Zeliade Systems

Model Validation: theory, practice and perspectives

Zeliade White Paper May, 2011
Zeliade Systems

ZWP-0006
Zeliade Systems SAS Zeliade Systems

56 rue Jean-Jacques Rousseau 75001 Paris France

Phone : +(33) 1 40 26 17 29 Fax : +(33) 1 40 26 17 81
e-mail : [email protected]

T ITLE : Model Validation: theory, practice and perspectivesa

N UMBER : ZWP-0006
N UMBER OF PAGES : 14
F IRST VERSION : January, 2010
C URRENT VERSION : May, 2011
R EVISION : 1.0.9
a C. Martini and P. Henaff (Département LUSSI, Télécom Bretagne)

This document must not be published, transmitted or reproduced without permission from Zeliade Systems SAS.
Copyright c 2007–2011 Zeliade Systems SAS. All Rights Reserved.
Trademarks: Zeliade Systems is a registered trademark of Zeliade Systems SAS. All other company and product names referenced in this document are used for identification
purposes only and may be trade names or trademarks of their respective owners.
Legal disclaimer: ZELIADE SYSTEMS SAS AND/OR ITS SUPPLIERS MAKE NO REPRESENTATIONS ABOUT THE SUITABILITY OF THE INFORMATION CONTAINED
IN THIS DOCUMENT FOR ANY PURPOSE. THIS DOCUMENT IS PROVIDED ”AS IS”, WITHOUT WARRANTY OF ANY KIND. ZELIADE SYSTEMS SAS AND/OR ITS
SUPPLIERS HEREBY DISCLAIM ALL WARRANTIES AND CONDITIONS OF WITH REGARD TO THIS INFORMATION, INCLUDING ALL IMPLIED WARRANTIES AND
CONDITIONS OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, TITLE AND NON INFRINGEMENT. IN NO EVENT SHALL ZELIADE SYSTEMS SAS
BE LIABLE FOR ANY SPECIAL, INDIRECT OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFIT
ARISING OUT OF OR IN CONNECTION WITH THE USE OR PERFORMANCE OF INFORMATION AVAILABLE IN THIS DOCUMENT.
Contents

The Nature of Model Risk 1

Model Validation teams 2

Qualitative Assessment of Model Risk 4

Derman classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
When everything else fails . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

Empirical Assessment of Model Risk 6

Dynamic hedging simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
The calibration conundrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
The need for simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

Quantification of Model Risk 8

The robustness of Black Scholes . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
The seminal work of Avellaneda and Lyons on model uncertainty . . . . . . . 9
Recent proposals and open issues . . . . . . . . . . . . . . . . . . . . . . . . . . 9

Conclusion 11
In July of 2009, the Basel Committee on Banking Supervision issued a directive [21]
requiring that financial institutions quantify model risk. The Committee further stated
that two types of risks should be taken into account: “The model risk associated with
using a possibly incorrect valuation, and the risk associated with using unobservable
calibration parameters”. The resulting adjustments must impact Tier I regulatory cap-
ital, and the directive must be implemented by the end of 2010.
On the surface, this seems to be a simple adjustment to the market risk framework,
adding model risk to other sources of risk that have already been identified within
Basel II. In fact, quantifying model risk is much more complex because the source of
risk (using an inadequate model) is much harder to characterize. Twelve months away
from the deadline, there is no consensus on this topic. There is fortunately a growing
body of literature, both from the academia and the industry, and the purpose of this
paper is to summarize the development of the notion of “model risk” and present the
current state of the art, before outlining open issues that must be resolved in order to
define a consistent framework for measuring model risk.

The Nature of Model Risk

Financial assets can be divided into two categories. In the first category, we find the
assets for which a price can be directly observed in the financial market place. These are
the liquid assets for which there are either organized markets (e.g. Futures exchanges)
or a liquid OTC market (e.g. interest rate swaps). For the vast majority of assets,
however, price cannot be directly observed, but needs to be inferred from observable
prices of related instruments. This is typically the case for financial derivatives whose
price is related to various features of the primary assets, depending on a model. This
process is known as “marking-to-model”, and involves both a mathematical algorithm
and subjective components, thus exposing the process to estimation error. Beyond this
generic observation, the notion of Model Risk has been interpreted in at least three
manners.

• A first interpretation of Model Risk stems from the observation that various mod-
els can be calibrated to perfectly price a set of liquid instruments, but produce in-
consistent estimates for an exotic product. If one accepts that there is one “true”
model, then model risk refers to the risk of mis-specification.

• A second interpretation focuses on the operational use of a model, that is used not
only to compute a price, but equally importantly to compute risk indicators for
dynamic hedging. In a perfect world where the true process for each risk factor
is known, and where hedge instruments are readily available, we know that to

c Zeliade Systems SAS 1/14 ZWP-0006– May, 2011

 each derivative corresponds a dynamic replicating portfolio. Thus “model-risk”
can be assessed by observing the hedging error, i.e. the discrepancy between the
payoff of a derivative and the value of the replicating portfolio.

• Well publicized events in the banking industry have highlighted a third manifes-
tation of Model Risk. When liquidity is low, how should a product be “marked to
model” so as to minimize the risk of discrepancy when a transaction can finally
be observed? To quote Rebonato [23]:

Model risk is the risk of occurrence of a significant difference between

the mark-to-model value of a complex and/or illiquid instrument, and
the price at which the same instrument is revealed to have traded in the
market.

In summary, Model Risk can lead to both mis-pricing and mis-management of the
hedging strategy. Mis-pricing clearly will have the most spectacular consequences,
but mis-hedging is an equally serious issue.
The importance of Model Risk has gradually emerged in the industry: in first-tier
banks, Model Validation teams have been created as an independent power besides
front-office trading/quant teams. These organizations are now briefly reviewed.

Model Validation teams

These teams are usually organized by asset class (equity, commodities, FX, fixed in-
come in a broad sense, credit, cross-currency). The day-to-day duty of all these model
validation teams is to qualitatively and quantitatively validate the models used by the
front-office. A closer look, however, reveals that their roles vary widely across institu-
tions:

• In its narrowest interpretation, model validation is an almost pure theoretical

and qualitative matter, which requires a solid mathematical finance background
and a good sense for the business domain. It focuses on whether a model or
methodology is adequate for the purpose it is used for by the front-office:

– do not use local volatility models for pricing Cliquet options since they will
be sensitive to the forward volatility which the local volatility models do not
encompass,
– do not use the SABR time-slice approximation for small strikes because it
will yield artificial arbitrage opportunities.

c Zeliade Systems SAS 2/14 ZWP-0006– May, 2011

 • In a broader definition, the role involves the validation not only of the models or
methodologies at hand, but also of their implementations. This involves review-
ing the code provided by the front-office, running simulation stress tests, and
developing alternative implementations as a cross-check. This new dimension
brings a fresh set of challenges however, as the validation team must now deal
with the varied technologies used by the front-office groups (Excel/VBA, Mat-
lab, C++). This challenge is compounded by the fact that these teams may have
a very strong Quant culture, but usually less than a strong IT discipline. If they
decide to develop an in-house analytics library, they will often rely on third-party
components.

• A still wider interpretation of the role is to validate not only the model imple-
mentation, but the whole process where the model comes into play:

– What the model is used for: either getting a delta from a calibrated price
for short-term arbitrage, or providing a reference price for an exotic product
which is to stay in the book; whether the complex product at hand is hedged
in various ways with indicators provided by the model. In fact the right
point of view is very often to validate not a model, but a strategy performed
by a market actor which involves model-computed quantities at some stage.
A typical example in the Vanna-Volga strategy, mostly used on FX markets,
which is not a model in the traditional sense.
– The market data and other data sources on which the strategy depends: do
you calibrate your models on the right data? For instance there are several
different fields in Bloomberg for an option price: do you use the right one?
Regarding dividends and repos on equity markets, where does your data
come from? How do you handle implied dividends and/or repos stemming
from the call/put parity in force on the vast majority of equity indices?
– At which horizon do you use the model which have been calibrated on 3-
years market data? How is the model extrapolated?
– At which frequency do you re-calibrate your model? How do you handle
the discrepancy between the implied volatility/correlation markets and the
historical ones?
– What happens in case of a Credit event for one of the counterparties?

The academic literature has also recognized the issue early on, starting with contri-
butions by Merton and Scholes themselves [18], shortly after the publication of the
Black-Scholes model. Academic contributions to the analysis of Model Risk can be or-
ganized into three strands of literature. The first strand provides a typology of the

c Zeliade Systems SAS 3/14 ZWP-0006– May, 2011

risks involved, and has been used by regulators and auditors to define best practice.
The second strand focuses on empirical tests of model robustness, while the most re-
cent current attempts to define a measure that would be homogeneous to market risk,
and thus could be directly used to define capital requirement. These contributions are
summarized in the next 3 sections.

Qualitative Assessment of Model Risk

Derman classification

An early description of Model Risk is due to E. Derman [8]. In this paper, Derman lists
6 types of risk:

1. Inapplicability of modeling: a mathematical model may not be relevant to de-

scribe the problem at hand.

2. Incorrect model: the risk of using a model that does not accurately describe the
reality being modeled. This is the most common interpretation of Model Risk.

3. Correct model, incorrect solutions, leading, for example, to inconsistent pricing

of related claims.

4. Correct model, inappropriate use: the risk related to an inaccurate numerical

solution of an otherwise correct model. For example, the risk related to Monte-
Carlo calculations with too few simulations.

5. Badly approximated solutions. This risk appears when numerical methods are
used to solve a model.

6. Software and hardware bugs.

7. Unstable data. Financial data is of notoriously poor quality. Models must be

robust with respect to errors in input data.

Derman’s recommendations for coping with model risk are mostly organizational: use
a comprehensive approach to model validation, test exhaustively, create good inter-
faces to minimize user error, etc. This typology of model risk is the foundation of the
processes used by auditors and consulting firms to validate the use of models inside
financial institutions. It also forms the framework, and provides a series of tests to be
used for assessing model risk from a legal perspective [17]. Most importantly, these
categories have inspired regulators. In the paper “Supervisory guidance for assessing

c Zeliade Systems SAS 4/14 ZWP-0006– May, 2011

banks’ financial instrument fair value practices” of April 2009 [20], the Basel Commit-
tee on Banking Supervision issues recommendations on controlling model and opera-
tional risk. The elements of a “rigorous model validation” are described as follows:

Validation includes evaluations of:

• The model’s theoretical soundness and mathematical integrity.

• The appropriateness of model assumptions, including consistency with
market practices and consistency with relevant contractual terms of
transactions.
• Sensitivity analysis performed to assess the impact of variations in
model parameters on fair value, including under stress conditions.
• Benchmarking of the valuation result with the observed market price
at the time of valuation or independent benchmark model.

It further states the need to assess a valuation adjustment that will reflect model risk
([21], 718(cxi-1)):

For complex products including, but not limited to, securitisation ex-
posures and n-th-to-default credit derivatives, banks must explicitly assess
the need for valuation adjustments to reflect two forms of model risk: the
model risk associated with using a possibly incorrect valuation methodol-
ogy; and the risk associated with using unobservable (and possibly incor-
rect) calibration parameters in the valuation model.

Auditors finally use the same categories to describe best-practice guidelines [22].

When everything else fails

One area where the qualitative assessment of model risk is of major importance is
that of very illiquid assets. This covers most of the structured Credit asset-back-type
securities, like ABS and RBMS and their derivatives, as well as hyper-exotic tailor-
made hybrid products. In such cases it is already a challenge to identify the various
risk factors. Among these the vast majority will be non-tradeable so that the pure
model price is the only benchmark. The model parameters will often be very roughly
calibrated to dubious-and-not-always-meaningful data.
In such situations, a parsimonious model with economically meaningful parameters
is an adequate tool (cf [12]). The qualitative behavior of the product payoff and of
the product price with respect to the model parameter is then key to the assessment

c Zeliade Systems SAS 5/14 ZWP-0006– May, 2011

of Model Risk . More quantitative tools can be used; yet only large simulation engines
will produce reliable quantities in many circumstances - the input of which will mostly
be stylised, qualitative parameters.
Model risk is not the ultimate clue to illiquidity: especially in that context, Model Risk
needs to be assessed as part of a comprehensive risk management policy. In less ex-
treme contexts, the framework outlined above provides a guideline for spelling out
compliance requirements, but does not provide a quantitative measure of Model Risk.
How to provide such a measure is the topic of the research reviewed in the next section.

Empirical Assessment of Model Risk

There is a large body of literature on the empirical assessment of Model Risk, start-
ing with contributions by Merton and Scholes themselves [18]. The general idea is to
simulate the pricing and hedging policy of a derivative product under a number of
assumptions in order to provide an empirical measure of the associated risk.

Dynamic hedging simulation

In 1985, Rubinstein [25] used this simulation method to document a major challenge
to the Black-Scholes model, which is the empirical observation of a “volatility smile”.
Among many other publications, we note the paper by Bakshi, Cao and Chen [3] which
performs a comprehensive assessment of ten years of development in pricing models
for equity derivatives, covering stochastic volatility models, models with jumps and
with stochastic interest rate. Interestingly, they find that the most sophisticated model
is not the most effective tool for hedging:

Overall, incorporating stochastic volatility and jumps is important for

pricing and internal consistency. But for hedging, modeling stochastic volatil-
ity alone yields the best performance.

Also of interest is the range of tests that the authors apply to compare models:

1. Internal consistency of implied parameters/volatility with relevant time-series

data,

2. Out-of-sample pricing, and

3. Quality of dynamic hedging.

c Zeliade Systems SAS 6/14 ZWP-0006– May, 2011

Note that the validation of a model through dynamic hedging simulation is what Jar-
row and Yildirim suggest too in the context of the inflation markets in their celebrated
paper [24].

The calibration conundrum

Following the development of the local volatility model, which allows to calibrate all
the available vanilla quotes with a single model (which may have a complex functional
form), a perfect model calibration to a set of vanilla options has become the expected
standard in the industry.
This apparent success, however, still leaves a number of fundamental issues open:

1. The vanilla quotes on a given day are only a part of the available information,
which contains at least the history of the underlying, and also the history of the
vanilla quotes.

2. The vanilla quotes are in theory equivalent to the marginal distributions of the
underlying under the market implied probability measure. This said, they con-
tain no information on the joint laws of the underlying process, whereas most of
the payoffs of exotic structured products will be highly sensitive to functionals
related to these joint laws.

The first issue is partially dealt with in mixed historical/market calibration procedures
where an historical prior is used in the market calibration strategy. Nevertheless, there
is plenty of room for research on robust calibration algorithms that exploit the whole
information available.
The second issue has been discussed in deep empirical calibration studies, mostly con-
ducted on the equity market. In [27], Schoutens and alii study the relative behavior
of several models which calibrate equally well to the vanilla quotes, and show that
they yield vastly different results when pricing path dependent or moment derivatives
such as variance swaps. The same kind of work has been performed by Dilip Madan
and Ernst Eberlein in [11], who have pioneered the class of Sato processes as a natural
choice for the joint dynamics underlying the primary asset process.

The need for simulation

The distribution of the Profit and Loss arising from a dynamic hedging simulation, or
the range of prices for a structured product within various ’perfectly’ calibrated models
are of primary interest to empirically assess Model Risk.

c Zeliade Systems SAS 7/14 ZWP-0006– May, 2011

The easier it is to perform such simulations, the better. The challenge is to get a re-
usable simulation infrastructure, modular enough to design quickly new tests, alto-
gether with a clean reporting. Such a simulation facility is already very useful in a
purely simulated world alleviated from the technicalities of real-life contracts (with
calendaring, market conventions, market data issues): in fact many phenomenon will
be more easily isolated, tackled, hopefully understood, in this simplified environment.
Yet it becomes an invaluable tool if it can handle real-life contracts and market data.
A true back-testing simulation infrastructure, in so far as it comes after the simplified
prototyping step, might be the ultimate tool in some sense. Very few systems offer
such facilities, at least with the required degree of modularity and transparency.
Regarding back-testing procedures, the pertaining choice of the time interval on which
performing the simulation is a crucial matter. Beyond this issue, there will always be
a single path available in the history of the underlying; the simulation of a pertaining
richer set of paths in a parsimonious manner (bootstrapping) is a field under active
study, where Markov Chain Monte Carlo methods give promising results (cf [4], [7]).
Simulation studies provide the most comprehensive assessment of model risk. Yet,
they do not address an important requirement of the Basel directive, that is the need for
a normalized measure of model risk that is homogeneous to other risk measures. How
to create this measure, that would be in currency unit, is a recent but active research
topic which is next reviewed.

Quantification of Model Risk

The robustness of Black Scholes

The mis-specification risk of the Black-Scholes model has been well understood since
the late 70’s: if the actual volatility that is experienced while delta hedging a derivative
is lower than the pricing volatility, then the replicating portfolio will under-perform
the derivative, and vice-versa. This result was expanded in a rigorous analysis by El
Karoui and al. [13]. The authors show that for Call and Puts, and more generally for
convex payoffs, the option seller will make a profit as soon as the selling and hedging
volatility is larger than the actual realized volatility, which can bear any kind of stochas-
ticity. Moreover, the Profit can be computed: it will be proportional to the accumulation
of the running option Gamma times the difference of the squared volatilities.
This property has been known in practice for a long time and it has played a crucial
role in the practical usage of the Black-Scholes formula. The robustness of Gaussian
hedging strategies in a fixed income context has been studied in [1].

c Zeliade Systems SAS 8/14 ZWP-0006– May, 2011

Therefore the Model Risk is well understood and quantified in the case of the Black-
Scholes pricing/hedging strategies for convex payoffs like Calls, Puts, or convex com-
binations of those.

The seminal work of Avellaneda and Lyons on model uncertainty

This very nice robustness property does not hold in general for exotic or non-convex
payoffs (as a basic call-spread). In 2 simultaneous seminal works, Avellaneda et al.
[16] and T.Lyons [15] solved the related problem of finding out the cheapest (resp.
highest) price of a super-replicating (resp. sub-replicating) strategy given the fact that
the volatility is unknown - with the assumption that it lies in a given range.
This Uncertain Volatility Model is in some sense the right generalization of the Black-
Scholes model, in so far as it extends the previous robustness property to any kind
of payoff. This comes to some cost of course: the corresponding ”super price” solves
a (fully) non-linear PDE. Even if can be efficiently solved in a basic trinomial tree al-
gorithm, this non linearity is a fundamental matter and the management in the UVM
framework of a book of derivatives is a quite tricky matter.
This largely accounts for the relatively cold reception made to the UVM model by the
practitioners, coupled with the fact that the ”super price” is in general very high as a
consequence of the generalized robustness property. Note that it seems that the UVM
approach applied to correlation uncertainty (as presented in [26]) is used for pricing
purposes in practice.
Even though the UVM ”super price” may not be pertaining for trading purposes, yet
the UVM ”super price”, or more precisely the UVM ”super price” minus the ”sub
price” is a very pertaining model risk indicator.

Recent proposals and open issues

More recently, the Basel II market risk framework ([21], 718(cxi) and 718(cxii)) explicitly
mandates a valuation adjustment, that impacts Tiers I regulatory capital, to account for
model risk. Therefore, one needs to design a method for measuring model risk that is
both a coherent risk measure, and can be expressed in currency terms. To frame the
issue, it is useful to consider the analogy with the VaR method for computing market
risk. The calculation of VaR involves two steps:

• The identification of the market risk factors and the estimation of the dynamic of
these risk factors: the classical VaR framework assumes a multivariate log-normal
distribution for asset prices

c Zeliade Systems SAS 9/14 ZWP-0006– May, 2011

 • The definition of a risk measure, for example the 99.5% confidence interval for a
10-day holding period.

What would be the equivalent when considering Model Risk? In this case, the risk fac-
tors include the risk of model mis-specification (leaving out important sources of risk,
mis-specifying the dynamic of the risk factors), and the risk of improper calibration,
even though the chosen model may be perfectly calibrated to a set of liquid instru-
ments. The second step involves defining a reasonable family of models over which
the risk should be assessed. The family of models is restricted to the models that can
be precisely calibrated to a set of liquid instruments. This constraint alone still defines
such a large set of models that further restrictions need to be applied. Intuitively, one
needs to define a meaningful notion of “distance” between models, in order to define
a normalized measure of Model Risk. Recent research is attempting to provide a rig-
orous answer to these challenging questions. Rama Cont [5] frames the problem as
follows.

• Let I be a set of liquid instruments, with Hi∈I being the corresponding payoffs,
∗
and Ci∈I the mid-market prices, with Ci∗ ∈ [Cibid , Ciask ].

• Let Q be a set of models, consistent with the market prices of benchmark instru-
ments:
Q ∈ Q ⇒ E Q [Hi ] ∈ [Cibid , Ciask ], ∀i ∈ I (1)

Define next the upper and lower price bounds over the family of models, for a payoff
X:

π = sup E Qj [X] π = inf E Qj [X]

j=1,...,n j=1,...,n

The risk measure is finally the range of values caused by model uncertainty:

µQ = π(X) − π(X)

Note the analogy with the UVM framework: in fact the UVM Model Risk indicator
sketched in the previous section perfectly fits this framework, with Q the set of models
with a volatility in a given range. The crucial aspect of this procedure is the definition
of a suitable set Q. There are many ways of defining it:

• Choose a class of models, and construct a set of models by varying some un-
observable parameters, while ensuring that each model calibrates to the set of
benchmark instruments.

c Zeliade Systems SAS 10/14 ZWP-0006– May, 2011

 • Select several types of model (local volatility, stochastic volatility, etc), and cali-
brate each model to the same set of benchmark instruments.

The first approach is followed by Dumont and Lunven [10]. They tested the method on
basket options, priced in a multivariate log-normal framework. The model is calibrated
to vanilla options on each element of the basket. This leaves the correlation undeter-
mined. The authors compute an historical distribution of this correlation matrix, and
sample it to construct the set Q. The optimization is finally performed heuristically.
The second approach is followed in [27], [11] where numerous mainstream academic
equity models are dealt with. As expected [9], model mis-specification induces more
variability than calibration error.
It is clear that the variability of models forming the set Q needs to be normalized. In
the same way as one computes “99.5% VaR for a 10 day holding period”, one needs
a normalizing factor to qualify model risk. In other words, one needs to define the
aforementioned notion of “distance” between models. A possible approach could be
found in [2],[6] and [19], where the relative entropy between two processes is used
to regularize the calibration problem. For payoffs that are not path dependent, one
can measure the relative entropy of the two distributions at expiry. This provides a
measure of distance that is independent of the nature of the underlying process.
Contrasting with the above analytical approach, Kerkhof and al. [14] develop a method
for accounting for Model Risk, using a data-driven method: They estimate an empiri-
cal distribution of model outcome, using historical data. The same information is used
to compute market risk and model risk, and model risk is expressed as an adjustment
(multiplicative factor) to the VaR estimate.
As this brief survey suggests, the quantification of Model Risk is under active study. As
with the other sources of risk, the crucial issue is to define a normalized measure that
will enable a comparison among exotic products, and ultimately among asset classes
and financial institutions.

Conclusion

Model validation encompasses a variety of disciplines and issues. The role of model
validation teams has been recognized to be crucial in the recent turmoil. They can
tackle at the same time the validation of well-defined components and the validation
of a whole business process and meanwhile contribute to a much better understanding
of the risk profile of the strategies conducted by the front office.
Under the pressure of the regulators, Model Validation teams will be asked to compute
standardized Model Risk indicators, spanning all asset classes, in a documented and

c Zeliade Systems SAS 11/14 ZWP-0006– May, 2011

reproducible manner. This is a significant technological challenge. To meet this chal-
lenge, the model validation teams will need to be equiped with a technology platform
that includes:

• Model libraries that can be validated independently of the librairies created by

the front office (i.e. developed independently by the Model Validation team,
sourced from third party vendors, or both)

• A simulation framework that enables to quickly test new models, payoffs and
strategies in a standardized and reproducible manner.

More broadly, the derivatives industry will ultimately need to develop a set of stan-
dardized methods for measuring Model Risk. Given the complexity of such measures,
their credibility will require an independent calculation framework with the ability
to provide transparent, reproducible and auditable results, for internal Model Validation
teams, auditors and regulators. This will ease the way towards the practice of mean-
ingful prices and sane risk policies with a quantified impact of model uncertainty.

References

[1] L. Schlogl A. Dudenhausen, E. Schlogl. Robustness of gaussian hedges and the

hedging of fixed income derivatives. Research Paper Series 19, Quantitative Fi-
nance Research Centre, University of Technology, Sydney, 1999.

[2] Marco Avellaneda, Craig Friedman, Richard N. Holmes, and Dominick J. Sam-
peri. Calibrating volatility surfaces via relative-entropy minimization. SSRN eLi-
brary, 1996.

[3] Gurdip Bakshi, Charles Cao, and Zhiwu Chen. Empirical performance of alterna-
tive option pricing models. The Journal of Finance, 52(5):2003–2049, 1997.

[4] C.Andrieu, N.DeFreitas, A.Doucet, and M.I.Jordan. An introduction to mcmc for

machine learning. Machine Learning, 2003.

[5] Rama Cont. Model uncertainty and its impact on the pricing of derivative instru-
ments, August 2004.

[6] Rama Cont and Peter Tankov. Calibration of Jump-Diffusion Option Pricing Mod-
els: A Robust Non-Parametric Approach. SSRN eLibrary, 2002.

[7] P. Del Moral. Feynman-Kac formulae. Genealogical and interacting particle systems with
applications. (Probability and its Applications). Springer Verlag, New York, 2004.

c Zeliade Systems SAS 12/14 ZWP-0006– May, 2011

 [8] Emmanuel Derman. Model risk. Goldman, Sachs & Co. Equity Product Strategy,
pages –, April 1996.

[9] Kai Detlefsen and Wolfgang Hardle. Calibration risk for exotic options in the
heston model.

[10] Philippe Dumont and Christophe Lunven. Risque de modèle et options multi
sous-jacents, 2005.

[11] D.Madan E. Eberlein. Sato processes and the valuation of structured products.
2007.

[12] Frederic Patras Jean-Pierre Lardy and Francois-Xavier Vialard. Correlation, cdos
of abs and the subprime crisis, in: Financial risks: New developments in struc-
tured product and credit derivatives. Economica (to appear).

[13] Nicole E. Karoui, Monique Jeanblanc-Picquè, and Steven E. Shreve. Robustness

of the black and scholes formula. Mathematical Finance, 8(2):93–126, 1998.

[14] Jeroen Kerkhof, Bertrand Melenberg, and Hans Schumacher. Model risk and cap-
ital reserves, August 2009.

[15] T.J. Lyons. Uncertain volatility and the risk-free sysnthesis of derivatives. Applied
Mathematical Finance, 1995.

[16] A. Paras M. Avellaneda, A. Levy. Pricing and hedging derivative securities in

markets with uncertain volatilities. Applied Mathematical Finance, 1995.

[17] Dr. Susan M. Mangiero. Financial model risk looms large. The Investment Lawyer,
9(11), November 2002.

[18] Robert C. Merton, Myron S. Scholes, and Mathew L. Gladstein. The returns and
risk of alternative call option portfolio investment strategies. The Journal of Busi-
ness, 51(2):183, 1978.

[19] Suhas Nayak and George Papanicolaou. Stochastic volatility surface estimation.
May 2006.

[20] Basel Committee on Banking Supervision. Supervisory guidance for assessing

banks’ financial instrument fair value practices - final paper.

[21] Basel Committee on Banking Supervision. Revisions to the basel ii market risk
framework - final version. July 2009.

[22] Richard Pace. The state of model risk management. Bank Accounting & Finance,
pages –, July 2008.

c Zeliade Systems SAS 13/14 ZWP-0006– May, 2011

[23] R. Rebonato. Theory and practice of model risk management. Modern Risk Man-
agement: A History, RiskWaters Group, London, pages 223–248–223–248, 2002.

[24] Y.Yildirim R.Jarrow. Pricing treasury inflation protected securities and related
derivative securities using an hjm model. Journal of Financial and Quantitative Anal-
ysis, 38(2), 2003.

[25] Mark Rubinstein. Nonparametric tests of alternative option pricing models using
all reported trades and quotes on the 30 most active cboe option classes from
august 23, 1976 through august 31, 1978. The Journal of Finance, 40(2):455–480,
1985.

[26] S.Romagnoli and T.Vargiolu. Robustness of the black-scholes approach in the case
of options on several assets. Finance and Stochastics, 2000.

[27] Jurgen Tistaert Wim Schoutens, Erwin Simons. A perfect calibration! now what?
Wilmott magazine, 2004.

c Zeliade Systems SAS 14/14 ZWP-0006– May, 2011