90%
23.89%
0.90
23.99%
23.99%
0.42
20.89%
20.87%
2.68
21.72%
21.66%
5.61
22.59%
22.55%
3.84
23.31%
23.28%
2.80
23.58%
23.56%
2.19
23.92%
23.91%
0.82
24.03%
24.02%
0.63
24.18%
24.18%
0.26
21.22%
21.22%
0.27
22.04%
22.03%
0.72
22.97%
22.96%
0.88
23.70%
23.70%
0.47
23.99%
23.99%
0.22
24.23%
24.23%
0.04
24.34%
24.34%
-0.01
24.42%
24.42%
-0.03
21.89%
21.90%
-0.82
22.53%
22.53%
-0.08
23.62%
23.64%
-1.65
24.25%
24.27%
-1.42
24.48%
24.49%
-1.15
24.59%
24.59%
-0.59
24.68%
24.68%
-0.48
24.70%
24.71%
-0.32
22.71%
22.78%
-7.12
23.71%
23.81%
-9.75
24.72%
24.82%
-9.85
25.16%
25.22%
-6.19
25.15%
25.19%
-3.93
25.15%
25.17%
-2.06
25.12%
25.13%
-1.30
25.06%
25.07%
-0.79
23.13%
23.20%
-7.19
24.31%
24.40%
-8.77
25.41%
25.46%
-4.37
25.65%
25.68%
-3.19
25.63%
25.64%
-1.70
25.46%
25.47%
-0.94
25.39%
25.39%
-0.30
25.29%
25.29%
-0.07
23.54%
23.55%
-1.20
24.47%
24.45%
2.21
24.80%
24.81%
-1.19
24.93%
24.93%
-0.67
24.83%
24.84%
-0.92
24.74%
24.74%
-0.49
24.68%
24.68%
-0.30
24.63%
24.64%
-0.17
consider i 1. Then by differentiating the payoff, one gets:

RT
T
y
i PBS
x+ 0 (rt qt )dt 2 +
0 rt dt
i
(x,
y)
=
E[e
(K
e
x
xi
R
= xi1 E[1
(e
q
RT
y
y
W
x+ 0 (rt qt )dt +
2
T T K)
y
T WT
)+ ]
q
RT
y
y
x 0 qt dt 2 + T WT
= xi1 E[(x + G)]
where is a bounded function (by K) and G is a Gaussian variable with zero mean and
R
(zx)2 /(2y)
variance equal to y. For such a function, we write E[(x + G)] = R (z) e
dz
2y
and from this, it follows by a direct computation that

C
xi1 E[(x + G)] i1
y 2
for any x and y. We have proved the estimate for j = 0 and i 1.
4.2. Positivity of the squared volatility process v. For a complete review related to time homogeneous CIR processes, we refer the reader to [GJY03]. For time
dependent CIR process, see [Mag96] where the existence and representation using
squared Bessel processes are provided.
To prove the positivity of the process v, we show that it can be bounded from
below by a suitable time homogeneous CIR process, time scale being the only difference (see definition 5.1.2. in [RY99]). The arguments are quite standard, but since we
17
T 3.7
3M
6M
1Y
2Y
3Y
5Y
hal-00370717, version 1 - 24 Mar 2009
7Y
10Y
30.00
30.00
-0.05
40.00
40.00
-0.11
50.01
50.01
-0.20
60.02
60.02
-0.20
70.01
70.01
-0.12
80.01
80.01
-0.06
90.00
90.00
-0.01
90.01
90.01
-0.02
20.06
20.06
-0.03
30.05
30.05
-0.19
40.08
40.08
-0.20
50.15
50.15
-0.04
60.11
60.11
-0.07
70.11
70.11
-0.02
70.36
70.36
0.08
80.20
80.19
0.02
10.81
10.81
0.30
20.45
20.44
0.49
21.72
21.71
0.83
31.94
31.94
0.73
41.58
41.57
0.53
43.68
43.67
0.35
53.04
53.04
0.25
55.13
55.13
0.15
4.23
4.23
0.05
6.21
6.21
0.20
9.14
9.14
0.35
13.31
13.31
0.26
16.46
16.46
0.15
21.36
21.35
0.03
25.25
25.25
-0.01
30.06
30.06
-0.04
1.22
1.22
-0.12
2.82
2.82
-0.02
3.26
3.26
-0.53
4.96
4.96
-0.67
6.02
6.03
-0.67
8.83
8.84
-0.48
9.69
9.70
-0.45
11.91
11.91
-0.37
0.28
0.29
-0.43
0.48
0.49
-0.92
0.64
0.65
-1.25
0.97
0.98
-1.18
1.44
1.45
-1.06
1.87
1.88
-0.74
2.26
2.27
-0.57
2.96
2.96
-0.45
0.13
0.13
-0.25
0.19
0.20
-0.43
0.22
0.22
-0.26
0.39
0.39
-0.32
0.49
0.49
-0.21
0.75
0.75
-0.18
0.84
0.84
-0.07
1.04
1.04
-0.02
0.06
0.06
-0.02
0.07
0.07
0.05
0.10
0.10
-0.04
0.11
0.11
-0.03
0.16
0.16
-0.05
0.21
0.21
-0.04
0.26
0.26
-0.03
0.34
0.34
-0.02
need a specific statement that is not available in the literature, we detail the result
and its proof. The time change t 7 At is defined by
Z At
t=
2s ds.
0
Because In f > 0, A is a continuous, strictly increasing time change and its inverse
A1 enjoys the same properties.
L 4.2. Assume (P) and v0 > 0. Denote by (ys )0sA1 the CIR process defined by
T
1
dyt = ( 2 yt )dt + yt dB t , y0 = v0 ,
2
In f
where B is the Brownian motion given by

AT
Z
B t =
s dBs .
(4.1)
Then, a.s. one has vt yA1 for any t [0, T]. In particular, (vt )0tT is a.s. positive.
t
Proof. Note that (B t )0tA1 is really a Brownian motion because by Levys CharacT
Bi
t = t (see Proposition
terization Theorem, it is a continuous local martingale with hB,
5.1.5 [RY99] for the computation of the bracket). Now that we have set v t = vAt , our
aim is to prove that v t yt for t [0, A1
]. Using Propositions 5.1.4 and 5.1.5 [RY99],
T
we write
Z t
Z At
( 2 (As v s)ds + v s dB s ).
((s vs )ds + s vs dBs ) = v0 +
v t = v0 +
0 A
0
s
18
T 3.8
and = 50%.
3M
6M
1Y
2Y
3Y
hal-00370717, version 1 - 24 Mar 2009
5Y
7Y
10Y
26.13%
25.57%
56.55
27.47%
26.89%
58.13
27.96%
27.67%
29.08
27.56%
27.52%
4.11
27.53%
27.42%
11.28
27.11%
27.01%
9.64
27.35%
27.03%
31.65
26.40%
26.36%
4.15
24.29%
24.43%
-14.06
25.81%
25.97%
-16.68
26.57%
26.75%
-18.08
26.51%
26.65%
-14.03
26.56%
26.66%
-9.11
26.25%
26.31%
-5.22
25.67%
25.71%
-3.57
25.66%
25.68%
-2.43
22.60%
22.63%
-2.51
24.31%
24.44%
-12.19
24.34%
24.39%
-5.01
24.93%
24.98%
-4.75
25.22%
25.26%
-4.59
24.83%
24.84%
-1.23
24.92%
24.93%
-1.09
24.70%
24.70%
-0.35
21.11%
21.11%
0.19
21.85%
21.84%
0.82
22.68%
22.66%
1.53
23.34%
23.33%
1.43
23.61%
23.60%
1.06
23.83%
23.82%
0.62
23.93%
23.93%
0.43
24.01%
24.00%
0.29
19.95%
19.90%
4.35
20.92%
20.89%
3.38
21.51%
21.43%
7.49
22.31%
22.25%
5.50
22.66%
22.62%
3.97
23.10%
23.08%
1.98
23.23%
23.21%
1.46
23.40%
23.39%
0.93
19.22%
18.99%
23.24
19.80%
19.50%
29.46
20.49%
20.24%
24.84
21.30%
21.15%
14.43
21.81%
21.71%
9.79
22.28%
22.23%
5.14
22.55%
22.52%
3.26
22.82%
22.80%
2.02
19.03%
18.91%
11.67
19.55%
19.61%
-5.28
20.19%
20.77%
-58.18
20.95%
21.19%
-23.17
21.39%
21.53%
-14.43
21.94%
21.98%
-4.04
22.22%
22.25%
-3.91
22.50%
22.53%
-2.65
18.92%
19.57%
-64.22
19.47%
21.11%
-164.16
20.11%
21.73%
-162.76
20.73%
22.20%
-146.81
21.16%
22.04%
-88.86
21.66%
22.14%
-47.56
21.98%
22.28%
-30.07
22.29%
22.48%
-18.89
Now we apply a comparison result for SDEs twice (see Proposition 5.2.18 in [KS88]).
1. First, one gets vt nt , where (ns )s is the (unique) solution of
Z t
2 ns ds + ns dB s ,
nt = 0 +
A
0
s
because v0 0 and
2A
(As x) 2
A
x, for all x R and s [0, A1

]. Of course nt = 0,
T
thus vt is non-negative.
we only need to compare drift
2. Secondly, using the non-negativity of v,
coefficients for the non-negative variable x. Under (P), since
1
(As x) 2 x x 0, s [0, A1
T ],
2
2
A
In f
s
we obtain v t yt for t [0, A1

] a.s.
T
Moreover, the positivity of y (and consequently that of v) is standard: indeed, y is
a 2-dimensional squared Bessel process with a time/space scale change (see [GJY03],
or the proof of Lemma 4.3 below).
4.3. Upper bound for negative moments of the integrated squared volatility
RT
process 0 vt dt.
L 4.3. Assume (P). Then for every p > 0, one has:
Z T
C
vt dt)p ] p .
sup E[(
T
0
01
19
T 3.9
3M
6M
1Y
2Y
3Y
5Y
hal-00370717, version 1 - 24 Mar 2009
7Y
10Y
30.01
30.01
0.21
40.02
40.02
0.37
50.04
50.04
0.36
60.08
60.08
0.09
70.05
70.05
0.17
80.03
80.03
0.11
90.00
90.00
0.06
90.02
90.02
0.03
20.14
20.15
-0.47
30.15
30.15
-0.59
40.21
40.22
-0.88
50.33
50.34
-1.00
60.25
60.25
-0.54
70.23
70.23
-0.30
70.54
70.55
-0.41
80.30
80.30
-0.18
11.01
11.02
-0.31
20.70
20.71
-1.33
22.11
22.12
-1.17
32.38
32.39
-1.32
41.99
42.00
-1.21
44.06
44.07
-0.53
53.40
53.40
-0.44
55.42
55.42
-0.20
4.21
4.21
0.04
6.16
6.15
0.23
9.03
9.02
0.61
13.11
13.10
0.80
16.20
16.19
0.72
21.01
21.00
0.54
24.84
24.84
0.43
29.57
29.57
0.34
0.95
0.94
0.57
2.43
2.42
0.81
2.59
2.57
2.27
4.06
4.03
2.47
4.98
4.96
2.20
7.65
7.64
1.54
8.36
8.35
1.32
10.43
10.42
1.04
0.12
0.11
0.82
0.19
0.17
1.59
0.22
0.21
1.67
0.39
0.37
1.59
0.69
0.67
1.73
0.99
0.98
1.29
1.28
1.27
1.04
1.82
1.81
0.89
0.03
0.03
0.16
0.04
0.04
-0.09
0.03
0.05
-1.05
0.08
0.09
-0.84
0.13
0.13
-0.74
0.25
0.26
-0.38
0.31
0.32
-0.45
0.44
0.44
-0.42
0.01
0.01
-0.36
0.01
0.02
-1.05
0.01
0.03
-1.69
0.02
0.04
-2.00
0.03
0.05
-1.80
0.06
0.07
-1.50
0.09
0.10
-1.32
0.13
0.14
-1.17
Before proving the result, we mention that analogous estimates appear in [BD07]
(Lemmas A.1 and A.2): some exponential moments are stated under stronger conditions than those in assumption (P). In addition, the uniformity of the estimates w.r.t.
(or equivalently w.r.t. ) is not emphasized. In our study, it is crucial to get uniform
estimates w.r.t. .
Proof. Fix p 21 (for 0 < p < 12 , we derive the result from the case p = 21 using the
Holder
inequality). The proof is divided into two steps. We first prove the estimates
in the case of constant coefficients , , with = 21 , = 1 and = 1. Then, using
the time change of Lemma 4.2, we derive the result for (vt )t . The critical point is to
get estimates that are uniform w.r.t. .
Step 1. Take t , t 1, = 12 , = 1 and consider
1
dyt = ( yt )dt + yt dBt , y0 = v0 ,
2
for a standard Brownian motion B. We represent y as a time space transformed
squared Bessel process (see [GJY03])
yt = et z (et 1)
4
where z is a 2-dimensional squared Bessel process. Therefore, using a change of

variable and the explicit expression of Laplace transform for the integral of z (see
20
T 3.10
expressed as a function of maturities in fractions of years and relative strikes. Parameters: = 6%, = 10, = 1
and = 50%.
3M
31.51%
30.68%
82.46
31.45%
30.83%
62.40
30.09%
29.87%
21.52
28.45%
28.46%
-0.53
28.08%
27.98%
9.78
27.40%
27.31%
9.15
27.56%
27.24%
32.00
26.53%
26.49%
4.02
6M
1Y
2Y
3Y
hal-00370717, version 1 - 24 Mar 2009
5Y
7Y
10Y
28.04%
28.99%
-94.66
28.86%
29.59%
-73.58
28.30%
28.72%
-42.32
27.27%
27.47%
-20.08
27.05%
27.16%
-11.59
26.52%
26.58%
-5.98
25.84%
25.88%
-3.83
25.77%
25.80%
-2.57
24.74%
24.95%
-21.22
26.52%
26.98%
-46.52
25.44%
25.54%
-10.69
25.51%
25.57%
-6.39
25.61%
25.66%
-5.41
25.04%
25.05%
-1.31
25.06%
25.08%
-1.14
24.80%
24.80%
-0.36
21.83%
21.71%
12.10
22.69%
22.58%
11.30
23.34%
23.28%
6.02
23.73%
23.71%
2.42
23.88%
23.86%
1.39
24.00%
23.99%
0.71
24.05%
24.05%
0.47
24.09%
24.09%
0.31
19.94%
19.38%
56.44
21.36%
21.09%
26.99
21.89%
21.70%
19.45
22.58%
22.50%
8.11
22.86%
22.81%
4.91
23.23%
23.21%
2.20
23.33%
23.31%
1.57
23.47%
23.46%
0.97
19.45%
18.05%
140.23
20.11%
19.14%
97.22
20.76%
20.30%
46.13
21.48%
21.28%
19.97
21.96%
21.83%
12.13
22.38%
22.33%
5.85
22.63%
22.59%
3.57
22.88%
22.86%
2.15
19.58%
19.76%
-18.10
20.05%
20.64%
-59.12
20.49%
21.65%
-115.72
21.12%
21.42%
-30.34
21.51%
21.67%
-16.04
22.03%
22.07%
-3.93
22.29%
22.33%
-3.88
22.55%
22.58%
-2.64
19.85%
22.93%
-308.17
20.20%
24.03%
-383.12
20.45%
23.17%
-271.22
20.90%
22.75%
-184.76
21.27%
22.30%
-102.46
21.75%
22.26%
-51.20
22.05%
22.36%
-31.56
22.34%
22.53%
-19.49
[BS02] p.377), one obtains for any u 0

T
Z
2T
(eT 1)
4
yt dt)] E[exp(4ue
zs ds)]
0
2u(1 eT ) 1
2u(1 eT )
T
) exp( 2ue v0 tanh(
)).
cosh(
2
2
R
RT
1
up1 eux du for x = 0 yt dt, one gets:
Combining this with the identity xp = (p)
0
E[exp(u
Z
E[(
0
1
yt dt) ]
(p)
p1
cosh(
2u(1 eT ) 1
) exp( 2ueT v0 tanh(
2
2u(1 eT )
))du.
2
(eT 1)
2u(1eT )
= v0 eT 2 2u.
Define the parameter 2 = 2v and the new variable n =
2
0
It readily follows that
Z T
Z
tanh(n)n
eT
yt dt)p ] C( 2 )2p
E[(
n2p1 cosh(n)1 exp(
)dn,
2
0
0
where C is a constant depending only on v0 and p. We upper bound the above integral
differently according to the value of .
(i) If 1, then
Z
Z T
eT 2p 2p1
p
n
cosh(n)1 dn Ce2pT .
(4.2)
yt dt) ] C( 2 )
E[(
0
0
21
T 3.11
of maturities in fractions of years and relative strikes. Parameters: = 6%, = 10, = 1 and = 50%.
3M
6M
1Y
2Y
3Y
5Y
hal-00370717, version 1 - 24 Mar 2009
7Y
10Y
30.05
30.04
0.99
40.06
40.05
0.92
50.08
50.07
0.41
60.10
60.10
-0.01
70.06
70.05
0.17
80.04
80.03
0.11
90.00
90.00
0.06
90.02
90.02
0.03
20.30
20.35
-4.95
30.28
30.32
-3.90
40.31
40.33
-2.60
50.39
50.40
-1.58
60.28
60.28
-0.73
70.25
70.25
-0.36
70.56
70.57
-0.44
80.31
80.31
-0.19
11.28
11.31
-2.80
20.96
21.02
-5.83
22.37
22.40
-2.58
32.54
32.56
-1.82
42.09
42.11
-1.46
44.15
44.16
-0.57
53.46
53.46
-0.47
55.47
55.48
-0.20
4.35
4.33
2.41
6.40
6.36
3.18
9.29
9.26
2.39
13.33
13.31
1.35
16.38
16.37
0.94
21.15
21.15
0.61
24.96
24.96
0.47
29.67
29.67
0.36
0.95
0.87
7.37
2.54
2.47
6.51
2.71
2.65
5.95
4.18
4.14
3.67
5.09
5.06
2.74
7.76
7.74
1.72
8.45
8.44
1.42
10.51
10.50
1.09
0.13
0.08
4.62
0.21
0.15
5.23
0.24
0.21
3.19
0.41
0.38
2.26
0.71
0.69
2.19
1.02
1.01
1.49
1.31
1.30
1.16
1.85
1.84
0.95
0.04
0.05
-0.31
0.05
0.06
-1.28
0.04
0.07
-2.52
0.09
0.10
-1.17
0.13
0.14
-0.86
0.26
0.27
-0.38
0.32
0.32
-0.46
0.44
0.45
-0.42
0.01
0.05
-3.51
0.01
0.06
-4.72
0.02
0.06
-3.89
0.02
0.05
-2.89
0.03
0.06
-2.22
0.06
0.08
-1.68
0.09
0.10
-1.42
0.13
0.14
-1.23
(ii) If 1, split the integral into two parts, n arctanh() and n arctanh().
For the first part, simply use n tanh(n) for any n. For the second part, use tanh(n)
and cosh(n)1 1. This gives
Z T
T Z arctanh()
tanh2 (n)
e
yt dt)p ] C ( 2 )2p
E[(
n2p1 cosh(n)1 exp(
)dn
2
0
0
Z

n
eT
n2p1 exp( )dn := C[T1 + T2 ].
+ ( 2 )2p
(4.3)
arctanh()
We upper bound the two terms separately.
tanh(n)
1. First term T1 . Using the change of variable m = , one has:
Z 1
T1 e2pT 4p+1
arctanh(m)2p1 cosh(arctanh(m)) exp(m2)dm.
0
Because of 1, we have the following inequalities for m [0, 1[:

arctanh(m) arctanh(m),
cosh(arctanh(m)) cosh(arctanh(m)).
Using 2p 1 0, it readily follows that

Z 1
e2T
arctanh(m)2p1 cosh(arctanh(m)) exp(m2 )dm.
T1 ( 2 )p
0
2. Second term T2 . Clearly, we have
Z
Z
eT 2p 2p1
n
e2T p 2p1 v
T2 ( 2 )
n
exp( )dn = ( 2 )
v
e dv.
0
0
(4.4)
(4.5)
22
T 3.12
and = 0%.
3M
6M
1Y
2Y
3Y
hal-00370717, version 1 - 24 Mar 2009
5Y
7Y
10Y
23.27%
22.35%
92.35
24.10%
22.52%
158.79
23.96%
22.20%
175.41
22.72%
21.40%
132.35
22.44%
20.74%
170.16
21.56%
20.03%
153.81
21.93%
19.51%
241.42
20.21%
19.24%
96.63
21.25%
21.48%
-22.93
22.05%
22.26%
-20.74
22.01%
22.14%
-12.81
21.05%
21.20%
-14.49
20.84%
20.67%
16.92
20.09%
19.88%
20.49
19.01%
19.09%
-7.47
18.92%
18.88%
4.46
19.59%
19.56%
2.23
20.22%
20.50%
-28.41
18.89%
18.99%
-10.17
18.61%
18.83%
-22.09
18.66%
18.93%
-27.04
17.86%
17.92%
-5.89
17.88%
17.95%
-6.53
17.58%
17.60%
-1.61
18.86%
18.85%
1.62
18.21%
18.14%
7.08
17.60%
17.45%
14.90
17.24%
17.10%
14.34
17.16%
17.06%
10.16
17.16%
17.10%
5.27
17.17%
17.14%
3.16
17.20%
17.18%
1.76
Combining (4.3), (4.4) and (4.5), we obtain E[(

inequality (ex 1 x, x 0), we have 2 =
E[(
0
19.47%
19.43%
3.51
18.68%
18.59%
9.08
18.51%
18.48%
2.60
18.04%
18.06%
-1.35
17.88%
17.91%
-3.17
17.61%
17.62%
-0.54
17.60%
17.62%
-1.53
17.53%
17.54%
-0.93
RT
(eT 1)
2v0
yt dt)p ] C
20.64%
20.83%
-18.90
20.75%
21.16%
-40.69
20.84%
21.42%
-57.41
20.26%
20.72%
-46.40
19.60%
19.96%
-36.03
19.08%
19.28%
-19.86
18.76%
18.88%
-12.47
18.42%
18.49%
-7.64
21.25%
21.48%
-22.93
21.78%
22.10%
-32.03
22.23%
22.20%
2.27
21.42%
21.32%
9.96
20.84%
20.67%
16.92
19.94%
19.83%
11.43
19.54%
19.39%
14.41
19.09%
18.97%
11.97
21.85%
21.94%
-9.79
22.72%
22.51%
20.96
22.85%
22.30%
54.89
22.42%
21.42%
100.04
21.67%
20.79%
87.99
20.75%
20.03%
72.25
20.16%
19.58%
58.41
19.61%
19.16%
44.80
2T
yt dt)p ] C( e2 )p . In view of the

T
2v0 ,
which gives
e2pT
,
Tp
(4.6)
available when 1.
To sum up (4.2) and (4.6), we have proved that
Z
E[(
0
yt dt)p ] Ce2pT (1 +
1
),
Tp
(4.7)
for a constant C depending only on p and v0 .

Step 2. Take ]0, 1]. We apply Lemma 4.2 to v , in order to write vt y 1 where
A,t
RT
R A,t
p
1
y dB , y = y0 . Thus, we get
v dt
t=
(s ) ds and dy = (
2 y )dt +
0
(In f )
23
T 3.13
3M
30.00
30.00
0.12
40.01
40.00
0.29
50.01
50.00
0.55
60.01
60.01
0.63
70.01
70.00
0.36
80.00
80.00
0.15
90.00
90.00
0.01
90.00
90.00
0.04
6M
1Y
2Y
3Y
5Y
hal-00370717, version 1 - 24 Mar 2009
7Y
10Y
20.06
20.07
-0.46
30.05
30.05
-0.37
40.06
40.06
-0.27
50.07
50.07
-0.39
60.04
60.04
0.28
70.03
70.02
0.26
70.07
70.08
-0.24
80.03
80.03
0.06
10.67
10.67
0.24
20.32
20.35
-2.15
20.98
21.00
-1.80
30.89
30.93
-4.17
40.60
40.64
-4.17
41.45
41.47
-1.76
51.04
51.06
-1.63
51.95
51.96
-0.64
3.76
3.76
0.32
5.13
5.11
1.99
7.01
6.95
5.92
9.70
9.62
8.03
11.82
11.75
6.94
15.21
15.16
4.61
17.97
17.94
3.25
21.43
21.41
2.14
0.88
0.88
0.45
1.91
1.89
2.06
1.73
1.73
0.70
2.28
2.29
-0.51
2.56
2.58
-1.42
3.75
3.75
-0.34
3.78
3.80
-1.07
4.55
4.56
-0.80
0.18
0.18
-0.88
0.25
0.27
-2.72
0.25
0.29
-4.43
0.28
0.33
-4.44
0.36
0.41
-4.48
0.38
0.40
-2.79
0.39
0.41
-1.94
0.46
0.48
-1.44
0.08
0.08
-0.58
0.09
0.10
-1.01
0.08
0.08
0.07
0.10
0.10
0.40
0.10
0.09
0.71
0.11
0.11
0.58
0.10
0.09
0.70
0.10
0.09
0.61
T 3.14
Equivalent averaged parameters.
R A1
,t
0
v0
3M
6M
1Y
2Y
3Y
5Y
7Y
10Y
4%
3.97 %
3.28 %
4.64 %
56.24 %
28.58 %
84.92 %
14.54 %
4%
4.04 %
4.38 %
4.02 %
4.04 %
2.68 %
0.59 %
4.57 %
30 %
30.12 %
30.89 %
31.12 %
32.10 %
33.63 %
35.41 %
39.98 %
-20 %
-19.93 %
-19.72 %
-18.95 %
-18.20 %
-16.52 %
-14.80 %
-12.32 %
ys ds)/(Sup )2 and in view of (4.7), it follows that

Z
E[(
0
A1
,T
Z
vt dt)p ] (Sup )2p E(
0
2p
C(Sup )2p e
2
Sup
2p 2
Ce
In f
ys ds)p
A1
(In f )2 ,T
2p
2p
(Sup +
where we have used 2 2In f T A1

2 2Sup T.
,T
(1 +
Sup 1
)
p
2p
T
In f
1
[A1
]p
,T
0.03
0.03
-0.13
0.04
0.03
0.29
0.05
0.04
0.98
0.04
0.03
1.61
0.04
0.03
1.54
0.04
0.02
1.24
0.03
0.02
1.00
0.03
0.02
0.76
24
T 3.15
Implied Black-Scholes volatilities of the closed formula, of the approximation formula and of the averaging
formula, expressed as a function of maturities in fractions of years and relative strikes. Piecewise constant parameters.
3M
23.45%
22.73%
23.45%
24.09%
23.09%
24.09%
23.95%
23.12%
23.95%
23.26%
22.84%
23.26%
23.28%
22.81%
23.28%
23.22%
22.88%
23.22%
23.86%
23.25%
23.86%
23.59%
23.46%
23.59%
6M
1Y
2Y
3Y
5Y
hal-00370717, version 1 - 24 Mar 2009
7Y
10Y
21.88%
21.96%
21.88%
22.59%
22.60%
22.59%
22.66%
22.66%
22.66%
22.30%
22.33%
22.30%
22.40%
22.38%
22.40%
22.46%
22.44%
22.46%
22.36%
22.39%
22.37%
22.96%
22.98%
22.96%
20.58%
20.60%
20.58%
21.30%
21.43%
21.30%
20.76%
20.81%
20.76%
21.01%
21.04%
21.01%
21.27%
21.33%
21.27%
21.34%
21.35%
21.34%
21.81%
21.82%
21.81%
22.30%
22.30%
22.30%
19.70%
19.69%
19.70%
19.63%
19.61%
19.63%
19.70%
19.68%
19.70%
19.99%
19.96%
19.98%
20.26%
20.24%
20.26%
20.77%
20.77%
20.77%
21.26%
21.26%
21.26%
21.97%
21.97%
21.97%
19.39%
19.35%
19.39%
19.33%
19.30%
19.33%
19.37%
19.32%
19.37%
19.66%
19.62%
19.66%
19.96%
19.93%
19.96%
20.54%
20.52%
20.54%
21.06%
21.05%
21.06%
21.82%
21.81%
21.82%
19.55%
19.53%
19.55%
19.58%
19.58%
19.58%
19.69%
19.78%
19.69%
19.83%
19.90%
19.83%
20.02%
20.04%
20.02%
20.54%
20.55%
20.54%
21.06%
21.07%
21.06%
21.83%
21.84%
21.83%
19.74%
19.84%
19.74%
19.92%
20.19%
19.92%
20.12%
20.62%
20.12%
20.09%
20.43%
20.09%
20.23%
20.47%
20.23%
20.65%
20.76%
20.64%
21.16%
21.23%
21.15%
21.92%
21.96%
21.92%
19.97%
20.28%
19.97%
20.31%
20.93%
20.31%
20.36%
21.05%
20.35%
20.37%
21.02%
20.37%
20.43%
20.90%
20.42%
20.80%
21.09%
20.79%
21.27%
21.45%
21.26%
22.02%
22.12%
22.01%
Note that the upper bound does not depend on ]0, 1]. For = 0, the upper
bound in Lemma 4.3 is also true because (v0t )t is deterministic and
max(v0 , Sup ) v0t min(v0 , In f ) > 0.
(4.8)
4.4. Upper bound for residuals of the Taylor development of g() defined in
(1.4). Throughout the following paragraph, we assume that (P) is in force. We define
the variables:
T
Z
PT =
Z
t (t 0,t )dBt
2t
2
Z
(vt v0,t )dt,
QT =
(1 2t )(vt v0,t )dt.
R T 2
RT
RT
RT q
Notice that (x0 + 0 t v0,t dBt 0 2t v0,t dt, 0 (12t )v0,t dt)+(P1T, Q1T ) = (x0 + 0 t v1t dBt
RT
R T 2
t 1
v dt, 0 (1 2t )v1t dt).
0 2 t
The main result of this subsection is the following proposition, the statement of which
uses the notation introduced at the beginning of Section 4.
25
T 3.16
Put prices of the closed formula, of the approximation formula and of the averaging formula, expressed as a
function of maturities in fractions of years and relative strikes. Piecewise constant parameters.
3M
6M
1Y
2Y
3Y
5Y
hal-00370717, version 1 - 24 Mar 2009
7Y
10Y
30.00
30.00
30.00
40.01
40.00
40.01
50.01
50.01
50.01
60.02
60.01
60.02
70.01
70.01
70.01
80.01
80.01
80.01
90.00
90.00
90.00
90.01
90.01
90.01
20.07
20.08
20.07
30.06
30.06
30.06
40.07
40.07
40.07
50.11
50.11
50.11
60.07
60.07
60.07
70.07
70.07
70.07
70.24
70.24
70.24
80.14
80.14
80.14
10.78
10.78
10.78
20.41
20.42
20.41
21.33
21.35
21.33
31.38
31.39
31.38
41.07
41.08
41.07
42.64
42.64
42.64
52.22
52.22
52.22
54.13
54.13
54.13
3.93
3.93
3.93
5.53
5.53
5.53
7.85
7.84
7.85
11.23
11.23
11.23
13.92
13.92
13.92
18.37
18.36
18.37
22.15
22.15
22.15
27.17
27.16
27.17
0.87
0.87
0.87
2.06
2.05
2.06
1.97
1.95
1.97
2.92
2.90
2.92
3.55
3.54
3.55
5.74
5.72
5.74
6.46
6.45
6.46
8.71
8.70
8.71
0.13
0.13
0.13
0.18
0.18
0.18
0.17
0.18
0.17
0.24
0.25
0.24
0.41
0.42
0.41
0.61
0.61
0.61
0.86
0.86
0.86
1.42
1.42
1.42
0.05
0.05
0.05
0.05
0.05
0.05
0.03
0.04
0.03
0.06
0.07
0.06
0.08
0.09
0.08
0.15
0.16
0.15
0.21
0.21
0.21
0.35
0.36
0.35
0.02
0.02
0.02
0.01
0.02
0.01
0.02
0.02
0.02
0.01
0.02
0.01
0.02
0.03
0.02
0.04
0.04
0.04
0.06
0.07
0.06
0.11
0.12
0.11
P 4.4. One has the following estimates for every p 1

kP1T kp C(Sup T) T,
1
kRP2,T kp C(Sup T)3 T,
(P1 )2
kR2,T kp C(Sup T)3 T,
kQ1T kp C(Sup T)T,
Q1
kR2,T kp C(Sup T)3 T,
(Q1 )2
kR2,T kp C(Sup T)3 T2 ,
3
P 1 Q1
kR2,T kp C(Sup T)3 T 2 .
To estimate the derivatives and the residuals for the variables PT and QT , we need
first to
the existence of the derivatives and the residuals of the volatility process
pprove
t = vt and its square v . Finally we prove Proposition 4.4.

4.4.1. Upper bounds for derivatives of and v . Under assumption (P), the
volatility process t is governed by the SDE:
dt = ((
2 2
t
t t 1
) )dt +
dBt , 0 = v0 ,
2
8 t 2 t
2
(4.9)
where we have used Itos Lemma and positivity of vt (see Lemma 4.2).
In order to estimate R0,t , we are going to prove that it verifies a linear equation
26
(Lemma 4.5) from which we deduce an a priori upper bound (Proposition 4.6). We
iterate the same analysis for the residuals R1,t (Proposition 4.7) and R2,t (Proposition
4.8). Analogously, we give upper bounds for the residuals of vt (Proposition 4.9).
L 4.5. Under (P), the process (R0,t = t 0t )0tT is given by
R0,t
Z
=
Ut
(Us )1 (
22s
s
ds +
dBs ),
80,s
2
where
dUt = t Ut dt, Ut = 1,
t = (
2 2
t t
1
)
+ .
2
8 t 0,t 2
hal-00370717, version 1 - 24 Mar 2009
Proof. From the definition (0,t )t = (0t )t and the equation (4.9), one obtains the
SDE
d0,t = (
t
0,t )dt, 0,0 = v0 .
20,t 2
Substitute this equation in (4.9) to obtain
dR0,t = t R0,t dt
2 2t
80,t
dt +
dBt , R0,0 = 0.
2
(4.10)
Note that R0,. is the solution of a linear SDE. Hence, it can be explicitly represented
using the process U (see Th. 52 in [Pro90]):
R0,t
Z
=
Ut
(Us )1 (
22s
s
ds +
dBs ).
80,s
2
P 4.6. Under (P), for every p 1 one has
k(R0,. )t kp CSup t.
In particular, the application t is continuous5 at = 0 in Lp .

Proof. At first sight, the proof seems to be straightforward from Lemma 4.5. But
actually, the difficulty lies in the fact that one can not uniformly in upper bound Ut
in Lp (because of the term with 1/t in t ).
Rt
Using Lemma 4.5 and Itos formula for the product (Ut)1 ( 0 2 s dBs ), one has
R0,t
Z
=
Ut
(Us )1 (
2 2
s
80,s
Z
ds) +
0
s
dBs Ut
2
Z
(
u
dBu )d(Us)1 .
2
5 Note that from the upper bound (4.11) in the proof, we easily obtain that the continuity also holds
a.s., and not only in Lp . Since only the latter is needed in what follows, we do not go into detail.
27
Under (P), one has t /2 > 0, which implies that t 7 Ut is decreasing and t 7 (Ut)1
is increasing. Thus, 0 Ut (Us)1 1 for s [0, t]. Consequently, we deduce
Z
|R0,t |
2 2s
ds + (
80,s
2 2s
80,s
ds + (
0
s
dBs )t + (
2
Z
0
s
dBs )t (1 Ut)
2
s dBs )t .
(4.11)
Now we easily complete the proof by observing that 0,s min( In f , v0 ) and
R.
k( 0 s dBs )t kp CSup t.
We define
Z t
s
0
(Us0 )1 dBs .
1,t = Ut
2
0
hal-00370717, version 1 - 24 Mar 2009
Therefore, (1,t )0tT solves the following SDE:

d1,t = (
t
t
+ )1,t dt + dBt , 1,0 = 0,

2
2(0,t )2 2
and for every p 1
k(1,. )t kp CSup t.
(4.12)
(4.13)
P 4.7. Under (P), the process (R1,t = t 0t 1,t )0tT fulfills the equality:
R1,t
Z
=
Ut
(Us)1 (
2 2s
2 2s
+ 1,s (( s
)R0,s +
))ds.
80,s
0,s 20,s
80,s
Moreover, for every p 1, one has
k(R1,. )t kp C(Sup t)2 .
In particular, the application t is C1 at = 0 in Lp sense with the first derivative at = 0
equal to 1,t (justifying a posteriori the definition R1,. ).

Proof. From Equations (4.10) and (4.12), it readily follows that
dR1,t = t R1,t dt 1,t(t
2 2t
t
)dt
dt, R1,0 = 0.
80,t
2(0,t )2 2
Because of the identity

(t
t
2 2t
),
)
=
((
)R
+
0,t 20,t 0,t 8(0,t )2
2(0,t )2 2
one deduces the equality
R1,t
Z
=
Ut
(Us)1 (
22s
2 2s
))ds.
+ 1,s (( s
)R0,s +
80,s
0,s 20,s
8(0,s )2
28
Then
t
2 2s
2 2s
))ds
+ |1,s |(( s +
)|R0,s | +
80,s
0,s 20,s
8(0,s )2
0
Z t
Z t
2 2
2 2s
1 s
1 s
Ut (Us ) (
U
(U
)
))ds
+
+ |1,s |(
|R0,s | +
|1,s kR0,s |ds
s
t
2
80,s
20,s
0,s
8(0,s )
0
0
Z t 2 2
2
2
1,.R0,.
s
s
))ds + (
+ |1,s|(
|R0,s | +
) ,
(
2
20,s
0,. t
8(0,s )
0 80,s
Rt
where we have used Ut (Us )1 1 for every s [0, t] and Ut 0 s (Us )1 ds = 1 Ut 1
for the third inequality. Apply Proposition 4.6 and Inequality (4.13) to complete the
proof of the estimate of k(R1,. )t kp .

We define (2,t )0tT as the solution of the linear equation
|R1,t |
Ut (Us )1 (
hal-00370717, version 1 - 24 Mar 2009
d2,t = ((
2t
(1,t )2
t
)
+
)dt, 2,0 = 0.
t
2,t
2(0,t )2 2
(0,t )3 40,t
Clearly, for p 1, we have
k(2,. )t kp C(Sup t)2 .
(4.14)
(4.15)
2
P 4.8. Under (P), the process (R2,t = t 0t 1,t 2 2,t )0tT fulfills the
equality:
Z t
2,s (1,s )2
2 2s
(Us)1 [2 (( s
R2,t = Ut
)(
)R0,s +
)
0,s 20,s
0,s
8(0,s )2 2
0
+ ((
2 2s
s
)1,s ]ds.
)R1,s +
0,s 20,s
8(0,s )2
Moreover, for every p 1, one has
k(R2,. )t kp C(Sup t)3 .
In particular, the application t is C2 at = 0 in Lp sense with the second derivative at

= 0 equal to 2,t .
Proof. The equality is easy to check. The estimate is proved in the same way as
in the proof of Proposition 4.7, we therefore skip the details.
C 4.9. The application vt is C2 at = 0 in Lp sense. The residuals for the
squared volatility satisfy the following inequalities: for every p 1, one has
k(Rv0,. )t kp CSup t,
k(Rv1,. )t kp C(Sup t)2 ,
k(Rv2,. )t kp C(Sup t)3 .
Proof. Note that vt = (t )2 = (0,t + R0,t )2 = v0,t + 20,t R0,t + (R0,t )2 . Thus, we have
p
Rv0,t = 20,t R0,t +(R0,t )2 , which leads to the required estimate using 0,t max( v0 , Sup )
and Proposition 4.6. The other estimates are proved analogously using Propositions
4.7 and 4.8 and Inequalities (4.13) and (4.15).
29
4.4.2. Proof of Proposition 4.4. We can write

T
Z
P1T
=
0
1
t R0,t dBt
2t
2
1
Rv0,t dt,
1
RP2,T
Z
=
0
1
t R2,t dBt
Z
0
2t
2
Rv2,t dt.
Then, using Propositions 4.6, 4.8 and Corollary 4.9, we prove the two first estimates
of Proposition 4.4. The others inequalities are proved in the same way.
4.5. Proof of Theorem 2.4. For convenience, we introduce the following notation
for [0, 1]:
Z

PBS () =PBS x0 +
0
t ((1 ) v0,t +
T
Z
,
hal-00370717, version 1 - 24 Mar 2009
Z
q
1
vt )dBt
2t
2
((1 )v0,t + v1t )dt

(1 2t )((1 )v0,t + v1t )dt ,
Z T
Z T 2
q

t
i+j PBS
i+j P BS
1
t ((1 ) v0,t + vt )dBt
x0 +
() =
((1 )v0,t + v1t )dt
j
j
i
i
x y
x y
0
0 2
Z T

2
1
,
(1 t )((1 )v0,t + vt )dt .
0
Notice that P BS (see (2.5)) is a particular case of P BS for = 0:

P BS = P BS (0),
i+j P BS i+j P BS
=
(0).
xi y j
xi y j
Now, we represent the error E in (2.12) using the previous notations. A second order
Taylor expansion leads to
1
g(1) = E(P BS (1)) = E(P BS (0) + P BS (0) + 2 P BS (0) +
2
d
0
(1 )2 3
PBS ()).
2
The first term E(P BS (0)) is equal to (2.6). Approximations of the three above derivatives contribute to the error E.
P
P
1. We have E( P BS (0)) = E( xBS P1T + yBS Q1T ). These two terms are equal to
(2.7) and (2.8) plus an error equal to
E(
P BS P1 P BS Q1
R +
R ).
x 2,T
y 2,T
2. Regarding the second derivatives, we have E( 12 2 P BS (0)) = E( 12
2
2 P BS 1 1
1 P BS
1 2
2 y2 (QT ) + xy PT QT ).
2 P BS
(P1T )2 +
x2
These terms are equal to (2.9), (2.10) and (2.11), plus an
error equal to
E(
1 2 P BS (P1 )2 1 2 P BS (Q1 )2 2 P BS P1 Q1
).
R
+
+
R
R
2 x2 2,T
2 y2 2,T
xy 2,T
3. The last term with 3 P BS is neglected and thus is considered as an error.
30
To sum up, we have shown that

E=
1
X
E[
i=0
Z
+
0
2
X
Ci2
2 P BS
1 P BS
(P1 )i (Q1 )1i
(P1 )i (Q1 )2i
E[
]
+
]
(0)R
(0)R2,T
2,T
i
1i
i
2i
2
x y
x y
i=0
3
(1 )2 X i
3 P BS
C3 E[ i 3i ()(P1T )i (Q1T )3i ]d.
2
x y
i=0
Using Lemma 4.1 and Assumption (R), one obtains for all [0, 1]
k
i+j P BS
()k2 Ck(
xi y j
Z
0
((1 )v0,t + v1t )dt)

Z
C((1 )k(
v0,t dt)
(2 j+i1)+
2
(2 j+i1)+
2
k4
Z
k4 + k(
v1t dt)
(2 j+i1)+
2
k4 )
hal-00370717, version 1 - 24 Mar 2009
where we have applied a convexity argument. Finally, apply Lemma 4.3 with = 0
and = 1 to conclude that
k
C
i+j P BS
()k2
,
j
i
x y
( T)(2 j+i1)+
uniformly w.r.t. [0, 1]. Combining this with Proposition 4.4 yields that
1
2
3
X
X
X
T2i/2
T1i/2
T3i/2
3
3
3
+
+
(Sup T)
(Sup T)
|E| C (Sup T)
( T)1i i=0
( T)3i i=0
( T)5i
i=0
C(Sup T)3 T.
Theorem 2.4 is proved.
5. Proof of Proposition 2.2 and Theorem 2.3.
5.1. Preliminary results. In this section, we bring together the results (and their
proofs) which allow us to derive the explicit terms in the formula (2.13).
In the following, t (resp. t ) is a square integrable and predictable process (resp.
deterministic) and l is a smooth function with derivatives having, at most, exponential
growth.
For the next Malliavin calculus computations, we freely use standard notations from
[Nua06].
L 5.1. (Lemma 1.2.1 in [Nua06]) Let G D1, (). One has
Z t
Z t
s DBs (G)ds],
s dBs ] = E[
E[G
0
where
DB (G) =
(DBs(G))s0 is the first Malliavin derivative of G

RT
= l( 0 t v0,t dBt ) gives the following result.
Taking G
L 5.2. One has:
Z T
Z T
Z
t v0,t dBt )] = E[(

t dBt )l(
E[(
0
w.r.t. B.
Z
(1)
t v0,t t dt)l (
0
t v0,t dBt )].
31
L 5.3. For any deterministic integrable function f and any continuous semimartingale Z vanishing at t=0, one has:
T
Z
f (t)Zt dt =
0
(0, f )
t,T dZt .
(0, f )
Proof. This is an application of the Ito formula to the product t,T Zt .

L 5.4. One has:
Z T
Z T
Z T
(,v0,. ),(,)
(1)
t v0,t dBt )],
E[l (
t v0,t dBt )
t v1,t dt] = 0,T
E[l(
0
0
T
Z
E[l(
t v0,t dBt )
Z
0
(2,2 v
t v21,t dt] = 0,T
0,. ),(2,)
t v0,t dBt )]
E[l(
0
(,v ),(,v0,. ),(2,)

E[l(2) (
+ 20,T 0,.
T
hal-00370717, version 1 - 24 Mar 2009
E[l(
t v0,t dBt )
Z
0
(,v0,. ),(0,),(,)
t v2,t dt] = 0,T
t v0,t dBt )
Z
t v1,t dt] = E[l(
E[l(2) (
t v0,t dBt )],
t v0,t dBt )].
Proof. Using Lemmas 5.2 ( f (t) = et t , Zt =

E[l(
Rt
0
es s v0,s dBs ) and 5.3, one has:
t v0,t dBt )
t v0,t dBt )
= E[l(
0
T
Z
= E[l(1) (
t v0,t dBt )]
(,)
t,T et t
es s v0,s dBs dt]
0
T
Z
et t
v0,t dBt ]
(,) t
t,T
e t t v0,t dt,
which gives the first equality. The second and the third are proved in the same way.
L 5.5. One has
i+j PBS
i+j P BS
]
=
(x0 ,
E[
xi y j
xi y j
v0,t dt).
0
Proof. One has

Z T 2
Z T
Z T
t
i P BS
i
E[
t v0,t dBt
(1 2t )v0,t dt)]
] = x=x0 E[PBS (x0 +
v0,t dt,
xi
0 2
0
0
Z T
i PBS
(x
,
v0,t dt).
=
0
xi
0
Since P BS verifies the following relation
P BS 1 2 P BS P BS
= (
),
y
2 x2
x
we immediately obtain the result.
(5.1)
32
5.2. Proof of Proposition 2.2. One has
E[
P BS
(
x
t (
0
v1,t
2(v0,t )
1
2
v2,t
4(v0,t )
1
2
v21,t
Z
)dBt
3
8(v0,t ) 2
2t
(v1,t +
2
Z T 2 v2
Z T
v2,t
1 2 P BS P BS
2P BS
= E[ (
2t (v1,t +
)
)dt] E[
2
2 x
x
2
x2 0
0
Z
Z
2 2
v2,t
2P BS T t v1,t
P BS T 2
t (v1,t +
)dt] E[
dt],
= E[
y 0
2
x2 0 8v0,t
v2,t
)dt)]
2
t 1,t
8v0,t
dt]
hal-00370717, version 1 - 24 Mar 2009
where we have used Lemma 5.2 at the first equality and identity (5.1) at the second
one. Plugging this relation into the approximation (2.12) and summing the second
and third line, one has
Z
v2,t
P BS T
g(1) =E[PBS ] + E[
)dt]
(v1,t +
2
y 0
Z T 2
Z T
Z
2 2
t
v1,t
1 2 P BS
2 P BS T t v1,t
dt] + E[
dBt
v1,t dt)2 ]
E[
t
(
1
2
2
2
x
x
0 2
0
0 8v0,t
2(v0,t ) 2
Z T
1 2 P BS
+ E[
(1 2t )v1,t dt)2]
(
2
y2
0
Z T
Z T 2
Z T
t
v1,t
2 P BS
2
t
(1 t )v1,t dt)(
dBt
(
v1,t dt)] + E.
(5.2)
+E[
1
xy
0
0 2
0
2(v ) 2
0,t
In addition, one has
2t v21,t
Z T 2
Z T
t
v1,t
1 2 P BS
dt] + E[
v1,t dt)2 ]
dBt
t
(
1
2
2
x
0 2
0
0 8v0,t
2(v0,t ) 2
Z Z t
Z t 2
2t
v1,s
s
v1,t
2 P BS T
= E[
(
dB
v
ds)(
dB
v1,t dt)]
s
s
t
t
1,s
1
1
2
x2 0
0
0 2
2(v0,s ) 2
2(v0,t ) 2
Z T Z t
Z t 2
v1,s
s
1 3 P BS 2 P BS
s
(
)
v1,s ds)2t v1,t dt]
dBs
= E[ (
1
2
2 x3
x
0
0
0 2
2(v0,s ) 2
Z t 2
Z Z t
v1,s
s
2 P BS T
s
(
dB
v1,s ds)2t v1,t dt],

= E[
s
1
xy 0
2
0
0
2
2(v0,s )
2 P BS
E[
x2
where we have used Itos Lemma for the square at the first equality, Lemma 5.2
at the second and Identity (5.1) at the third one. Substituting this relation in the
approximation (5.2) and summing the second and fourth line, one gets
Z
v2,t
P BS T
g(1) =E[PBS ] + E[
(v1,t +
)dt]
y 0
2
Z T Z t
Z t 2
v1,s
s
2 P BS
s
(
+E[
(
v1,s ds)2t v1,t dt
dB
s
1
xy
2
0
0
0
2(v0,s ) 2
Z T 2
Z T
Z T
t
v1,t
2
dBt
v1,t dt))]
t
(1 t )v1,t dt)(
+(
1
0 2
0
0
2(v0,t ) 2
Z T
1 2 P BS
(
(1 2t )v1,t dt)2] + E.
+ E[
2
y2
0
33
(5.3)
We now study the second term of (5.3). In the computations below, we use Itos
Lemma for the second equality, Lemma 5.2 and Identity (5.1) for the third equality
hal-00370717, version 1 - 24 Mar 2009
and Lemma 5.1 (G =
2 P BS
v )
xy 1,t
for the fourth one; it gives
Z T Z t
Z t 2
v1,s
s
2 P BS
s
(
(
dB
v1,s ds)2t v1,t dt

s
1
xy
2
0
0
0
2(v0,s ) 2
Z T 2
Z T
Z T
t
v1,t
2
dBt
v1,t dt))]
(1 t )v1,t dt)(
+(
t
1
0 2
0
0
2(v0,t ) 2
Z t 2
Z T Z t
s
v1,s
2 P BS
dB
v1,s ds)(2t + 1 2t )v1,t dt

(
=E[
(
s
s
1
2
xy
2
0
0
0
2(v0,s )
Z T Z t
2
v1,t
+
( (1 2s )v1,s ds)(t
dBt t v1,t dt))]
1
2
0
0
2(v0,t ) 2
Z
Z
Z T
t
t 2
v1,s
s
2 P BS
s
v1,t (
dB
v1,s ds)]dt
E[
=
s
1
xy
0
0 2
0
2(v0,s ) 2
Z Z t
2 P BS T
( (1 2s )v1,s ds)2t v1,t dt]
+ E[
y2 0
0
Z T
Z t 2
Z t
s
v1,s
2 P BS
E[
=
s
(v1,t (
v1,s ds) +
DB v1,t ds)
xy
2 v0,s s
0
0 2
0
Z t 2
Z Z t
s
3 P BS
2P BS T
+
( (1 2s )v1,s ds)2t v1,t dt].
v1,t
v1,s ds]dt + E[
x2 y
y2 0
0 2
0
From Equation (2.3), one has DBs v1,t = ekt eks s v0,s . Hence it is deterministic. Thus,
using Identity (5.1) and Lemma 5.2 for the first equality and Equation (2.4) for the
second equality, one has:
Z t
Z t
Z
2 P BS T
2
(1 2s )v1,s ds)2t v1,t dt)]
v
ds)v
dt
+
(
((
A =E[
1,t
s 1,s
y2 0
0
0
Z Z t
v1,s kt ks
P BS T
+ E[
(
e e s v0,s dBs )dt]
y 0
0 2v0,s
Z t
Z t
Z
Z
P BS T v2,t
2 P BS T
2
2
2
(1
)v
ds)
v
dt)]
+
E[
=E[
v
ds)v
dt
+
(
((
dt].
1,t
s 1,s
s 1,s
t 1,t
y 0 2
y2 0
0
0
A =E[
34
Now, plug this last equality into (5.3) and use the identity
T
((
0
0
T
((
0
0
T
((
0
Z
2s v1,s ds)v1,t dt + (
Z
2s v1,s ds)v1,t dt + (
(1 2s )v1,s ds)2t v1,t dt) +
1
(
2
Z
0
(1 2t )v1,t dt)2 =
(1 2s )v1,s ds)(2t + 1 2t )v1,t dt) =
(2s + 1 2s ))v1,s ds)v1,t dt =
1
(
2
v1,t dt)2 ;
it immediately gives the result.

5.3. Proof of Theorem 2.3. Proof. Step 1: We show the equality
hal-00370717, version 1 - 24 Mar 2009
P BS
E[
y
(v1,t + v2,t )dt] =
2
X
ai,T
i+1 PBS (x0 ,
RT
0
xi y
i=1
v0,t dt)
where
(,v0,. ),(,1)
a1,T = 0,T
(,v0,. ),(0,),(,1)
a2,T = 0,T
Actually, the result is an immediate application of Lemma 5.4 and Lemma 5.5.
Step 2: We show the equality
1 2 P BS
(
E[
2
y2
1
X
v1,t dt) ] =
0
b2i,T
i=0
2i+2 PBS (x0 ,
RT
0
x2i y2
v0,t dt)
where
(2,2 v0,. ),(,1),(,1)
b0,T = 0,T
(,v0,. ),(,1),(,v0,. ),(,1)
b2,T = 0,T
(,v0,. ),(,v0,. ),(,1),(,1)
+ 20,T
a21,T
2
Indeed, one has

1 2 P BS
(
E[
2
y2
Z
0
v1,t dt)2 ] = E[
2 P BS
y2
v1,s ds)v1,t dt]
(
0
Z Z T
Z t
2 P BS T
s
t 2
t
= E[
(
e ds)(e v1,t dt + t v0,t e (
v1,s ds)dBt )]
y2 0
t
0
Z Z T
Z
3P BS T (,v0,. ),(,1)
2 P BS T
s
t 2
e
ds)e
v
dt]
+
E[
v1,t dt],
(
= E[
1,t
y2 0
xy2 0 t,T
t
Rt
where we have used Lemma 5.3 ( f (t) = et , Zt = ( 0 v1,s ds)(et v1,t )) for the second
RT
Rt
equality and Lemmas 5.2 and 5.3 ( f (t) = ( t es ds)t t v0,t et , Zt = 0 v1,s ds) for the
last one.
An application of the first and second equality in Lemma 5.4 gives the announced
35
result. Actually, it remains to show that b2,T = a21,T /2. Indeed, consider two càdlàg
functions f and g : [0, T] R. Then
(
RT
0
ft (
RT
t
gs ds)dt)2
R TR T
0
ft1 (
RT
gt3 dt3 ) ft2 (
t1
t2
gt4 dt4 )dt2 dt1
2
T
Z
ft1 (
=
0
T
Z
0
TZ T
t1
Z
=2
0
Z
ft1
t1
Z
+
0
Z
ft1
T
t1
gt4 dt3 dt4 dt2 dt1
t3
Z
gt3
gt4 dt4 dt2 dt3 )dt1
t2
T
Z
gt3
t2
Z
ft2
t3
T
Z
ft2
gt3 gt4 dt3 dt4 dt2
Z
gt3
t1
T
gt4 dt4 )dt3 dt2 )dt1
t2
t2
t2
TZ T
Z
ft2
Z
gt3 ft2 (
t1
t1
Z T
ft1 (
hal-00370717, version 1 - 24 Mar 2009
RT
T
t3
Z
ft2
t2
gt4 dt4 dt2 dt3 dt1 .
Putting f (t) = t t v0,t ekt and g(t) = ekt in the previous equality readily gives b2,T =
which finishes the proof.
a21,T
2
6. Conclusion. We have established an approximation pricing formula for call/put

options in the time dependent Heston models. We prove that the error is of order
3 w.r.t. the volatility of volatility and 2 w.r.t. the maturity. In practice, taking the
Fourier method as a benchmark, the accuracy is excellent for a large range of strikes
and maturities. In addition, the computational time is about 100 to 1000 times smaller
than using an efficient Fourier method.
Following the arguments in [BGM09b], our formula extends immediately to
other payoffs depending on ST (note that the identities (2.2) and (5.1) are valid for
any payoff of this type). As explained in [BGM09b], the smoother the payoff, the
higher the error order w.r.t. T; the less smooth the payoff, the lower the error order
w.r.t. T. For digital options, the error order w.r.t. T becomes 3/2 instead of 2.
Extensions to exotic options and to the third order expansion formula w.r.t. the
volatility of volatility are left for further research.
7. Appendix: closed formulas in Heston model. There are few closed repreRt
sentations for the call/put prices written on the asset St = e 0 (rs qs )ds eXt in the Heston
model (defined in (1.1) and (1.2)). We focus on the Heston formula [Hes93] and on
the Lewis formula [Lew00]. Both of them rely on the knowledge of the characteristic
function of the log-asset price (Xt )t and on Fourier transform-based approaches.
(i) In [Hes93], Heston obtains a representation in a Black-Scholes form:
CallHeston (t, St , vt ; T, K) = St e
RT
t
qs ds
P1 Ke
RT
t
rs ds
P2 ,
where both probabilities P1 and P2 are equal to a one-dimensional integral of characteristic functions.
(ii) In [Lew00], Lewis takes advantage of the generalized Fourier transform,
by using an integration along a straight line in the complex plane parallel to the real
36
axis. It is important to detect the strip where the integration is safe. Lewis suggests
the use of complex numbers z such that Im(z) = 21 . His formula writes
RT
CallHeston (t, St , vt ; T, K) =
where X = log
RT
St e t qs ds
RT
Ke t rs ds
RT
St e t
rs ds
qs ds Ke t
i
2 +
i
2
eizX T (z)
dz
z2 iz
!
and T (z) = E(ez(XT Xt ) |Ft ). Then, the above integral is
hal-00370717, version 1 - 24 Mar 2009
evaluated by numerical integration.

Using PDE arguments in combination with affine models, we can obtain an explicit
formula for T (z) in the case of constant Heston parameters. In addition, it can
be computed without discontinuities in z, following the arguments in [JK05]. For
piecewise constant parameters, the characterictic function T (z) can be computed
recursively using nested Riccatti equations with constant coefficients: we refer to the
work by Mikhailov and Nogel [MN03].
In our numerical tests, we prefer the Lewis formula which gives better numerical
results, in particular for very small or very large strikes, compared to the Heston
formula.
REFERENCES
[Alf05]
[Alf08]
[ALV07]
[And08]
[AS09]
[BBF04]
[BD07]
[BGM09a]
[BGM09b]
[BK06]
[BS02]
[Dio03]
[Duf01]
[Eli08]
[For08]
[FPS00]
[GJY03]
A. Alfonsi. On the discretization schemes for the CIR (and Bessel squared) processes. Monte
Carlo Methods and Applications, 11(4):355384, 2005.
A. Alfonsi. High order discretization schemes for the CIR process: application to Affine
Term Structure and Heston models. ssrn working paper, 2008.
E. Alos,
` J. A. Leon,
and J. Vives. On the short-time behaviour of the implied volatility for
jump-diffusion models with stochastic volatility. Finance and Stochastics, 11(4):571589,
2007.
L. Andersen. Efficient simulation of the Heston stochastic volatility model. ssrn working
paper, 2008.
F. Antonelli and S. Scarlatti. Pricing options under stochastic volatility: a power series
approach. Finance and Stochastics, 13(2):269303, 2009.
H. Berestycki, J. Busca, and I. Florent. Computing the implied volatility in stochastic
volatility models. Comm. Pure Appl. Math., 10:13521373, 2004.
M. Bossy and A. Diop. An efficient discretization scheme for one dimensional SDEs with a
diffusion coefficient function of the form |x|a , a in [1/2,1). Technical report, INRIA,preprint
RR-5396 - version 2, 2007.
E. Benhamou, E. Gobet, and M. Miri. Closed forms for European options in a local volatility
model. To appear in International Journal of Theoretical and Applied Finance, 2009.
E. Benhamou, E. Gobet, and M. Miri. Smart expansion and fast calibration for jump
diffusion. To appear in Finance and Stochastics, 2009.
M. Broadie and O. Kaya. Exact simulation of stochastic volatility and other affine jump
diffusion processes. Operations Research, 54(2):217231, 2006.
A.N. Borodin and P. Salminen. Handbook of Brownian motion-facts and formulae. Probability
and its Applications, Birkhauser, second edition, 2002.
A. Diop. Sur la discrètisation et le comportement a` petit bruit dEDS multidimensionnelles dont les coefficients sont a` derivees singulières. PhD thesis, INRIA,(available at
http://www.inria.fr/rrrt/0785.html), 2003.
D. Dufresne. The integrated square-root process. Working paper,University of Montreal, 2001.
A. Elices. Models with time-dependent parameters using transform methods: application
to Hestons model. arXiv:0708.2020v2, 2008.
M. Forde. The small-time behavior of diffusion and time-changed diffusion processes on
the line. arXiv:math.PR/0609117, 2008.
J.P. Fouque, G. Papanicalaou, and R. Sircar. Derivatives in financial Markets with stochastic
volatility. Cambridge University Press, 2000.
A. Going-Jaeschke
and M. Yor. A survey and some generalizations of Bessel processes.

Bernoulli, 9(2):313349, 2003.

[Hes93]
[HKLW02]
[JK05]
[Klu02]
[KS88]
[Lab05]
[Lew00]
[Lew07]
[Lip02]
[Mag96]
[Mir09]
[MN03]
hal-00370717, version 1 - 24 Mar 2009
[Nua06]
[Osa07]
[Pit05a]
[Pit05b]
[Pro90]
[RT96]
[RY99]
[Smi08]
37
S. Heston. A closed-form solutions for options with stochastic volatility with applications
to bond and currency options,. Review of Financial Studies, 6:327343, 1993.
P.S. Hagan, D. Kumar, A.S. Lesniewski, and D.E. Woodward. Managing smile risk. Willmott
Magazine, 15:84108, September 2002.
P. Jackel and C. Kahl. Not-So-Complex Logarithms in the Heston Model. Wilmott Magazine,
19:94103, September 2005.
T. Kluge. Pricing derivatives in stochastic volatility models using the Finite Difference method.
PhD thesis, Technical University, Chemnitz, 2002.
I. Karatzas and S.E. Shreve. Brownian motion and stochastic calculus. Springer-Verlag, New
York, 1988.
P. Henry Labordère. A general asymptotic implied volatility for stochastic volatility models.
arXiv:condmat/04317, 2005.
A.L. Lewis. Option valuation under stochastic volatility with Mathematica code. Finance Press,
Newport Beach, California, 2000.
A. Lewis. Geometries and smile asymptotics for a class of stochastic volatility models.
www.optioncity.net, 2007.
A. Lipton. The vol smile problem. Risk Magazine, 15(2):6165, 2002.
Y. Maghsoodi. Solution of the extended CIR term structure and bond option valuation.
Mathematical finance, 6(1):89109, 1996.
M. Miri. in preparation. PhD thesis, Universite de Grenoble, 2009.
S. Mikhailov and U. Nogel. Hestons stochastic volatility model implementation, calibration
and some extensions. Wilmott magazine, pages 7479, July 2003.
D. Nualart. Malliavin calculus and related topics. Springer-Verlag, Berlin, second edition,
2006.
Y. Osajima. The asymptotic expansion formula of implied volatility for dynamic SABR
model and FX hybrid model. Available at http://ssrn.com/abstract=965265, 2007.
V.V. Piterbarg. Stochastic volatility model with time-dependent skew. Applied Mathematical
Finance, 12(2):147185, 2005.
V.V. Piterbarg. Time to smile. Risk Magazine, 18(5):7175, 2005.
P. Protter. Stochastic integration and differential equations. Springer, Heidelberg, 1990.
E. Renault and N. Touzi. Option hedging and implied volatilities in a stochastic volatility
model. Mathematical Finance, 6:279302, 1996.
D. Revuz and M. Yor. Continuous martingales and Brownian motion. Springer-Verlag, Hardcover, 3rd edition, 1999.
R.D. Smith. An almost exact simulation method for the Heston model. Journal of Computational Finance, 11(1):115125, 2008.

Time Dependent Heston Model: Abstract

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Time Dependent Heston Model: Abstract

Uploaded by

Copyright:

Available Formats

TIME DEPENDENT HESTON MODEL

E. BENHAMOU , E. GOBET , AND M. MIRI

hal-00370717, version 1 - 24 Mar 2009

AMS subject classifications. 60J75, 60HXX

1. Introduction. Stochastic volatility modeling has emerged in the late nineties

Pricing Partners, 204 rue de Crimee 75019, Paris, FRANCE (eric.benhamou@pricingpartners.com).

and Yor [GJY03],

Electronic copy available at: http://ssrn.com/abstract=1367955

E. BENHAMOU, E. GOBET AND M. MIRI

hal-00370717, version 1 - 24 Mar 2009

Electronic copy available at: http://ssrn.com/abstract=1367955