See discussions, stats, and author profiles for this publication at: [Link]

net/publication/227559301

Model Calibration

Chapter · May 2010

DOI: 10.1002/9780470061602.eqf08002

CITATIONS READS

3 2,251

1 author:

Rama Cont
University of Oxford
237 PUBLICATIONS 18,787 CITATIONS

SEE PROFILE

All content following this page was uploaded by Rama Cont on 19 July 2019.

The user has requested enhancement of the downloaded file.

Model Calibration Option prices being evaluated as expectations,
this inverse problem can also be interpreted as a
(generalized) moment problem for the law
of risk-
neutral process given a finite number of option prices,
The fundamental theorem of asset pricing (see Fun-
it is typically an ill-posed problem and can have
damental Theorem of Asset Pricing) shows that, in
many solutions. However, the number of observed
an arbitrage-free market, market prices can be rep-
options can be large (100 − 200 for index options)
resented as (conditional) expectations with respect to
and finding even a single solution is not obvious and
a martingale measure
: a probability measure

requires efficient numerical algorithms.
on the set
of possible trajectories (St )t∈[0,T ] of the
In the Black–Scholes model (see Black–
underlying asset such that the asset price St /Nt dis-
Scholes Formula), calibration amounts to picking
counted by the numeraire Nt is a martingale. The
the volatility parameter to be equal to the implied
value Vt (HT ) of a (discounted) terminal payoff HT
volatility of a traded option. However, if more than
at T is then given by
one option is traded, the Black–Scholes model cannot
Vt (HT ) = E
[B(t, T )HT |Ft ] (1) be calibrated to market prices, since in most options
markets implied volatility varies across strikes and
where B(t, T ) = Nt /NT is the discount factor. For maturities; this is the volatility smile phenomenon.
example, the value under the pricing rule
of Therefore, to solve the calibration problem, we need
a call option with strike K and maturity T is more flexible models, some examples of which are
given by E
[B(t, T )(ST − K)+ |Ft ]. However, this given here.
result does not say how to construct the pricing
measure
. Given that data sets of option prices have Example 1 [Diffusion Model (see Local Volatility
become increasingly available, a common approach Model)] If an asset price is modeled as a diffusion
for selecting a pricing model
is to choose, given process
a set of liquidly traded derivatives with (discounted)
dSt = St [µ dt + σ (t, St ) dWt ] (4)
terminal payoffs (H i )i∈I and market prices (Ci )i∈I ,
a pricing measure
compatible with the observed
parameterized by a local volatility function
market prices:
σ : (t, S) → σ (t, S) (5)
Problem 1 [Calibration Problem] Given market
prices (Ci )i∈I (say at date t = 0) for a set of options
then the values of call options can be computed by
with discounted terminal payoffs (Hi )i∈I , construct a
solving the Dupire equation (see Implied Volatility
probability measure
on
such that
Surface)
• the (discounted) asset price (St )t∈[0,T ] is a mar-
tingale under
∂C0 ∂C0 K 2 σ 2 (T , K) ∂ 2 C0
+ Kr − =0
∂T ∂K 2 ∂K 2
T ≥ t ≥ u ≥ 0 ⇒ E
[St |Fu ] = Su (2)
∀K ≥ 0, C0 (T = 0, K) = (S − K)+ (6)
• the pricing rule implied by
is consistent with
market prices The corresponding inverse problem is to find a
(smooth) volatility function σ : [0, T ] × + → +
∀i ∈ I , E
[Hi ] = Ci (3) such that C σ (Ti , Ki ) = C ∗ (Ti , Ki ) where C σ is the
solution of equation (6) and C ∗ (Ti , Ki ) are the market
where, for ease of notation, we have set discount prices of call options.
factors to 1 (prices are discounted) and E[.] denotes
the conditional expectation given initial information Example 2 In an exponential-Lévy model St =
F0 . Thus, a pricing rule
is said to be calibrated to exp Xt , where Xt is a Lévy process (see Exponential
the benchmark instruments Hi if the value of these Lévy Models) with diffusion coefficient σ > 0 and
instruments, computed in the model, correspond to Lévy measure ν, call prices C σ,ν (t0 , S0 ; Ti , Ki ) are
their market prices Ci . easily computed using Fourier-based methods (see
2 Model Calibration

Fourier Methods in Options Pricing). The calibra- Inversion formulas

tion problem is to find σ, ν such that
In the theoretical situation where prices of European
∀i ∈ I, C σ,ν (t0 , S0 ; Ti , Ki ) = C ∗ (Ti , Ki ) (7) options are available for all strikes and maturities,
the calibration problem can sometimes be explicitly
This is an example of a nonlinear inverse problem solved using an inversion formula.
where the parameter lies in a space of measures. For the diffusion model in Example 1, the Dupire
formula [25] (see Dupire Equation):
Example 3 In the LIBOR market model, a set

 ∂C0
of N interest rates (LIBOR rates) is modeled as  + Kr ∂C0
a diffusion process Lt = (Lit )i=1..N with constant σ (T , K) =  ∂T 2 2 ∂K (11)
K ∂ C0
covariance matrix  =t σ.σ ∈ Sym+ (n × n): 2 ∂K 2

dLit = µit dt + Lit σi . dWt

j
(8) allows to invert the volatility function from call
option prices. Similar formulas can be obtained in
This model can then be used to analytically credit derivative pricing models, for inverting port-
folio default rates from collateralized debt obligation
price caps, floors, and swaptions (using a lognormal
(CDO) tranche spreads [16] and pure jump models
approximation), whose prices depend on the entries
with state-dependent jump intensity (“‘local Levy”
of the covariance matrix . The calibration problem
model) [12]. No such inversion formula is avail-
is to find a symmetric semidefinite positive matrix
able in the case of American options (see American
 ∈ Sym+ (n × n) such that the model prices C 
Options). The Dupire formula (11) has been widely
match market prices
used by practitioners for recovering the local volatil-
ity function from call/put option prices by interpolat-
∀i ∈ I, C  (Ti , Ki ) = C ∗ (Ti , Ki ) (9)
ing in strike and maturity and applying equation (11).
This problem can be recast as a semi-definite However, since equation (11) involves differentiating
programming problem [2]. the inputs, it suffers from instability and sensitivity to
Other examples include the construction of yield small changes in inputs, as shown in Figure 1. This
curves from bond prices (see Bond Options) calibra- instability deters one from using inversion formulas
tion of term structure models (see Term Structure such as equation (6) even in the rare cases where they
Models) to bond prices, recovering the distribution of exist.
volatility from option prices [28] calibration to Amer-
ican options in diffusion models [1] and recovery of
portfolio default rates from market quotes of credit Least-squares Formulation
derivatives [16, 18].
These problems are typically ill-posed in the sense Typically, if the model is misspecified, the observed
that, either solutions may not exist (model class is option prices may not lie within the range of prices
too narrow to reproduce observations) or solutions attainable by the model. Also, option prices are
are not unique (if data is finite or sparse). In practice, defined up to a bid–ask spread: a model may generate
existence of a solution is restored by formulating the prices compatible with the market but may not
problem as an optimization problem exactly fit the mid-market prices for any given θ ∈ E.
For these reasons, one often reformulates calibration
inf F (C θ − C) (10) as a least-squares problem
θ∈E

where E is the parameter space and F is a loss 

I
inf J0 (θ), J0 (θ) = wi |Ci (θ) − Ci |2 (12)
function applied to the discrepancy C θ − C between θ∈E
i=1
market prices and model prices. An algorithm is then
used to retrieve one solution and the main issue is the where Ci are mid-market quotes and wi > 0 are a set
stability of this reconstructed solution as a function of weights, often chosen inversely proportional to the
of inputs (market prices). (squared) bid–ask spread of Ci .
 Model Calibration 3

u1(t,x) s1(t,x)

1 1

0.5 0.5

0 0
1 1
1.5 1.5
0.5 1 0.5 1
t 0 0.5 x 0 0.5

u2(t,x) s2(t,x)

1 1

0.5 0.5

0 0
1 1
1.5 1.5
0.5 1 0.5 1
t 0 0.5 x 0 0.5

Figure 1 Extreme sensitivity of Dupire formula to noise in the data. Two examples of call price function (left) and their
corresponding local volatilities (right). The prices differ through IID noise ∼ U N I F (0, 0.001), representing a bid–ask
spread

In most models, the call prices are computed to its ease of calibration using the Hagan formula
numerically via Fourier transform (see Fourier [30].
Methods in Options Pricing) or by solving a par- In most cases, option prices Ci (θ) depend contin-
tial differential equation (PDE) (see Partial Dif- uously on θ and E is a subset of a finite dimensional
ferential Equations). However, in many situations space (i.e., there are a finite number of bounded
(short or long maturity, small vol–vol, etc.) approx- parameters), so the least-squares formulation always
imation formulae for implied volatilities (Ti , Ki ) admits a solution. However, the solution of equation
of call options are available [5, 10, 11, 30] in (12) need not be unique: J0 may, in fact, have several
terms of model parameters (see Implied Volatility global minima, when the observed option prices do
in Stochastic Volatility Models; Implied Volatility: not uniquely identify the model. Figures 2 and 3 show
Volvol Expansion; Implied Volatility: Long Matu- examples of the function J0 for some popular para-
rity Behavior; SABR Model). In these situations, metric option pricing models, computed using a data
parameters are calibrated by a least-squares fit to the set of DAX index options prices on May 11, 2001.
approximate formula: The pricing error in the Heston stochastic volatil-
ity model (see Heston Model), shown in figure as

I
a function of the “volatility of volatility” and the
inf wi |(Ti , Ki ; θ) −  ∗ (Ti , Ki )|2 (13)
θ∈E mean reversion rate, displays a line of local minima.
i=1
The pricing error for the variance gamma model (see
An example is the SABR model (see SABR Variance-gamma Model) in Figure 3 displays a non-
Model), whose popularity is almost entirely due convex profile, with two distinct minima in the range
4 Model Calibration

Pricing error in heston model: SP500 options data, 2000.

Log error
6

3
5
M
ea
n
re 10
ve
rs 1.5
ion 1
pa 0.5
ram 15 y
0 of volatilit
et
er Volatility

Figure 2 Error surface for the Heston stochastic volatility model, DAX options

× 105
2
1.8
1.6
1.4
1.2 A
1
0.8 0.25
0.6
8 0.2
7
6
5 0.15 s
4
3
k 2 1
0 0.1

Figure 3 Error surface for variance gamma (pure jump) model, DAX options

of observed values. These examples show that, even term, to the pricing error and solve the auxiliary
if the number of observations (option prices) is much problem:
higher than the number of parameters, this does not inf Jα (θ) (14)
imply identifiability of parameters. θ∈E

Regularization methods can be used to overcome

where
this problem [27]. A common method is to have
a convex penalty term R, called the regularization Jα (θ) = J0 (θ) + αR(θ) (15)
 Model Calibration 5

The functional (16) consists of two parts: the regu- This regularized formulation has the advantage
larization term αR(θ) which is convex in its argument that its solution exhibits continuous dependence on
and the quadratic pricing error which measures the market prices and with respect to the choice of the
precision of calibration. The coefficient α, called prior model [21, 22].
regularization parameter, defines the relative impor- Simpler regularization methods can be used in
tance of the two terms: it characterizes the trade- settings where prices are computed using analytical
off between prior knowledge and the information transform methods. Belomestny & Reiss [8] pro-
contained in option prices. Jα (.) is usually minimized pose a spectral regularization method for calibrating
by gradient-based methods, where the crux of the exponential-Lévy models. Aspremont [3] formulates
algorithm is an efficient computation of the gradient the calibration of LIBOR market models (Exam-
∇θ J . ple 3) as semidefinite programming problems under
When parameter is a function (such as the local constraints.
volatility function), the regularization term is often Different regularization terms select different solu-
chosen to be a smoothness (e.g., Sobolev) norm. tions: Tikhonov regularization approximates the least-
This method, called Tikhonov regularization (see squares solution with smallest norm [27] while
Tikhonov Regularization) has been applied to diffu- entropy-based regularization selects the minimum-
sion models [1, 2, 13, 23, 26] and to exponential-Lévy entropy least-squares solution [22].
models [19].
Another popular choice of regularization term is
the relative entropy (see Entropy-based Estimation) Entropy Minimization Under Calibration
R(θ) = H (
θ |) with respect to a prior probabil- Constraints
ity measure . In continuous-time models, relative
entropy can be used as regularization criterion only An alternative approach to regularization is to select a
if the prior possesses a nonempty class of equiva- pricing model
by minimizing the relative entropy
lent martingale measures, that is, it corresponds to an (see Entropy-based Estimation) of the probability
incomplete market model (see Complete Markets). measure
with respect to a prior, under calibration
From a calibration perspective, market incomplete- constraints
ness (i.e., the nonuniqueness of equivalent martingale
measure) is therefore an advantage: it allows to con-
inf H (
|) under Ci = E
[Hi ] for i ∈ I
ciliate compatibility with option prices and equiva-
∼
lence with respect to a reference probability measure. (17)
Examples are provided by jump processes (see Jump
Processes; Exponential Lévy Models) or reduced- Relative entropy being strictly convex, any solu-
form credit risk models (see Reduced Form Credit tion of equation (17) is unique and can be computed
Risk Models): one can modify the jump size distri- in a stable manner using Lagrange multiplier (dual)
bution (Lévy measure) or the default intensity while methods [24] (see Convex Duality).
preserving equivalence (see Equivalence of Prob- Application of these ideas to a set of scenarios
ability Measures) of measures [18, 20]. For Lévy leads to the weighted Monte Carlo algorithm (see
processes (see Exponential Lévy Models), the rela- Weighted Monte Carlo) [6]: one first simulates N
tive entropy term H (ν) is computable in terms of the sample paths
N = {ω1 , ..ωN } from a prior model
Lévy measure ν [21]. The calibration problem then  and then solves the above problem (AV) using
takes the following form: as prior the uniform distribution on
N . The idea
is to weight the paths in order to verify the cali-
Problem 2 Given a prior Lévy process with law
0 bration constraints. The weights (
N (ωi ), i = 1..N )
and characteristics (σ0 , ν0 ), find a Lévy measure ν are constructed by minimizing relative entropy under
which minimizes calibration constraints

N
Jα (ν) = αH (ν) + wi (C0ν (Ti , Ki ) − C0 (Ti , Ki ))2 
N

N (ωi )
i=1 inf
N (ωi ) ln under
(16)
N ∈P(
N )
i=1
N (ωi )
6 Model Calibration


N programming techniques. Consider a Markovian

N (ωi )Gj (ωi ) = Cj (18) model where the state variable St (asset price, interest
i=1 rate,..) follows a stochastic differential equation
This constrained optimization problem is solved
by duality [6, 24]: the dual has an explicit solution, in dXt =µθ (t) dt + σθ (t, St ) dWt
the form of a Gibbs–Boltzmann measure [4, 6] (see 
Entropy-based Estimation). A (discounted) payoff + γθ (t, Xt− )µ(dt dz) (21)
X is then priced using the same set of simulated paths
via
where W is a Wiener process and µ a com-
pensated Poisson random measure with intensity

N
E
N [X] =
N (ωi )X(ωi ) νθ (dz)λθ (t) dt. The coefficients of the model are
i=1
parameterized by some parameter θ ∈ E; in a non-
parametric setting, θ is just the coefficient itself and
1 
N (ωi )
N
E is a functional space. Denote the law of the solution
= X(ωi ) (19)
N i=1 N (ωi ) by
θ . Consider now the case where the calibration
criterion J (.)
 Tcan be expressed as an expected value
The benchmark payoffs (calibration instruments) J (θ) = E θ [ 0 φ(Xt ) dt] with a strictly convex func-
play the role of biased control variates, leading to tion φ(.). A classical approach to solve the calibration
variance reduction [29]: problem
 

I I
inf J (θ), under E θ [Hi ] = Ci (22)

N
N
E [X] = E X− αi Hi + αi Ci (20) θ∈E
i=1 i=1
is to introduce the Lagrangian functional
This method yields as a by-product, a static
hedge portfolio αi∗ , which minimizes the variance in 
L(θ, λ) = J (θ) − λi (E θ [Hi ] − Ci )
equation (20) [3, 6, 17].
i∈I
A drawback is that the martingale property is  
lost in this process since it would correspond to an T 
infinite number of constraints. As a result, derivative =E θ
φ(Xt ) dt − λi (Hi − Ci )
0 i∈I
prices computed with the weighted Monte Carlo
algorithm may fail to verify arbitrage relations across (23)
maturities (e.g. calendar spread relations), especially
when applied to forward-starting contracts. where λi is the Lagrange multiplier associated to the
These arbitrage constraints can be restored by calibration constraint for payoff Hi . The dual problem
representing
as a random mixture of martingales associated to the constrained minimization problem
the law of random mixture being chosen via relative (22) is given by
entropy minimization under calibration constraints
 T
[17]. This results in an arbitrage-free version of the
weighted Monte Carlo approach, which is applied inf L(θ, λ) = inf E θ φ(Xt ) dt
θ∈E θ∈E 0
to recovering covariance matrices implied by index 
options in [15]. − λi (Hi − Ci ) (24)
i∈I

Stochastic Control Methods

It can be viewed as a stochastic control problem
In certain continuous-time models, the relative (see Stochastic Control) with running cost φ(t, Xt )
entropy minimization approach can be mapped, and terminal cost .
via a duality argument, into a stochastic control This original formulation of the calibration prob-
problem, which can then be solved using dynamic lem was first presented by Avellaneda et al. [7] in the
 Model Calibration 7

context of diffusion model with unknown volatility models that are compatible with the market data
(Cibid , Ciask )i=1..I . An evolutionary algorithm simu-
dSt = St σ (t, St ) dWt (25) lates an inhomogeneous Markov chain (Xn )n≥1 in
E N which undergoes mutation–selection cycles [9]
The calibration criterion in [7] was chosen to be
designed such that as the number of iterations n
 T  grows, the components (θ1N , ..., θnN ) of Xn converge
J (σ ) = E σ dt η(σ 2 (t, Xtσ )) (26) to the Gδ , yielding a population of points (θk ) which
0
converges to a sample of model parameters compati-
where η is a strictly convex function. Duality between ble with the market data (Cibid , Ciask )i=1..I in the sense
(22) and (24) is not obvious in this case since the that J0 (θk ) ≤ δ. We thus obtain a population of N
Lagrangian is not convex with respect to its argument model parameters calibrated to market data, which
[31]. The stochastic control approach can also be can be different especially if the initial problem has
applied in the context of model calibration by relative multiple solutions.
entropy minimization for classes of models where Figure 4 shows a sample of local volatility func-
absolute continuity is preserved under a change of tions obtained using this approach [9]. These exam-
parameters, such as models with jumps. Cont and ples illustrate that precise reconstruction of local
Minca [18] use this approach for retrieving the default volatility from call option prices is at best illusory; the
rate in a portfolio from CDO tranche spreads indexed parameter uncertainty is too important to be ignored,
on the portfolio. especially for short maturities where it does not affect
the prices very much; short-term volatility hovers
anywhere between 15% and 30%. These observa-
Stochastic Algorithms tions cast a doubt on the volatility content of very
short-term options in terms of volatility and ques-
Objective functions used in calibration (with the
tions whether one can solely rely on short maturity
exception of entropy-based methods) are typically
asymptotics (see SABR Model) in model calibration.
nonconvex, event after regularization, leading to
multiple minima and lack of convergence in gradient-
based methods. Stochastic algorithms known as
evolutionary algorithms, which contain simulated Parameter Uncertainty
annealing as a special case, have been widely used for
global nonconvex optimization are natural candidate Model calibration is usually the first step in a pro-
for solving such problems [9]. cedure whose ultimate purpose is the pricing and
Suppose, for instance, we want to minimize the hedging of (exotic) options. Once the model param-
pricing error eter θ is calibrated to market prices, it is used to
compute a model-dependent quantity f (θ)—price of

I
an exotic option or a hedge ratio—using a numerical
J0 (θ) = wi |Ciθ − Ci |, θ ∈ E (27)
procedure. Given the ill-posedness of the calibration
i=1
problem and the resulting uncertainty on the solution
where Ciθ are model prices and Ci are observed θ, one question is the impact of this uncertainty on
(transaction or mid-market) prices for the benchmark such model-dependent quantities. This aspect is often
options. Now define the a priori error level δ as neglected in practice and many users of pricing mod-
els view the calibrated parameter as fixed, equating

I
calibration with a curve-fitting exercise.
δ= wi |Cibid − Ciask | (28)
Particle methods yield, as a by-product, a way to
i=1
analyze model uncertainty. While calibration algo-
Given the uncertainty on option values due to rithms based on deterministic optimization yield a
bid–ask spreads, one cannot meaningfully distin- point estimate for model parameters, particle meth-
guish a “perfect” fit J0 (θ) = 0 from any other fit ods yield a population Q = {
θ1 , ...,
θk } of pricing
with J0 (θ) ≤ δ. Therefore, all parameter values in models, all of which price the benchmark options
the level set Gδ = {θ ∈ E, J0 (θ) ≤ δ} correspond to with equivalent precision E
(Hi ) ∈ [Cibid , Ciask ]. The
8 Model Calibration

Confidence intervals for local volatility : DAX options.

0.35

0.3

0.25

0.2

0.15

0.1
0.5
1
0.2
1.5 0.15
2 0.1
S/S0 0.05
2.5 0 t

Figure 4 A sample of local volatility surfaces calibrated to DAX options

heterogeneity of this population reflects the uncer- with a portfolio containing αi units of benchmark
tainty in model parameters, which are left undeter- instrument Hi ,
mined by the benchmark options. This idea can be 
exploited to produce a quantitative measure of model H = α0 + αi Hi (30)
i∈I
uncertainty compatible with observed market prices 
of benchmark instruments [14], by considering the the cost α0 + αi Ci of setting up the hedge is
interval of prices automatically equal to the model price E
[H ].
  Calibration does not entail that prices, hedge
ratios, or risk parameters generated by the model
inf E
[X], sup E
[X] (29)

∈Q
∈Q
are “correct” in any sense. This requires a correct
model specification with realistic dynamics for risk
for a payoff X in the various calibrated models. factors. Indeed, many different models may calibrate
Another approach is to calibrate several different the same prices of, say, a set of call options but lead
models to the same data and compare the value to very different prices of hedge ratios for exotics
of the exotic option across models [14, 32]. Model [14, 32]. For example, any equity volatility smile can
uncertainty in derivative pricing is further discussed be reproduced by a one-factor diffusion model (see
in [14]. Example 1) via an appropriate specification of the
local volatility surface, but there is ample evidence
that volatility itself should be modeled as a risk factor
Relation with Pricing and Hedging (see Stochastic Volatility Models) and a one-factor
diffusion may lead to an underestimation of volatility
Calibrating a model to market prices simply ensures risk and unrealistic dynamics [30].
that model prices of benchmark instruments reflect However, a model that is not calibrated to market
current “mark-to-market” values. It also ensures that prices of liquidly traded derivatives is typically not
the cost of a static hedge (see Static Hedging) using easy to use. For example, even if a payoff can
these benchmark instruments is correctly reflected in be statically hedged with traded derivatives using
model prices: if a payoff H can be statically hedged an initial capital V0 , the model price will not be
 Model Calibration 9

equal to V0 . Thus, model prices will, in general, [14] Cont, R. (2006). Model uncertainty and its impact on the
be inconsistent with hedging costs if the model is pricing of derivative instruments, Mathematical Finance
not calibrated. Thus, calibration seems a necessary 16(3), 519–547.
[15] Cont, R. & Deguest, R. (2009). What do index options
but not sufficient condition for choosing a model for
imply about the dependence among stock returns? Col-
pricing and hedging. umbia University Financial Engineering Report 2009-
06,[Link].
[16] Cont, R., Deguest, R. & Kan, Y.H. (2009). Default
References Intensities Implied by CDO Spreads: Inversion Formula
and Model Calibration. Columbia University Financial
Engineering Report 2009-04, [Link].
[1] Achdou, Y. (2005). An inverse problem for a parabolic [17] Cont, R. & Léonard, Ch. (2008). A Probabilistic
variational inequality arising in volatility calibration Approach to Inverse Problems in Option Pricing. Work-
with American options, SIAM Journal on Control and
ing Paper.
Optimization 43, 1583–1615.
[18] Cont, R. & Minca, A. (2008). Recovering Portfolio
[2] Achdou, Y. & Pironneau, O. (2002). Volatility smile
Default Intensities Implied by CDO Tranches. Financial
by multilevel least square, International Journal of
Engineering Report 2008-01, Columbia University.
Theoretical and Applied Finance 5(2), 619–643.
[19] Cont, R. & Rouis, M. (2006). Recovering Lévy Processes
[3] d’Aspremont, A. (2005). Risk-management methods for
from Option Prices by Tikhonov Regularization. Working
the Libor market model using semidefinite program-
Paper.
ming, Journal of Computational Finance 8(4), 77–99.
[20] Cont, R. & Tankov, P. (2004). Financial Modelling with
[4] Avellaneda, M. (1998). The minimum-entropy algorithm
Jump Processes, Chapman and Hall/CRC Press, Boca
and related methods for calibrating asset-pricing mod-
Raton.
els, Proceedings of the International Congress of Math-
[21] Cont, R. & Tankov, P. (2004). Nonparametric calibration
ematicians, Documenta Mathematica, Berlin, Vol. III,
of jump-diffusion option pricing models, Journal of
pp. 545–563.
Computational Finance 7(3), 1–49.
[5] Avellaneda, M., Boyer-Olson, D., Busca, J. & Friz, P.
(2002). Reconstructing the smile, Risk Magazine [22] Cont, R. & Tankov, P. (2005). Recovering Lévy pro-
October. cesses from option prices: regularization of an ill-posed
[6] Avellaneda, M., Buff, R., Friedman, C., Grandchamp, N., inverse problem, SIAM Journal on Control and Opti-
Kruk, L. & Newman, J. (2001). Weighted Monte Carlo: mization 45(1), 1–25.
a new technique for calibrating asset-pricing mod- [23] Crépey, S. (2003). Calibration of the local volatility in
els, International Journal of Theoretical and Applied a trinomial tree using Tikhonov regularization, Inverse
Finance 4, 91–119. Problems 19, 91–127.
[7] Avellaneda, M., Friedman, C., Holmes, R. & Sam- [24] Csiszár, I. (1975). I-divergence geometry of probability
peri, D. (1997). Calibrating volatility surfaces via distributions and minimization problems, The Annals of
relative entropy minimization, Applied Mathematical Probability 3, 146–158.
Finance 4, 37–64. [25] Dupire, B. (1994). Pricing with a smile, Risk 7, 18–20.
[8] Belomestny, D. & Reiss, M. (2006). Spectral calibration [26] Engl, H. & Egger, H. (2005). Tikhonov regulariza-
of exponential Lévy Models, Finance and Stochastics tion applied to the inverse problem of option pricing:
10(4), 449–474. convergence analysis and rates, Inverse Problems 21,
[9] Ben Hamida, S. & Cont, R. (2004). Recovering volatility 1027–1045.
from option prices by evolutionary optimization, Journal [27] Engl, H.W., Hanke, M. & Neubauer, A. (1996). Reg-
of Computational Finance 8(3), 43–76. ularization of Inverse Problems, Mathematics and its
[10] Berestycki, H., Busca, J. & Florent, I. (2004). Comput- Applications, Kluwer Academic Publishers, Dordrecht,
ing the implied volatility in stochastic volatility mod- The Netherlands, Vol. 375.
els, Communications on Pure and Applied Mathematics [28] Friz, P. & Gatheral, J. (2005). Valuing Volatility Deriva-
57(10), 1352–1373. tives as an Inverse Problem, Quantitative Finance,
[11] Bouchouev, I., Isakov, V. & Valdivia, N. (2002). Recov- December 2005.
ering a volatility coefficient by linearization, Quantita- [29] Glasserman, P. & Yu, B. (2005). Large sample prop-
tive Finance 2, 257–263. erties of weighted Monte Carlo estimators, Operations
[12] Carr P., Geman H., Madan D.B. & Yor M. (2004). Research 53(2), 298–312.
From local volatility to local Lévy models, Quantitative [30] Hagan, P., Kumar, D., Lesniewski, A.S. & Wood-
Finance 4(5), 581–588. ward, D.E. Managing smile risk, Wilmott Magazine
[13] Coleman, T., Li, Y. & Verma, A. (1999). Reconstructing September, 84–108.
the unknown volatility function, Journal of Computa- [31] Samperi, D. (2002). Calibrating a diffusion model with
tional Finance 2(3), 77–102. uncertain volatility, Mathematical Finance 12, 71–87.
 10 Model Calibration

[32] Schoutens, W., Simons, E. & Tistaert, J. (2004). A per- Related Articles
fect calibration! Now what? Wilmott Magazine March.
Black–Scholes Formula; Convex Duality; Dupire
Further Reading Equation; Entropy-based Estimation; Exponential
Lévy Models; Implied Volatility in Stochastic
Biagini, S. & Cont, R. (2006). Model-free representation of Volatility Models; Implied Volatility: Large Strike
pricing rules as conditional expectations, in Stochastic Pro-
Asymptotics; Jump Processes; Local Volatility
cesses and Applications to Mathematical Finance, J. Aka-
hori, S. Ogawa and S. Watanabe, eds, World Scientific,
Model; Markov Functional Models; SABR Model;
Singapore, pp. 53–66. Stochastic Volatility Models; Weighted Monte
Harrison, J.M. & Pliska, S.R. (1981). Martingales and stochas- Carlo; Yield Curve Construction.
tic integrals in the theory of continuous trading, Stochastic
Processes and their Applications 11, 215–260. RAMA CONT

View publication stats