M. Sbai
Joint work with B. Jourdain
le-logo
Sbai (UPE-CERMICS)
CNF 2009
1 / 32
Outline
Introduction
Model Specification
Calibration
Numerical experiments
Conclusion
le-logo
Sbai (UPE-CERMICS)
CNF 2009
2 / 32
Our objective : a new modeling approach allowing for a good fit of both Index
and stocks.
le-logo
Sbai (UPE-CERMICS)
CNF 2009
3 / 32
Another viewpoint : a factor model (the index represents the market and
influences the stocks).
A new correlation structure. Correlation risk
(In discrete time) Cizeau, Potters and Bouchaud [2001] show that it is
possible to capture the essential features of stocks cross-correlations by a
simple non-Gaussian one factor model, specially in extreme market
conditions :
ri (t) = i rm (t) + i (t)
where ri (t) =
Si (t)
Si (t1)
Sbai (UPE-CERMICS)
CNF 2009
4 / 32
Outline
Introduction
Model Specification
Calibration
Numerical experiments
Conclusion
le-logo
Sbai (UPE-CERMICS)
CNF 2009
5 / 32
wj Stj,M
ItM =
j=1
dStj,M
Stj,M
CNF 2009
6 / 32
i j 2 (t, ItM )
i2 2 (t, ItM ) + i2 (t, Sti,M )
Note that they depend not only on the stocks but also on the index.
le-logo
Sbai (UPE-CERMICS)
CNF 2009
7 / 32
j,M
M
j=1 j wj St
dt +
j,M
M
j=1 j wj St
j,M
j,M
j
M
j=1 wj St j (t, St )dWt
(2)
ItM .
j=1 j wj St
For large M, we will show that the term
neglected.
j
j
j
M
j=1 wj St j (t, St )dWt
can be
= rj = j rI M + j W j + drift
where rj (resp. rI M ) is the log-return of the stock j (resp. the index).
The return of a stock is decomposed into a systemic part driven by the index,
le-logo
which represents the market, and a residual part.
Sbai (UPE-CERMICS)
CNF 2009
8 / 32
A simplified Model
We look at the asymptotics for a large number of underlying stocks.
Consider the limit candidate (It )t[0,T] solution of
I0 = I0M
(3)
Theorem 1
Let p . If
(H1) Kb s.t. (t, s), |(t, s)| + |j (t, s)| Kb
K s.t. (t, s1 , s2 ), |s1 (t, s1 ) s2 (t, s2 )| K |s1 s2 |.
then, CT a constant independent of M such that
0tT
CT
M
2
j=1 wj
+
Sbai (UPE-CERMICS)
M
j=1 wj |j
M
j=1 wj |j
2p
2p
le-logo
CNF 2009
9 / 32
Theorem 2
M
j
Denote by I t = M
j=1 wj St . Under the assumptions of Theorem 1 and if
(H2) K s.t. (t, s1 , s2 ), |s1 (t, s1 ) s2 (t, s2 )| K |s1 s2 |
KLip s.t. (t, s1 , s2 ), |(t, s1 ) (t, s2 )| KLip |s1 s2 |
then, j {1, . . . , M}, CTj s.t.
CTj
M
2
j=1 wj
0tT
Sbai (UPE-CERMICS)
M
j=1 wj |j
max1jM CTj
+
M
j=1 wj |j
2p
2p
M
j=1 wj
M
j=1 wj |j
2p
2p
M
2
j=1 wj
M
j=1 wj |j
CNF 2009
2p
le-logo
10 / 32
The limit M +
Corollary
Under the additional assumptions
(H3) A s.t. maxj1 (S0j,M )2 + (jM )2 + (jM )2 A,
M
M 2
j=1 (wj ) M 0,
M
M M
j=1 wj |j | M 0,
(H4) PM
w =
(H5) PM
=
M
M
wM
j |j | 0,
(H6) PM
=
j=1
2
M
wM
M 0.
j < then sup0tT |It I t |
M
sup
M j=1
le-logo
Sbai (UPE-CERMICS)
CNF 2009
11 / 32
PM
w
0.026
opt
0.975
2
(PM
opt )
0.0173
2
(PM
=1 )
0.0174
M 2
TABLE: Computation of PM
w , opt and (Popt ) for the Eurostoxx index at December
21, 2007. The beta coefficients are estimated on a two year history.
le-logo
Sbai (UPE-CERMICS)
CNF 2009
12 / 32
To sum up, under mild assumptions, when the number of underlying stocks is
large, our original model may be approximated by
j {1, . . . , M},
dStj
(4)
We end up with
A local volatility model for the index
A stochastic volatility model for each stock, decomposed into a systemic
part driven by the index level and an intrinsic part.
Careful ! Our simplified model is not valid for options written on the index
together with all its composing stocks since the index is no longer an exact,
but an approximate, weighted sum of the stocks. Instead, one should consider
M
j
le-logo
the reconstructed index I t = M
j=1 wj St or use the original model.
Sbai (UPE-CERMICS)
CNF 2009
13 / 32
Outline
Introduction
Model Specification
Calibration
Numerical experiments
Conclusion
le-logo
Sbai (UPE-CERMICS)
CNF 2009
14 / 32
dSt
= (r )dt + (t, It )dBt + (t, St )dWt
St
(5)
dIt
= (r I )dt + (t, It )dBt .
It
Fitting the index smile boils down to the calibration of a local volatility
model.
Fitting an individual stock smile is more tedious.
Our model gives an advantage to the fit of index option prices (index
options are usually more liquid than individual stock options).
le-logo
Sbai (UPE-CERMICS)
CNF 2009
15 / 32
Non-parametric estimation of
We have a relation between the local volatility and the stochastic volatility
(see Gyongi [4] or Dupire [3]) :
vloc (t, K) = 2 (t, K) + 2 2 (t, It ) | St = K
So,
(t, K) =
(6)
vloc can be calibrated with the best-fit of a parametric form to the stock
market smile.
Estimating the conditional expectation is more challenging (it depends
implicitly on as it is the case for (St , It )).
We investigate a simulation based approach yielding a non-parametric
estimation of .
le-logo
Sbai (UPE-CERMICS)
CNF 2009
16 / 32
dIt
= (r I )dt + (t, It )dBt
It
This SDE is non-linear in the sense of McKean.
Kernel estimators of the Nadaraya-Watson type :
N
2 (t, Iti )K
2 (t, It ) | St = s
i=1
N
K
i=1
s Sti
hN
s Sti
hN
CNF 2009
17 / 32
N
j=1
2 (t,Itj )K
N
j=1
j,N
i,N
St St
hN
j,N
i,N
St St
hN
dWti
Sbai (UPE-CERMICS)
CNF 2009
18 / 32
dStj,M
Stj,M
wi Sti,M
ItM =
i=1
CNF 2009
19 / 32
Outline
Introduction
Model Specification
Calibration
Numerical experiments
Conclusion
le-logo
Sbai (UPE-CERMICS)
CNF 2009
20 / 32
An acceleration technique
The simulation of the particle system is time consuming : a global
complexity of order O(nN 2 ) where n is the number of time steps in the
Euler scheme.
A possible acceleration technique : neglect particles which are far away
from each other.
How ? Sort the particles and stop the estimation of the conditional
expectation whenever the contribution of a particle is lower than some
fixed threshold.
We lose in precision but we gain much more in computation time.
le-logo
Sbai (UPE-CERMICS)
CNF 2009
21 / 32
Data :
Local volatilities of the Eurostoxx index and of Carrefour at December
21, 2007.
Beta coefficient estimated on a two years history ( = 0.7).
Short interest rate and dividend yields as of December 21, 2007.
Maturity T = 1.
Threshold for the accelerated technique :
1
N.
Smoothing parameter : hN = N 10 .
Number of time steps for the Euler scheme : n = 20.
Moneyness ( SK0 )
Error : |simul exact |
0.5
36
0.7
8
0.9
2
1
1
1.1
2
1.2
9
1.5
32
2
56
TABLE: Error (in bp) on the implied volatility with N = 200000 particles.
le-logo
Sbai (UPE-CERMICS)
CNF 2009
22 / 32
0.40
Exact Implied Vol.
N=10000
0.38
N=200000
0.36
0.34
0.32
0.30
0.28
0.26
0.24
0.5
0.6
0.7
0.8
0.9
1.0
1.1
1.2
1.3
1.4
1.5
Moneyness
Sbai (UPE-CERMICS)
CNF 2009
le-logo
23 / 32
dStj,M
= rdt + (t, ItM )dBt
Stj,M
i,M
ItM = M
i=1 wi St .
dStj
Stj
dIt
It
i
Reconstructed index I t = M
i=1 wi St .
The constant-correlation model
j {1, . . . , M},
dStj
Stj
= rdt +
le-logo
CNF 2009
24 / 32
le-logo
Sbai (UPE-CERMICS)
CNF 2009
25 / 32
0.42
Simplified
0.40
Market
Original
0.38
0.36
0.34
0.32
0.30
0.28
0.26
0.24
0.22
0.20
0.5
1.0
1.5
2.0
Moneyness
Sbai (UPE-CERMICS)
CNF 2009
26 / 32
0.40
Simplified
Market
Original
0.35
Simplified Reconstructed
0.30
0.25
0.20
0.15
0.5
1.0
1.5
2.0
Moneyness
Sbai (UPE-CERMICS)
CNF 2009
27 / 32
Outline
Introduction
Model Specification
Calibration
Numerical experiments
Conclusion
le-logo
Sbai (UPE-CERMICS)
CNF 2009
28 / 32
le-logo
Sbai (UPE-CERMICS)
CNF 2009
29 / 32
Thank you !
le-logo
Sbai (UPE-CERMICS)
CNF 2009
30 / 32
References I
M. Avellaneda, D. Boyer-Olson, J. Busca, and P. Friz.
Reconstructing volatility.
Risk, pages 8791, October 2002.
P. Cizeau, M. Potters, and J-P. Bouchaud.
Correlation structure of extreme stock returns.
Quantitative Finance, 1(2) :217222, February 2001.
B. Dupire.
Pricing with a smile.
Risk, pages 1820, January 1994.
I. Gyongy.
Mimicking the one-dimensional marginal distributions of processes
having an Ito differential.
Probab. Theory Relat. Fields, 71(4) :501516, 1986.
le-logo
Sbai (UPE-CERMICS)
CNF 2009
31 / 32
References II
le-logo
Sbai (UPE-CERMICS)
CNF 2009
32 / 32