Value-at-Risk Model Risk

Carol Alexander
Chair of Risk Management,
ICMA Centre, Henley Business School at the University of Reading, Reading,
RG6 6BA, UK.
Email: [Link]@[Link]
Phone: +44 118 3786431

José Marı́a Sarabia

Professor of Statistics,
Department of Economics, University of Cantabria, Spain
Email: [Link]@[Link]
Phone: +34 942 201635

The second author acknowledges support from the Ministerio de Educación of

Spain (PR2009-0200, Programa Nacional de Movilidad de Recursos Humanos, and
SEJ2007-65818) and would like to thank the ICMA Centre for its hospitality.

Electronic copy available at: [Link]

 Value-at-Risk Model Risk

Abstract

Large banks assess their regulatory capital for market risk using complex, firm-wide
Value-at-Risk (VaR) models. In their ‘bottom-up’ approach to VaR there are many
sources of model risk. A recent amendment to banking regulations requires addi-
tional market risk capital to cover all these model risks but, as yet, there is no
accepted framework for computing such an add-on. We introduce a top-down ap-
proach to quantifying VaR model risk in a rigorous statistical framework and derive
a corresponding adjustment to regulatory capital that is relatively straightforward
to implement.

Key Words: Basel II; Maximum entropy; Model risk; Quantile; Risk capital; Value-at-
Risk (VaR)

JEL Codes: C1, C19, C51, G17, G28

Electronic copy available at: [Link]

1 Introduction
The term ‘model risk’ is commonly applied to encompass various sources of uncertainty
in statistical models. Following Cairns (2000) we distinguish two sources of model risk:1

• Model choice, i.e. inappropriate assumptions about the form of the statistical model
for the random variable;

• Parameter uncertainty, i.e. estimation error in the parameters of the chosen model.2

This paper focuses on the model risk of quantile risk assessments with particular reference
to ‘Value-at-Risk’ (VaR) estimates, which are derived from quantiles of portfolio profit
and loss distributions. VaR corresponds to an amount that could be lost, with a specified
probability, if the portfolio remains unmanaged over a specified time horizon. It has
become the global standard for assessing risk in all types of financial firms: in fund
management, where portfolios with long-term VaR objectives are actively marketed; in
the treasury divisions of large corporations, where VaR is used to assess position risk;
and in insurance companies, who measure underwriting and asset management risks in a
VaR framework. But most of all, banking regulators remain so confident in VaR that its
application to computing market risk capital for banks, used since the 1996 amendment
to the Basel I Accord,3 will soon be extended to include stressed VaR under an amended
Basel II and the new Basel III Accords.4
The finance industry’s reliance on VaR has been supported by decades of academic
research. Especially during the last ten years there has been an explosion of articles
published on this subject – see Christoffersen (2009) for a survey. A popular topic is the
introduction of new VaR models,5 and another very prolific strand of literature focuses
on testing their accuracy.6 However, the stark failure of many banks to set aside sufficient
capital reserves during the banking crisis of 2008 sparked an intense debate on using
1
There is no consensus on the sources of model risk. For instance, Cont (2006) points out that both
these sources could be encompassed within a universal model, and Kerkhof et al. (2010) distinguish
‘identification risk’ as an additional source.
2
This includes sampling error ( parameter values can never be estimated exactly because only a finite
set of observations on a random variable are available) and optimization error (e.g. different numerical
algorithms typically produce slightly different estimates based on the same model and the same data).
3
See Basel Committee on Banking Supervision (1996).
4
See Basel Committee on Banking Supervision (2009).
5
Historical simulation (Hendricks, 1996) is the most popular approach amongst banks (Perignon and
Smith, 2010) but data-intensive and prone to pitfalls (Pritsker, 2006). Other popular VaR models assume
normal risk factor returns with the RiskMetrics covariance matrix estimates (RiskMetrics, 1997). More
complex VaR models are proposed by Hull and White (1998b), Mittnik and Paolella (2000), Christoffersen
et al. (2001), Ventner and de Jongh (2002), Angelidis et al. (2004), Hartz et al. (2006), Kuan et al.
(2009) and many others.
6
The coverage tests introduced by Kupiec (1995) are favoured by banking regulators, and these are
refined by Christoffersen (1998), Christoffersen et al. (2001) and Christoffersen and Pelletier (2004).
However Berkowitz et al. (2010) demonstrate that more sophisticated tests such as the conditional
autoregressive test of Engle and Manganelli (2004) may perform better.

1
VaR models for the purpose of computing the market risk capital requirements of banks.
Turner (2009) is critical of the manner in which VaR models have been applied and Taleb
(2007) even questions the very idea of using statistical models for risk assessment. Despite
the warnings of Turner, Taleb and other critics of VaR models,7 most financial institutions
continue to employ them as their primary tool for market risk assessment and economic
capital allocation.
For internal, economic capital allocation purposes VaR models are commonly built
using a ‘bottom-up’ approach. That is, VaR is first assessed at an elemental level, e.g.
for each individual trader’s positions, then is it progressively aggregated into desk-level
VaR, and VaR for larger and larger portfolios, until a final VaR figure for a portfolio that
encompasses all the positions in the firm is derived. This way the traders’ limits, and
risk budgets for desks and broader classes of activities can be allocated within a unified
framework. However, this bottom-up approach introduces considerable complexity to the
VaR model for a large bank. Indeed, it could take more than a day to compute the
full (often numerical) valuation models for each product over all the simulations in a VaR
model. Yet, for regulatory purposes VaR must be computed at least daily, and for internal
management intra-day VaR computations are frequently required.
To reduce complexity in the internal VaR system simplifying assumptions are com-
monly used, in the data generation processes for financial assets and interest rates and in
the valuation models used to mark complex products to market every day. For instance,
it is very common to apply normality assumptions in VaR models, along with lognormal,
constant volatility approximations for exotic options prices and sensitivities.8 Of course,
there is conclusive evidence that financial asset returns are not well represented by nor-
mal distributions. However, the risk analyst in a large bank may be forced to employ this
assumption for pragmatic reasons.
Another common choice is to base VaR calculations on simple historical simulation.
Many large commercial banks have legacy systems that are only able to compute VaR
using this approach, commonly basing calculations on at least 3 years of daily data for
all traders’ positions. Thus, some years after the credit and banking crisis vastly over-
inflated VaR estimates could still be produced by these models, even though markets
have returned to normal. The implicit and simplistic assumption that history will repeat
itself with certainty – that the banking crisis will recur within the risk horizon of the VaR
model – may well seem absurd to the analyst, yet he is constrained by the legacy system
to compute VaR using simple historical simulation.
7
Even the use of a quantile has been subject to criticism, as it is not necessarily sub-additive, and
related metrics such as conditional VaR may be preferred. See Beder (1995) and Artzner et al. (1999).
8
Indeed, model risk frequently spills over from one business line to another, e.g. normal VaR models
are often employed in large banks simply because they are consistent with the geometric Brownian motion
assumption that is commonly applied for option pricing and hedging.

2
 Thus, financial risk analysts are often constrained to employ a model that does not
comply with their beliefs about the data generation processes for financial returns, and
the data that they think is inappropriate.9 As a result they have long recognized model
risk due to model choice and parameter uncertainty. In the literature, Derman (1996),
Simons (1997), Crouhy et al. (1998), Green and Figlewski (1999), Kato and Yoshiba
(2000) and Rebonato (2001) all identify these two main causes of model risk in finance.
However, in those papers a formal, quantitative definition of model risk that allows its
assessment is elusive.
Surprizingly few papers deal explicitly with VaR model risk. Early work by Jorion
(1996) and Talay and Zheng (2002) investigated sampling error and Brooks and Persand
(2002) assessed the effect of applying different GARCH models to estimate VaR. At the
time of writing the only other published paper in this important area is by Kerkhof et
al. (2010). It addresses the VaR model risk stemming from both model choice and
parameter uncertainty and, by quantifying the adjustment to VaR that is necessary for
normal i.i.d. and GARCH models to pass regulatory backtests, the authors derive an
incremental market risk capital charge (henceforth simply called the risk capital ‘add-on’)
to cover VaR model risk. This issue is very important because recent revisions to the
Basel II market risk framework include the requirement that banks set aside additional
reserves to cover all sources of model risk in the internal models used to compute the
market risk capital charge.10
This paper introduces a new framework for measuring quantile model risk with an
elegant, intuitive and practical method for computing the risk capital add-on to cover VaR
model risk. In the model development section we employ the general ‘quantile’ rather than
‘VaR’ terminology because our methodology has potential applications to many areas, not
only to finance.11 In addition to the computation of a model risk ‘add-on’ for a given VaR
model and given portfolio,12 our approach can be used to assess which, of the available
VaR models, has the least model risk (relative to a given portfolio). Similarly, given a
specific VaR model, our approach can assess which portfolio has the least model risk.
9
Banking regulators recommend 3-5 years of data for historical simultation and require at least 1 year
of data for constructing the covariance matrices used in other VaR models.
10
See Basel Committee on Banking Supervision (2009), Section IV.
11
Quantile estimation is commonly applied to a variety of disciplines: a survey of their applictions to
insurance, actuarial science, hydrology and several other fields as well as finance is given by Reiss and
Thomas (1997). The probability of the occurrence of extreme events is of prime interest for actuaries,
where heavy-tailed distributions are used to model large claims and losses (for instance, see Matthys et
al., 2004); heavy-tailed distributions are frequently used for quantile estimation in hydrology and climate
change (Mkhandi et al. (1996), Katz et al. (2002)) in statistical process control for computing capability
indices (Anghel, 2001), for measuring efficiency (Wheelock and Wilson, 2008) and for reliability analysis
(Unnikrishnan Nair and Sankaran, 2009). The uncertainty surrounding quantile-based risk assessments
in these areas has long been recognised (see Peng and Qi, 2006).
12
A portfolio could be at any level of granularity, e.g. an individual trader’s positions, or all positions
in a desk, or the entire daily P&L of a bank.

3
 In the following: the benchmark for assessing model risk is discussed in Section 2;
Section 3 gives a formal definition of quantile model risk and outlines a framework for
its quantification. We present a statistical model for the probability α̂ that is assigned,
under the benchmark distribution, to the α quantile of the model distribution. Our
idea is to endogenize model risk by using a distribution for α̂ to generate a distribution
for the quantile. The mean of this model-risk-adjusted quantile distribution detects any
systematic bias in the model’s α quantile, relative to the α quantile of the benchmark
distribution. A suitable quantile of the model-risk-adjusted distribution determines an
uncertainty buffer which, when added to the bias-adjusted quantile gives a model-risk-
adjusted quantile that is no less than the α quantile of the benchmark distribution at a pre-
determined confidence level, this confidence level corresponding to a penalty imposed for
model risk; Section 4 presents an empirical example on the application of our framework
to VaR model risk, in which the degree of model risk is controlled by simulation; Section
5 summarizes and concludes.

2 The Benchmark
Model risk in finance has been approached in two different ways: examining all feasible
models and evaluating the discrepancy in their results, or specifying a benchmark model
against which model risk is assessed. Papers on the quantification of valuation model risk
in the risk-neutral measure exemplify each approach: Cont (2006) quantifies the model
risk of a complex product by the range of prices obtained under all possible valuation
models that are calibrated to market prices of liquid (e.g. vanilla) options; Hull and
Suo (2002) define model risk relative to the implied price distribution, i.e. a benchmark
distribution implied by market prices of vanilla options. In the context of VaR model risk
the benchmark approach, which we choose to follow, is more practical than the former.
Some authors identify model risk with the departure of a model from a ‘true’ dynamic
process: see Branger and Schlag (2004) for instance. Yet, outside of an experimental or
simulation environment, we never know the ‘true’ model for sure. In practice, all we can
observe are realisations of the data generation processes for the random variables in our
model. It is futile to propose the existence of a unique and measurable ‘true’ process
because such an exercise is beyond our realm of knowledge.
However, we can observe a maximum entropy distribution (MED). This is based on
a ‘state of knowledge’, i.e. no more and no less than the information available regarding
the random variable’s behaviour. This information includes the observable data that are
thought to be relevant plus any subjective beliefs. Shannon (1948) defined the entropy of

4
a probability density function g(x), x ∈ R as

H(g) = −Eg [log g(x)] = − g(x) log g(x)dx.
R

This is a measure of the uncertainty in a probability distribution and its negative is a

measure of information.13 The maximum entropy density is the function f (x) that maxi-
mizes H(g), subject to a set of conditions on g(x) which capture the testable information
that is available.14 The criterion here is to be as vague as possible (i.e. to maximize
uncertainty) given the constraints imposed by the testable information. This way, the
maximum entropy distribution (MED) represents no more (and no less) than this infor-
mation. If the testable information consists only of a historical sample on X of size n
then, in addition to the normalization condition, there are n conditions on g(x), one for
each data point. In this case, the MED is just the empirical distribution based on that
sample. Otherwise, the testable information consists of fewer conditions, which capture
only that sample information which is thought to be relevant, and any other conditions
imposed by subjective beliefs about the MED.
From the banking regulator’s perspective what matters is not the ability to aggregate
and disaggregate VaR in a bottom-up framework, but the adequacy of a bank’s total
market risk capital reserves, which are derived from the aggregate market VaR. Therefore,
regulators only need to define a benchmark VaR model to apply to the bank’s aggregate
daily profit and loss (P&L). This model will be the MED of the regulator, i.e. the model
that best represents the regulator’s state of knowledge regarding the accuracy of VaR
models. Our recommendation is that banks assess their VaR model risk by comparing
their aggregate VaR figure, which is typically computed using the ‘bottom-up’ approach,
with the VaR obtained using the regulator’s MED in a ‘top-down’ approach, i.e. calibrated
directly to the bank’s aggregate daily trading P&L. Typically this P&L contains marked-
to-model prices for illiquid products, in which case their valuation model risk is not
quantified in our framework.
Following the theoretical work of Shannon (1948), Zellner (1977), Jaynes (1983) and
many others it is common to assume the testable information is given by a set of moment
functions derived from a sample, in addition to the normalization condition. When only
the first two moments are deemed relevant, the MED is a normal distribution (Shannon,
13
For instance, if g is normal with variance σ 2 , H(g) = 21 (1 + log(2π) + log(σ)), so the entropy increases
as σ increases and there is more uncertainty and less information in the distribution. As σ → 0 and the
density collapses the Dirac function at 0, there is no uncertainty but −H(g) → ∞ and there is maximum
information. However, there is no universal relationship between variance and entropy and where their
orderings differ entropy is the superior measure of information. See Ebrahimi, Maasoumi and Soofi (1999)
for further insight.
14
A piece of information is testable if it can be determined whether F is consistent with it. One of
piece of information is always a normalization condition.

5
1948). More generally, when the testable information contains the first N sample mo-
ments, f (x) takes an exponential form. This is found by maximizing entropy subject to

the conditions µn = R xn g(x)dx for n = 0, . . . , N, where µ0 = 1 and
 µn (n = 1, ...,
 N) are
n=N
the moments of the distribution. The solution is f (x) = exp − n=0 λn xn where
the parameters 0 , . . . λn are obtained by solving the system of non-linear equations
 λ
 n
µn = x exp − n=N n=0 λn x
n
dx, for n = 0, . . . , N.
Rockinger and Jondeau (2002), Wu (2003) and several others have applied a four-
moment MED to various econometric and risk management problems. Park and Bera
(2009) and Chan (2009a) apply a four-moment MED to GARCH models and Chan (2009b)
extends this to the computation of VaR. But, perhaps surprizingly, none of these papers
consider the tail weight that is implicit in the use of a four-moment MED. In fact, moment-
based MEDs are only well-defined when N is even. For any odd value of N there will
be an increasing probability weight in one of the tails. Also, the four-moment MED has
lighter tails than a normal distribution, due to the presence of the term exp[−λ4 x4 ] with
non-zero λ4 in f (x). Indeed, the more moments included in the conditions, the thinner
the tail of the MED. But financial asset returns are typically heavy-tailed and it is likely
that this property will carry over to a bank’s aggregate daily P&L. Hence, we cannot
advocate the use of moment-based MEDs.
Nevertheless, there are advantages in choosing a parametric MED for the regulator’s
benchamrk. VaR estimates are quantiles of a forward-looking P&L distribution, but
to base model parameter estimates entirely on historical data limits beliefs about the
future to what has been experienced in the past. Parametric distributions are frequently
advocated for VaR estimation because the parameter values estimated from historical data
may be changed subjectively to accomodate beliefs about the future PL distribution.
The flexible class of generalized beta generated (GBG) distributions introduced by
Alexander et al. (2010) have three parameters that offer direct control over peakness, skew
and relative tail weights. The GBG distribution is defined by replacing the uniform U[0, 1]
distribution in the probability integral transform by a generalized beta distribution (of the
first kind, introduced by McDonald, 1984) denoted GB(a, b, c). This may be characterized
by its density function
fGB (u; a, b, c) = B(a, b)−1 [cuac−1 (1 − uc )b−1 ], 0 < u < 1, a, b, c > 0. (1)
Given any continuous parent distribution F (x), x ∈ R with density f (x), X has a GBG
distribution when the probability transformed variable U = F (X) has density (1). Then
the distribution of X may be characterised by its density function:
fGBG (x; a, b, c) = B(a, b)−1 f (x)[cF (x)ac−1 (1 − F (x)c )b−1 ], x ∈ R. (2)
GBGs are MEDs under just three, fairly general shape constraints, so an advantage of
using a GBG for the benchmark VaR model is that the parameter values that are estimated

6
from historical data may be adjusted subjectively, to reflect specific beliefs about the
peakness, skew and relative tail weights of future P&L. Another reason for our interest
in GBG distributions is that they will become central to our analysis in Section 3.
Following the study by Berkowitz and O’Brien (2002) on the aggregate performance
of VaR models for large commercial banks, regulators may prefer to specify a bench-
mark model for VaR, calibrated to aggregate daily P&L, from the familiar GARCH class.
Berkowitz and O’Brien found that most ‘bottom-up’ internal VaR models produced VaR
estimates that were too large, and insufficiently risk-sensitive, compared with the GARCH
VaR estimates that were derived using only the aggregate daily P&L. Filtered historical
simulation (FHS) of aggregate daily P&L would be another popular alternative, espe-
cially when combined with a volatility filtering that increases its risk sensitivity: see
Barone-Adesi et al. (1998) and Hull and White (1998a). Alexander and Sheedy (2008)
demonstrated empirically that GARCH volatility filtering combined with histroical sim-
ulation can produce very accurate VaR and conditional VaR estimates, even at extreme
quantiles. By contrast, the standard historical simulation approach failed many of their
backtests.

3 Quantile Model Risk

The α quantile of a continuous distribution F of a real-valued random variable X with
range R is denoted15
qαF = F −1 (α). (3)
In our statistical framework F is identified with the unique MED based on a state of
knowledge K which contains all testable information on F . We characterise a statistical
model as a pair {F̂ , K̂} where F̂ is a distribution and K̂ is a filtration which encompasses
both the model choice and its parameter values. The model provides an estimate F̂ of F ,
and uses this to compute the α quantile. That is, instead of (3) we use
qαF̂ = F̂ −1 (α). (4)

Quantile model risk arises because {F̂ , K̂} = {F, K}. Firstly, K̂ =
 K, e.g. K may include
the belief that only the last six months of data are relevant to the quantile today; yet K̂
may be derived from an industry standard that must use at least one year of observed data
in K̂;16 and secondly, F̂ is not, typically, the MED even based on K̂, e.g. the execution of
firm-wide VaR models for a large commerical bank may present such a formidable time
15
In practice, the probability α is often predetermined. Frequently it will be set by senior managers
or regulators and small or large values corresponding to extreme quantiles are very commonly used. For
instance, regulatory market risk capital is based on VaR models with α = 1% and a risk horizon of 10
trading days.
16
As is the case under current banking regulations for the use of VaR to estimate risk capital reserves
- see Basel Committee on Banking Supervision (1996).

7
challenge that F̂ is based on simplified data generation processes, as discussed in the
introduction.
In the presence of model risk the α quantile of the model is not the α quantile of the
MED, i.e. qαF̂ = qαF . The model’s α quantile qαF̂ is at a different quantile of F and we use
the notation α̂ for this quantile, i.e. qαF̂ = qα̂F , or equivalently,

α̂ = F (F̂ −1 (α)). (5)

In the absence of model risk α̂ = α for every α. Otherwise, we can quantify the extent of
model risk by the deviation of α̂ from α, i.e. the distribution of the quantile probability
errors
e(α|F, F̂ ) = α̂ − α. (6)

If the model suffers from a systematic, measurable bias at the α quantile then the mean
error ē(α|F, F̂ ) should be significantly different from zero. A significant and positive
(negative) mean indicates a systematic over (under) estimation of the α quantile of the
MED. Even if the model is unbiased it may still lack efficiency, i.e. the dispersion of
e(α|F, F̂ ) may be high. Several measures of dispersion may be used to quantify the
efficiency of the model. In section 4 we use the the root mean squared error (RMSE) but
the standard deviation of the errors, the mean absolute error (MAE) or the range are
common alternatives.17
We now regard α̂ = F (F̂ −1 (α)) as a random variable with a distribution that is
generated by our two sources of model risk, i.e. model choice and parameter uncertainty.
Because α̂ is a probability it has range [0, 1]; so we may approximate its distribution by
any distribution with support [0, 1]. The most general distribution of this type is the
generalized beta distribution GB(a, b, c) which has density (1).
Assuming α̂ ∼ GB(a, b, c), the α quantile of our model, adjusted for model risk,
becomes a GBG random variable:
Q(α|F, F̂ ) = F −1 (α̂), α̂ ∼ GB(a, b, c). (7)
That is, the model-risk-adjusted quantile Q(α|F, F̂ ) is a random variable whose GBG
distribution is generated from the MED F and the parameters of the generalized beta
representation of α̂. If GB denotes the distribution of GB(a, b, c), the distribution function
of Q(α|F, F̂ ) is
GF (v; a, b, c) = Pr(F −1 (α̂) ≤ v) = GB (F (v); a, p, q), v ∈ R, (8)
17
If the model has a significant bias the MAE, RMSE and range should be applied with caution because
they include the bias. Anyway, when the bias is large it is better that the model be inefficient: the worst
possible case is an efficient but biased model.

8
and, denoting by f the density of F , the mean E[Q(α|F, F̂ )] of Q(α|F, F̂ ) is given by

−1
E[Q(α|F, F̂ )] = B(a, b) c vf (v)F (v)ac−1[1 − F (v)c ]b−1 dv. (9)
R

This mean quantifies any systematic bias in the quantile estimates: e.g. if the MED has
heavier tails than the model then extreme quantiles qαF̂ will be biased: if α is close to zero
then E[Q(α|F, F̂ )] > qαF and if α is close to one then E[Q(α|F, F̂ )] < qαF . This bias can
be removed by adding the difference qαF − E[Q(α|F, F̂ )] to the model’s α quantile qαF̂ so
that the bias-adjusted quantile has expectation qαF .
Unfortunately, analytic expressions for E[Q(α|F, F̂ )] are only available for some par-
ticular choices of F and some values of a, b and c. However, it is always possible to
obtain an approximate expression. Upon expanding F −1 (α̂) in a Taylor series around the
point α̂m = E(α̂), we obtain the following a series expansion for F −1 (α̂) in terms of its
derivatives F −1(i) , for i = 1, . . . , 4:
1
F −1 (α̂) ≈ F −1 (α̂m ) + F −1(1) (α̂m )(α̂ − α̂m ) + F −1(2) (α̂m )(α̂ − α̂m )2
2
1 −1(3) 1
+ F (α̂m )(α̂ − α̂m )3 + F −1(4) (α̂m )(α̂ − α̂m )4 . (10)
6 24

Taking expectations in (10) yields

1
E[Q(α|F, F̂ )] ≈ F −1 (α̂m ) + F −1(2) (α̂m )σα̂2
2
1 −1(3) 1
+ F (α̂m )γ1 (α̂)σα̂3 + F −1(4) (α̂m )γ2 (α̂)σα̂4 , (11)
6 24

where σα̂2 = E[(α̂ − α̂m )2 ], γ1 (α̂) = σα̂−3 E[(α̂ − α̂m )3 ], and γ2 (α̂) = σα̂−4 E[(α̂ − α̂m )4 ].
The bias-adjusted α quantile estimate could still be far away from the maximum
entropy α quantile: the more dispersed the distribution of Q(α|F, F̂ ), the greater the
potential for qαF̂ to deviate from qαF . Because risk estimates are typically constructed to
be conservative, we introduce an uncertainty buffer to the bias-adjusted α quantile by
adding a quantity equal to the difference between the mean of Q(α|F, F̂ ) and G−1 F (y), the
y% quantile of Q(α|F, F̂ ), to the bias-adjusted α quantile estimate. This way, we become
(1 − y)% confident that the model-risk-adjusted α quantile is no less than qαF .
Finally, our point estimate for the model-risk-adjusted α quantile becomes:
bias adjustment uncertainty buffer
   
qαF̂ + {qαF − E[Q(α|F, F̂ )]} +{E[Q(α|F, F̂ )] − G−1 F̂ F −1
F (y)} = qα + qα − GF (y). (12)

The total model risk adjustment to the quantile estimate is thus qαF − G−1
F (y), and the
computation of E[Q(α|F, F̂ )] could be circumvented if the decomposition into bias and
uncertainty components is not required.

9
 The confidence level 1 − y reflects a penalty for model risk which is a matter for
subjective choice. When X denotes daily P&L and α is small (e.g. 1%), typically all
three terms on the right hand side of (12) will be negative. But the α% daily VaR is
minus the α quantile, so the model-risk-adjusted VaR estimate is −qαF̂ − qαF + G−1
F (y).
−1 F
The total adjustment to VaR is GF (y) − qα . This will be positive unless, in repeated
observations, VaR estimates are typically much greater than the benchmark VaR. In
that case there should be a negative bias adjustment, and this could be large enough to
outweigh the uncertainty buffer, especially when y is large, i.e. when we require only a
low degree of confidence for the model-risk-adjusted VaR to exceed the benchmark VaR.

4 Empirical Illustration
We now describe an experiment in which a portfolio’s returns are simulated based on
a known data generation process. This allows us to control the degree of VaR model
risk and to demonstrate that our framework yields intuitive and sensible results for the
bias and inefficiency adjustments described above. We also validate the accuracy of our
approximation (11).
It is widely accepted that of all the parsimonious discrete-time variance processes the
asymmetric GARCH class is most useful for capturing the volatility clustering that is
commonly observed in financial asset returns. Thus, we shall assume that our MED for
the returns Xt at time t is N (0, σt2 ), where σt2 follows an asymmetric GARCH process.
First the return xt from time t to t + 1 and its variance σt2 are simulated using:
σt2 = ω + α(xt−1 − λ)2 + βσt−1
2
, xt |It ∼ N (0, σt2), (13)

where ω > 0, α, β ≥ 0, α + β ≤ 1 and It = (xt−1 , xt−2 , . . . ).18 For the simulated returns
the parameters of (13) are assumed to be:
ω = 1.5 × 10−6 , α = 0.04, λ = 0.005, β = 0.95, (14)
and so the steady-state annualized volatility of the portfolio return is 25%.19 Then the
MED at time t is Ft = F (Xt |Kt ), i.e. the conditional distribution of the return Xt given
the state of knowledge Kt , which comprises the observed returns It and the knowledge
that Xt |It ∼ N (0, σt2 ).
At time t, a VaR model provides a forecast F̂t = F̂ (Xt |K̂t ) where K̂ comprises It plus
the model Xt |It ∼ N (0, σ̂t2). We now consider three different models for σ̂t2 . The first
model has the correct choice of model but uses incorrect parameter values: instead of (14)
18
We employ the standard notation α for the GARCH return parameter here; this should not be
confused with the notation α for the quantile of the returns distribution, which is also standard notation
in the VaR model literature.
19
The steady-state variance is σ̄ 2 = (ω + αλ2 )/(1 − α − β) and for the annualization we have assumed
returns are daily, and that there are 250 business days per year.

10
the fitted model is:
σ̂t2 = ω̂ + α̂(xt−1 − λ̂)2 + β̂ σ̂t−1
2
, (15)

with
ω̂ = 2 × 10−6 , α̂ = 0.0515, λ̂ = 0.01, β̂ = 0.92. (16)

The steady-state volatility estimate is therefore correct, but since α̂ > α and β̂ < β the
fitted volatility process is more ‘jumpy’ than the simulated variance generation process. In
other words, compared with σt , σ̂t has a greater reaction but less persistence to innovations
in the returns, and especially to negative returns since λ̂ > λ.
The other two models are chosen because they are commonly adopted by financial
institutions, having been popularized by the ‘RiskMetrics’ methodology introduced by JP
Morgan in the mid-1990’s – see RiskMetrics (1997). The second model uses a simplified
version of (13) with:
ω̂ = λ̂ = 0, α̂ = 0.06, β̂ = 0.94. (17)

This is the RiskMetrics exponentially weighted moving average (EWMA) estimator in

which a steady-state volatility is not defined. The third model is the RiskMetrics ‘Regu-
latory’ estimator in which:
250
1
α̂ = λ̂ = β̂ = 0, ω̂ = x2t−i . (18)
250 i=1
10,000
A time series of 10,000 returns {xt }t=1 is simulated from the ‘true’ model (13) with
parameters (14). Then, for each of the three models defined above we use this time series
to (a) estimate the daily VaR, which when expressed as a percentage of the portfolio value
is given by Φ−1 (α)σ̂t , and (b) compute the probability α̂t associated with this quantile
under the simulated returns distribution Ft = F (Xt |Kt ). Because Φ−1 (α̂t )σt = Φ−1 (α)σ̂t ,
this is given by
σ̂t
α̂t = Φ Φ−1 (α) . (19)
σt
Taking α = 1% for illustration, the empirical densities of the three VaR estimates are
depicted in Figure 1. The horizontal scale is the percentage VaR (i.e. Φ−1 (α)σ̂t ) multiplied
by 100. The AGARCH VaR is less variable than the EWMA VaR, and the Regulatory
VaR has a multi-modal distribution, a feature that results from the lack of risk-sensitivity
of this model.20 Now, for each VaR model, we use the true (i.e. simulated) distribution to
estimate α̂ at every time point in the simulations, using (19). For α = 0.1%, 1% and 5%,
Table 1, reports the mean and standard deviation of α̂ and the RMSE between α̂ and α.
The closer α̂ is to α, the smaller the RMSE and the less model risk there is in the VaR
model. The Regulatory model yields an α̂ with the highest RMSE, for every α, so this
has the greatest degree of model risk. The AGARCH model, which we already know has
20
Specifically, a single extreme negative return makes the VaR jump to a level which is sustained for
exactly 250 days, and then the VaR jumps down again when this return drops out of the in-sample period.

11
 Figure 1: Density of daily VaR estimates (α = 1%).
0.16

0.14 AGARCH
0.12 EWMA

0.1 Regulatory
TRUE
0.08

0.06

0.04

0.02

0
0 1 2 3 4 5 6 7 8

Figure 2: Density of quantile probabilities (α = 1%).

0.25

AGARCH
0.2
EWMA
Regulatory
0.15

0.1

0.05

0
0% 1% 2% 3% 4% 5% 6%

the least model risk of the three, produces a distribution for α̂ that has mean closest to
the true α and the smallest RMSE. These observations are supported by Figure 2, which
depicts the empirical distribution of α̂ and Figure 3, which shows the empirical densities
of the model-risk-adjusted VaR estimates F −1 (α̂) (here the horizontal axis is the same as
Figure 1, i.e. the VaR as a percentage of the portfolio value, multiplied by 100). Both
these figures take α = 1% for illustration.

12
 Figure 3: Density of model-risk-adjusted daily VaR (α = 1%).
0.35

0.3 AGARCH

0.25 EWMA
Regulatory
0.2

0.15

0.1

0.05

0
2 2.5 3 3.5 4 4.5 5 5.5 6

Table 1: Sample statistics for quantile probabilities

AGARCH EWMA Regulatory

0.10% Mean 0.11% 0.16% 0.23%
Stdev 0.0007 0.0013 0.0035
RMSE 0.07% 0.14% 0.37%
1% Mean 1.03% 1.25% 1.34%
Stdev 0.0042 0.0059 0.0119
RMSE 0.42% 0.64% 1.22%
5% Mean 4.97% 5.44% 5.27%
Stdev 0.0103 0.0123 0.0268
RMSE 1.03% 1.31% 2.66%

Table 2: Values of (â, b̂) for the distributions given in Table 1.

α AGARCH EWMA Regulatory

0.10% â 2.35 1.50 0.43
b̂ 2145.39 959.21 188.35
1% â 5.92 4.43 1.24
b̂ 570.98 350.90 91.04
5% â 22.10 18.29 3.61
b̂ 422.04 318.15 64.78

13
 We fit a beta distribution to α̂, for each model, and show that the approximation (11)
is valid. Table 2 reports the method of moment beta parameter estimates, denoted (â, b̂),
that are computed using the means and standard deviations given in Table 1. Since
â << b̂ all beta distributions have a large positive skew. Both â and b̂ are inversely
proportional to α, and to the degree of model risk. Thus, at one extreme we have the
highly-peaked beta density of the AGARCH model when α = 5%, and at the other we
have the ‘L’-shaped density of the Regulatory model when α = 0.1%.
Next we compare the empirical mean with the mean of our beta normal distribution
for the adjusted VaR. This is computed using numerical integration of (9) or via the
analytic approximation (11). All three means are subject to error, albeit in slightly
different ways, but the errors are small relative to the VaR model risk. The numerical
and analytic computations are based on the beta parameter estimates in Table 2. Note
that dφ(z)/dz = −zφ(z) and that the derivatives of Φ−1 (z) in (11) are given by:

Φ−1(1) (z) = φ(Φ−1 (z))−1 ,

Φ−1(2) (z) = Φ−1 (z)[φ(Φ−1 (z))]−2 ,
 2
Φ−1(3) (z) = [1 + 2 Φ−1 (z) ][φ(Φ−1 (z))]−3 ,
 2
Φ−1(4) (z) = Φ−1 (z)[7 + 6 Φ−1 (z) ][φ(Φ−1 (z))]−4 .

Table 3: Mean of the model-risk-adjusted VaR distribution

α Mean AGARCH EWMA Regulatory

0.10% (Emp.) 4.919 4.793 4.961
(9) 4.953 4.832 -
(11) 4.988 5.017 1.292
1% (Emp.) 3.703 3.608 3.735
(9) 3.722 3.603 3.729
(11) 3.710 3.614 4.067
5% (Emp.) 2.618 2.551 2.641
(9) 2.626 2.227 2.643
(11) 2.618 2.552 2.654

The results are reported in Table 3. The approximation (11) is best for the AGARCH
model, which has the smallest degree of model risk. The least accurate approximation is
for the Regulatory VaR model at extreme quantiles, when the bias in the VaR estimates
becomes quite large. At the 0.1% quantile, where the Regulatory model has a value of â
less than 1, it becomes impossible to integrate (9) using standard numerical methods.
A point estimate for model-risk-adjusted VaR (RaVaR, for short) is computed using

14
(12). Both RaVaR and the benchmark VaR (BVaR, for short) depend on the time they
are measured. For illustration, we select a point when the simulated volatility is at its
steady-state value of 25% – so the BVaR is 4.886%, 3.678% and 2.601% at the 0.1%,
1% and 5% levels, respectively. Drawing at random from the points when the simulated
volatility was 25%, we obtain AGARCH, EWMA and Regulatory volatility forecasts of
27.00%, 23.94% and 28.19% respectively.21 These volatilities determine the VaR estimates
that we shall now adjust for model risk.

Table 4: Components of the model risk adjustment

α AGARCH EWMA Regulatory

0.1% Mean 4.919 4.793 4.961
Bias -0.033 0.093 -0.075
Quantile 4.447 4.177 3.912
Uncert. Buffer 0.472 0.615 1.049
1% Mean 3.703 3.608 3.735
Bias -0.025 0.070 -0.056
Quantile 3.348 3.145 2.945
Uncert. Buffer 0.355 0.463 0.789
5% Mean 2.618 2.551 2.641
Bias -0.017 0.050 -0.040
Quantile 2.366 2.224 2.082
Uncert. Buffer 0.252 0.327 0.559

Table 4 summarizes the bias and the uncertainty buffer, for different levels of α, based
on the empirical distribution of Q(α|F, F̂ ).22 It reveals a general tendency for the EWMA
model to slightly underestimate VaR and the other models to slightly overestimate VaR.
Yet the bias is relatively small, since all models assume the same normal form as the
MED and the only difference between them is their volatility forecast. Although the bias
tends to increase as α decreases it is not significant for any model.23 Beneath the bias we
report the 5% quantile of the model-risk-adjusted VaR distributions, since we shall first
compute the RaVaR so that it is no less than the BVaR with 95% confidence.
Following the framework introduced in the previous section we now define:

bias adjustment uncertainty buffer

   
RaVaR(y) = VaR + (BVaR − E[Q(α|F, F̂ )]) +(E[Q(α|F, F̂ )] − G−1
F (y)) .
21
So the AGARCH and Regulatory models overestimated BVaR at this point and the EWMA model
underestimated BVaR; of course, this was not the case every time the simulated volatility was 25%.
22
Similar results based on the fitted distributions are not reported for brevity.
23
Standard errors of Q(α|F, F̂ ) are not reported, for brevity. They range between 0.157 for the
AGARCH at 5% to 0.891 for the Regulatory model at 0.1%, and are directly proportional to the degree
of model risk just like the standard errors on the quantile probabilities given in Table 1.

15
Table 5 sets out the RaVaR computation for y = 5%. The model’s volatility forecasts
are in the first row and the corresponding VaR estimates are in the first row of each cell,
for α = 0.1%, 1% and 5% respectively. The (small) bias is corrected by adding the bias
from Table 4 to each VaR estimate. The main source of model risk here concerns the
potential for a large (positive or negative) errors in the quantile probabilties, i.e. the
dispersion of the densities in Figure 2. To adjust for this we add to the bias-adjusted
VaR an uncertainty buffer equal to the difference between the BVaR and the 5% quantile
given in Table 4.24 This gives the RaVaR estimates shown in the third row of each cell.

Table 5: Computation of 95% RaVaR

α AGARCH EWMA Regulatory

Volatility 27.00% 23.94% 28.19%
0.10% VaR 5.277 4.678 5.509
Bias Adj. VaR 5.244 4.772 5.434
RaVaR 5.716 5.387 6.483
1% VaR 3.972 3.522 4.147
Bias Adj. VaR 3.948 3.592 4.091
RaVaR 4.303 4.055 4.880
5% VaR 2.809 2.490 2.932
Bias Adj. VaR 2.791 2.540 2.892
RaVaR 3.043 2.867 3.451

Since risk capital is a multiple of VaR, the percentage increase resulting from replacing
VaR by RaVaR(y) is:

BVaR − G−1
F (y)
% risk capital increase = . (20)
VaR

The penalty (20) for model risk depends on α, except in the case that both the MED and
VaR model are normal, and on the confidence level (1 − y)%.
Table 6 reports the percentage increase in risk capital due to model risk when RaVaR
is no less than the BVaR with (1 − y)% confidence. As expected, this turns out to be
directly proportional to the degree of model risk in the VaR model. At 95% confidence,
the first row of the table shows that risk capital based on the AGARCH model would be
increased by about 8%, increased by about 15% in the EWMA model and about 17.5% in
the Regulatory model. The other rows in Table 6 give the amount of additional risk captial
that would be required at other reasonable confidence levels. The add-on for VaR model
risk increases with the degree of model risk, and with the degree of confidence that the
regulator requires for the model-risk-adjusted VaR to be at least as great as the benchmark
24
Recall, the BVaR is 4.886%, 3.678% and 2.601% at the 0.1%, 1% and 5% levels, respectively.

16
 Table 6: Percentage increase in risk capital from model risk adjustment of VaR

1−y AGARCH EWMA Regulatory

95% 8.33% 15.15% 17.68%
90% 6.05% 12.62% 14.16%
85% 4.47% 10.99% 11.37%
80% 3.19% 9.57% 8.81%
75% 2.25% 8.38% 6.68%

VaR. Note that the risk capital estimate could require the greatest upwards adjustment
when based on EWMA VaR. Indeed, if we require only a low degree of confidence (80% or
less) that the RaVaR exceeds the BVaR, the additional capital required for a EWMA VaR
model exceeds that required for a Regulatory VaR model. This is because the EWMA
VaR is exceptionally low at the time of the adjustment.

5 Summary and Conclusions

National regulators will very soon require that banks set aside additional reserves to
cover all sources of model risk in the internal models used to compute the market risk
capital charge. This paper concerns VaR model risk, i.e. the risk of producing inaccurate
VaR estimates because of an inapproriate choice of VaR model (e.g. historical simulation
or parametric analytic) and/or inaccuracy in the VaR model parameter estimates. We
present a rigorous statistical methodology that provides an elegant and practical solution
to the problem of quantifying the regulatory capital that should be set aside to cover this
type of model risk.
We argue that there is no better choice of model risk benchmark than a maximum
entropy distribution since, by definition, this embodies the entirety of information and
beliefs, no more and no less. In the context of the model risk capital charge, the benchmark
should be specified by the regulator. Then VaR model risk is assessed using a top-down
approach, by comparing two time series: one for the benchmark VaR, which is derived
directly from the bank’s total daily P&L, and the other for the bank’s aggregate daily VaR
figures, upon which their market risk capital charge is based. The latter is typically derived
using a computationally intensive bottom-up approach, with many approximations and
simplifications being applied.
The main ideas are as follows: in the presence of model risk an α quantile is at a
different quantile of the benchmark model, and has an associated tail probability under
the benchmark that is stochastic. Modelling this uncertainty as generally as possible,
the model-risk-adjusted quantile becomes a generalized beta generated random variable,

17
and its distribution quantifies the bias and uncertainty due to model risk. A significant
model risk bias arises if the VaR estimates produced by the bank for the purposes of
risk capital charge calculation tend to be consistently above or below the VaR estimates
obtained by applying the benchmark model to their aggregate daily P&L. Even when the
bank’s VaR estimates show no bias, an adjustment for uncertainty is required because the
difference between the bank’s VaR and the benchmark VaR could vary considerably over
time. The bias and uncertainty in the VaR model, relative to the benchmark, determine a
risk capital adjustment for model risk whose size will also depend on the confidence level
regulators require for the adjusted risk capital to be no less than the risk capital based
on the benchmark model. An empirical example that considers three VaR models with
controlled degrees of model risk has been used to illustrate the framework.
Further research is required in conjunction with both regulators and banks – on the
development of benchmark VaR models to apply to aggregate P&L, and on backtesting
the model-risk-adjusted estimates for commonly-used VaR models. The extension of our
methodology to other quantile-based metrics used in finance, and to conditional VaR in
particular, should be straightforward.25 The fundamental idea of using a stochastic prob-
ability to generate another distribution would still hold, and the bias and uncertainty of
the generated ditribution could still be used to quantify the effect of model risk. Our work
also has potential applications to many other disciplines where quantile risk assessments
are used, such as insurance, hydrology, climate change, statistical process control and
reliability analysis.

25
Conditional VaR is also called ‘expected tail loss’ or ‘expected shortfall’ by a variety of authors.

18
References
Alexander, C., and Sheedy, E. (2008). Developing a stress testing framework based on market
risk models. Journal of Banking and Finance. 32(10), 2220-2236.

Alexander, C., Cordeiro, G., Ortega, E. and Sarabia, J-M. (2010). Generalized beta generated
distributions. ICMA Centre Discussion Papers in Finance.

Angelidis, T., Benos, A., and Degiannakis, S. (2004). The use of GARCH models in VaR
estimation. Statistical Methodology. 1(2), 105-128.

Anghel, C. (2001). Statistical process control methods from the viewpoint of industrial appli-
cation. Economic Quality Control. 16(1), 49 - 63.

Artzner, P., Delbaen F., Eber, J., and Heath, D. (1999). Coherent measures of risk. Mathe-
matical Finance. 9, 203-228.

Beder, T. (1995). VAR: Seductive but dangerous. Financial Analysts Journal. 51(5), 12-24.

Barone-Adesi, G., Bourgoin, F. and Giannopoulos, K. (1998). Don’t look back. Risk. 11(8),
100-103.

Basel Committee on Banking Supervision (1996). Amendment to the capital accord to incor-
porate market risks. Bank for International Settlements, Basel.

Basel Committee on Banking Supervision (2006). International convergence of capital measure-

ment and capital standards: A revised framework. Bank for International Settlements,
Basel.

Basel Committee on Banking Supervision (2009). Revisions to the Basel II market risk frame-
work. Bank for International Settlements, Basel.

Berkowitz, J., and O’Brien, J. (2002). How accurate are Value-at-Risk models at commercial
banks? Journal of Finance. 55, 1093-1111.

Berkowitz, J., Christoffersen, P. and Pelletier, D. (2010). Evaluating VaR models with desk-
level data. Management Science. Forthcoming

Bollerslev, T. (1987). A conditionally heteroskedastic time series model for speculative prices
and rates of return. Review of Economics and Statistics. 69(3), 542-547.

Branger, N. and Schlag, C. (2004). Model risk: A conceptual framework for risk measurement
and hedging. Working Paper, EFMA Basel Meetings. Available from SSRN.

Brooks, C., and Persand, G. (2002). Model choice and Value-at-Risk performance. Financial
Analysts Journal. 58(5), 87-98.

Cairns, A.. (2000). A discussion of parameter and model uncertainty in insurance. Insurance:
Mathematics and Economics. 27, 313-330.

Chan, F. (2009a). Modelling time-varying higher moments with maximum entropy density.
Mathematics and Computers in Simulation. 79(9), 2767-2778.

Chan, F. (2009b). Forecasting Value-at-Risk using Maximum Entropy. 18th World IMACS /
MODSIM Congress, Cairns, Australia.

19
Christoffersen, P. (1998). Evaluating interval forecasts. International Economic Review. 39,
841-862.

Christoffersen, P., Hahn, J. and Inoue A. (2001). Testing and comparing Value-at-Risk mea-
sures. Journal of Empirical Finance. 8, 325-342.

Christoffersen, P. and Pelletier, D. (2004). Backtesting Value-at-Risk: a duration-based ap-

proach. Journal of Financial Econometrics. 2(1), 84-108.

Christoffersen, P. (2009). Value-at-Risk models. In Handbook of Financial Time Series. Edited

by Mikosch, T., Kreiss, J.P., Davis R.A., and T.G. Andersen. Springer.

Crouhy, M., Galai, D. and Mark, R. (1998). Model risk. Journal of Financial Engineering. 7,
267-288.

Cont, R. (2006). Model uncertainty and its impact on the pricing of derivatiave instruments.
Mathematical Finance, 16(3), 519-547.

Derman, E. (1996). Model risk. Risk, 9:5 139-145.

Ebrahimi, N., Maasoumi, E. and Soofi, E. (1999). Ordering univariate distributions by entropy
and variance. Journal of Econometrics. 90, 317-336.

Engle, R. and Manganelli, S. (2004). CAViaR: Conditional autoregressive Value at Risk by

regression quantile. Journal of Business and Economic Statistics. 22, 367-381.

Green, T., and Figlewski, S. (1999). Market risk and model risk for a financial institution
writing options. Journal of Finance. 54, 1465-1499.

Hartz, C., Mittnik, S. and Paolella, M. (2006) Accurate Value-at-Risk forecasting based on the
normal-GARCH model. Computational Statistics and Data Analysis. 51(4), 2295-2312.

Hendricks, D. (1996). Evaluation of Value-at-Risk models using historical data. FRBNY

Economic Policy Review. April. 39-69.

Hull, J. and White, A. (1998a). Incorporating volatility updating into the historical simulation
method for Value-at-Risk. Journal of Risk. 1, 5-19.

Hull, J. and White, A. (1998b). Value at risk when daily changes in market variables are not
normally distributed. Journal of Derivatives. 5(3), 9-19.

Hull, J. and Suo, W. (2002). A methodology for assessing model risk and its application to the
implied volatility function model. Journal of Financial and Quantitative Analysis. 37(2),
297-318.

Jaynes, E. (1983) Papers on probability, statistics and statistical physics. Edited by R.D.
Rosenkrantz, Reidel Publishing Company, Dordrecht.

Jorion, P. (1996). Measuring the risk in value at risk. Financial Analysts Journal 52, 47-56.

Kato, T. and Yoshiba, T. (2000). Model risk and its control. Monetary and Economic Studies.
December. 129-157.

Katz, R., Parlange M. and Naveau P. (2002). Statistics of extremes in hydrology. Advances in
Water Resources. 25, 1287-1304.

20
Kerkhof, J., Melenberg, B. and Schumacher, H. (2010). Model risk and capital reserves. Journal
of Banking and Finance. 34, 267-279.

Kuan, C-M., Yeh, J-H. and Hsu, Y-C. (2009). Assessing Value-at-Risk with CARE: the con-
ditional autoregressive expectile models. Journal of Econometrics. 150(2), 261-270.

Kupiec, P. (1995). Techniques for verifying the accuracy of risk measurement models. Journal
of Derivatives. 3(2), 73-84.

Matthys, G., Delafosse, E., Guillou A. and Beirlant, J. (2004). Estimating catastrophic quantile
levels for heavy-tailed distributions. Insurance: Mathematics and Economics. 34, 517-537.

McDonald, J. (1984). Some generalized functions for the size distribution of income. Econo-
metrica. 52(3), 647-663.

Mittnik, S. and Paolella, M. (2000). Conditional density and Value-at-Risk: prediction of

Asian currency exchange rates. Journal of Forecasting. 19, 313-333.

Mkhandi, S., Kachroo, R. and Guo S. (1996). Uncertainty analysis of flood quantile estimates
with reference to Tanzania. Journal of Hydrology. 185, 317-333.

Mohammad-Djafari, A. (2001). A Matlab program to calculate maximum entropy distribu-

tions. arXiv:physics/0111126v1 [[Link]-an]

Park, S. and Bera, A. (2009). Maximum entropy autoregressive conditional heteroskedasticity

model. Journal of Econometrics. 150(2), 219-230.

Peng, L. and Qi, Y. (2006). Confidence regions for high quantiles of a heavy tailed distribution.
Annals of Statististics. 34(4), 1964-1986.

Perignon, C. and Smith, D. (2010). The level and quality of Value-at-Risk disclosure by
commercial banks. Journal of Banking and Finance. 34(2), 362-377.

Pritsker, M. (2006). The hidden dangers of historical simulation. Journal of Banking and
Finance. 30, 561-582.

RiskMetrics (1997). Technical Document. Available from the RiskMetrics website.

Rebonato, R. (2001). Managing Model Risk. In Volume 2 of Mastering Risk, C. Alexander

(ed.), Pearson UK.

Reiss, R. and Thomas, M. (1997). Statistical Analysis of Extreme Values with Applications to
Insurance, Finance, Hydrology and Other Fields. Birkhauser Verlag, Basel.

Rockinger, M. and Jondeau, E. (2002). Entropy densities with an application to autoregressive

conditional skewness and kurtosis. Journal of Econometrics. 106, 119-142.

Shannon, C. (1948). The mathematical theory of communication, Bell System Technical Jour-
nal July-Oct. Reprinted in: C.E. Shannon and W. Weaver, The mathematical theory of
communication University of Illinois Press, Urbana, IL 3-91.

Simons, K. (1997). Model error - evaluation of various finance models. New England Economic
Review. Nov-Dec, 17-28.

21
Talay, D. and Zheng, Z. (2002). Worst case model risk management. Finance and Stochastics.
6, 517-537.

Taleb, N. (2007). The black swan: The impact of the highly improbable. Penguin.

Turner, L. (2009). The Turner review: A regulatory response to the global banking crisis.
Financial Services Authority, London.

Unnikrishnan Nair, N. and Sankaran, P. (2009). Quantile-based reliability analysis. Commu-

nications in Statistics - Theory and Methods. 38(2), 222-232.

Venter, J. and de Jongh, P. (2002). Risk estimation using the normal inverse Gaussian distri-
bution. Journal of Risk. 4, 1-23.

Wheelock, D. and Wilson, P. (2008). Non-parametric, unconditional quantile estimation for

efficiency analysis with an application to Federal Reserve check processing operations.
Journal of Econometrics. 145, 209-225.

Wu, X. (2003). Calculation of maximum entropy densities with application to income distri-
bution. Journal of Econometrics. 115(2), 347-354

Zellner, A. (1977). Maximal data information prior distributions. In: A. Aykac and C. Brumat,
eds., New methods in the applications of Bayesian methods. North-Holland.

22