Mathematical Finance, Vol. 12, No. 3 (July 2002), 239-269: EY Ords

Mathematical Finance, Vol. 12, No.
3 (July 2002), 239–269
PORTFOLIO VALUE-AT-RISK WITH HEAVY-TAILED RISK

FACTORS
Paul Glasserman
Graduate School of Business, Columbia University
Philip Heidelberger
IBM Research Division, Yorktown Heights, NY
Perwez Shahabuddin
IEOR Department, Columbia University
This paper develops efficient methods for computing portfolio value-at-risk (VAR) when the
underlying risk factors have a heavy-tailed distribution. In modeling heavy tails, we focus on
multivariate t distributions and some extensions thereof. We develop two methods for VAR
calculation that exploit a quadratic approximation to the portfolio loss, such as the delta-gamma
approximation. In the first method, we derive the characteristic function of the quadratic
approximation and then use numerical transform inversion to approximate the portfolio loss
distribution. Because the quadratic approximation may not always yield accurate VAR estimates, we
also develop a low variance Monte Carlo method. This method uses the quadratic approximation to
guide the selection of an effective importance sampling distribution that samples risk factors so that
large losses occur more often. Variance is further reduced by combining the importance sampling with
stratified sampling. Numerical results on a variety of test portfolios indicate that large variance
reductions are typically obtained. Both methods developed in this paper overcome difficulties
associated with VAR calculation with heavy-tailed risk factors. The Monte Carlo method also extends
to the problem of estimating the conditional excess, sometimes known as the conditional VAR.
KEY WORDS: value-at-risk, delta-gamma approximation, Monte Carlo, simulation, variance
reduction, importance sampling, stratified sampling, conditional excess, conditional value-at-risk
1. INTRODUCTION
A central problem in market risk management is estimation of the profit-and-loss
distribution of a portfolio over a specified horizon. Given this distribution, the
calculation of specific risk measures is relatively straightforward. Value-at-risk (VAR),
for example, is a quantile of this distribution. The expected loss and the expected excess
loss beyond some threshold are integrals with respect to this distribution. The difficulty
in estimating these types of risk measures lies primarily in estimating the profit-and-loss
distribution itself, especially the tail of this distribution associated with large losses.
P. Glasserman and P. Shahabuddin were partially supported by NSF NYI Award DMI9457189 and NSF
Career Award DMI9625297, respectively.
Manuscript received August 2000; final revision received January 2001.
Address correspondence to Paul Glasserman, 403 Uris Hall, Columbia University, New York, NY 10027;
pg20@columbia.edu.
Ó 2002 Blackwell Publishing Inc., 350 Main St., Malden, MA 02148, USA, and 108 Cowley Road, Oxford,
OX4 1JF, UK.
239
240 P. GLASSERMAN, P. HEIDELBERGER, AND P. SHAHABUDDIN
All methods for estimating or approximating the distribution of changes in portfolio

value rely (at least implicitly) on two types of modeling considerations: assumptions
about the changes in the underlying risk factors to which a portfolio is exposed, and a
mechanism for translating these changes in risk factors to changes in portfolio value.
Examples of relevant risk factors are equity prices, interest rates, exchange rates, and
commodity prices. For portfolios consisting of positions in equities, currencies,
commodities, or government bonds, mapping changes in the risk factors to changes in
portfolio value is straightforward. But for portfolios containing complex derivative
securities this mapping relies on a pricing model.
The simplest and perhaps most widely used approach to modeling changes in
portfolio value is the variance-covariance method popularized by RiskMetrics (1996).
This approach is based on assuming (i) that changes in risk factors are conditionally
multivariate normal over a horizon of, say, one day, two weeks, or a month, and (ii) that
portfolio value changes linearly with changes in the risk factors. (‘‘Conditionally’’ here
means conditional on information available at the start of the horizon; the uncondi-
tional distribution need not be normal.) Under these assumptions, the portfolio profit-
and-loss distribution is conditionally normal; its standard deviation can be calculated
from the covariance matrix of the underlying risk factors and the sensitivities of the
portfolio instruments to these risk factors. The attraction of this approach lies in its
simplicity. But each of the assumptions on which it relies is open to criticism, and
research in the area has tried to address the shortcomings of these assumptions.
One line of work has focused on relaxing the assumption that portfolio value
changes linearly with changes in risk factors while preserving computational
tractability. This includes, in particular, the ‘‘delta-gamma’’ methods developed in
Britten-Jones and Schaefer (1999), Duffie and Pan (2001), Rouvinez (1997), and Wilson
(1999). These methods refine the relation between risk factors and portfolio value to
include quadratic as well as linear terms. Methods that combine interpolation
approximations to portfolio value with Monte Carlo sampling of risk factors are
considered in Jamshidian and Zhu (1997), Picoult (1999), and Shaw (1999). Low
variance Monte Carlo methods based on exact calculation of changes in portfolio value
are proposed in Cardenas et al. (1999), Glasserman, Heidelberger, and Shahabuddin
(2000), and Owen and Tavella (1999).
Another line of work has focused on developing more realistic models of changes in
risk factors. It has long been observed that market returns exhibit systematic deviation
from normality: across virtually all liquid markets, empirical returns show higher peaks
and heavier tails than would be predicted by a normal distribution, especially over
short horizons. Early studies along these lines include Mandelbrot (1963), Fama
(1965), Praetz (1972), and Blattberg and Gonedes (1974). More recent investigations,
some motivated by value-at-risk, include Bouchaud, Sornette, and Potters (1997),
Danielsson and de Vries (1997), Eberlein and Keller (1995), Eberlein, Keller, and
Prause (1998), Embrechts, McNeil, and Straumann (2001), Hosking, Bonti, and Siegel
(2000), Huisman et al. (1998), Koedijk, Huisman, and Pownall (1998), McNeil and
Frey (1999), Heyde (1999). Using different approaches to the problem and different sets
of data, these studies consistently find high kurtosis and heavy tails. Moreover, most
studies find that the tails in financial data are not so heavy as to produce infinite
variance (as would be implied by a nonnormal stable distribution), though higher order
moments (e.g., fifth and higher) may be infinite.
This paper contributes to both lines of investigation by developing methods for
calculating portfolio loss probabilities when the underlying risk factors are heavy
PORTFOLIO VAR WITH HEAVY-TAILED RISK FACTORS 241
tailed. Most of the literature documenting heavy tails in market data has focused on the
univariate case—time series for a single risk factor or in some cases a fixed portfolio.
There has been less work on modeling the joint distribution of risk factors with heavy
tails (recent work in this direction includes Embrechts et al. 2002 and Hosking et al.
2000). There has been even less work on combining heavy-tailed joint distributions for
risk factors with a nonlinear relation between risk factors and portfolio value, which is
the focus of this paper.
We model changes in risk factors using a multivariate t distribution and some
generalizations of this distribution. A univariate t distribution is characterized by a
parameter m, its degrees of freedom. The tails of the t density decay at a polynomial rate
of m þ 1, so the parameter m determines the heaviness of the tail and the number of finite
moments. Empirical support for modeling univariate returns with a t distribution or
t-like tails can be found in Blattberg and Gonedes (1974), Danielsson and de Vries
(1997), Hosking et al. (2000), Huisman et al. (1998), Hurst and Platen (1997), Koedijk
et al. (1998), and Praetz (1972). There are many possible multivariate distributions with
t marginals. We follow Anderson (1984), Tong (1990), and others in working with a
particular class of multivariate distributions having t marginals for which the joint
distribution is characterized by a symmetric, positive definite matrix R, along with the
degrees of freedom. The matrix R plays a role similar to that of the covariance matrix
for a multivariate normal; this facilitates modeling with the multivariate t and
interpretation of the model.
Because it is characterized by the matrix R, the multivariate t shares some attractive
properties with the multivariate normal while possessing heavy tails. This is important
in combining a realistic model of risk factors with a nonlinear relation between risk
factors and portfolio value, which is our goal. We use the structure of the
multivaratiate t to develop computationally efficient methods for calculating portfolio
loss probabilities capturing heavy tails and without assuming linearity of the portfolio
value with respect to changes in risk factors. While it may be possible to find other
multivariate distributions that are preferable on purely statistical grounds, the
advantage of such a model may be limited if it cannot be integrated with efficient
methods for calculating portfolio risk measures. The multivariate t balances tractability
with the empirical evidence for heavy tails. Moreover, we will see that some of the
methods developed apply to a broader class of distributions of which the multivariate
t is a particularly interesting special case.
We develop two methods for estimating portfolio loss probabilities in the presence of
heavy-tailed risk factors. The first method uses transform inversion based on a
quadratic approximation to portfolio value. It thus extends the delta-gamma
approximation developed in the multivariate normal setting. But the analysis here
differs from the normal case in several important ways. Because the t distribution has a
polynomial tail, it does not have a moment generating function; and whereas
uncorrelated multivariate normal random variables are necessarily independent, the
same is not true with the multivariate t. This means that the characteristic function for
a quadratic in t’s does not factor into a product of one-dimensional characteristic
functions (as it does in the normal case). Indeed, we never explicitly find the
characteristic function of a quadratic in t’s, which may be intractable. Instead, we use
an indirect transform analysis through which we are able to compute the distribution of
interest.
This method is fast, but a quadratic approximation to portfolio value is not always
sufficiently accurate to produce reliable VAR estimates. We therefore also develop an
efficient Monte Carlo procedure. This method builds on the first; it uses the transform
analysis to design highly efficient sampling procedures that are particularly well suited
to estimating the tail of the loss distribution. The method combines importance
sampling and stratified sampling in the spirit of Glasserman, Heidelberger, and
Shahabuddin (1999a, 1999b, 2000). But the methods in these studies assumed a
multivariate normal distribution and, as is often the case in importance sampling, they
applied an exponential change of measure. An exponential change of measure is
inapplicable to a t distribution, again because of the nonexistence of a moment
generating function. (Indeed, the successful application of importance sampling in
heavy-tailed settings is a notoriously difficult problem; see, e.g., Asmussen and
Binswanger 1997 Asmussen, Binswanger, and Højgaard 2002 and Juneja and
Shahabuddin 1999.) We circumvent this problem by an indirect approach to
importance sampling and stratified sampling. Through careful sampling of market
risk factors, we are able to substantially reduce the variance in Monte Carlo estimates
of loss probabilities and thus to reduce the number of samples needed to estimate a loss
probability to a specified precision. Both a theoretical analysis of the method and
numerical examples indicate the potential for enormous gains in computational speed
as a result of this approach. This therefore makes it computationally feasible to
combine the realism of heavy-tailed distributions and the robustness of Monte Carlo
simulation in estimating portfolio loss probabilities.
The rest of this paper is organized as follows. Section 2 describes the multivariate t
distribution and an extension of it that allows different marginals to have different
degrees of freedom. Section 3 develops the transform analysis for the quadratic (delta-
gamma) approximation to portfolio losses. Section 4 builds on the quadratic
approximation to develop an importance sampling procedure for efficient Monte
Carlo simulation. Section 5 provides an analysis of the importance sampling estimator
and discusses adaptations of the importance sampling procedure for estimating the
conditional excess. Section 6 extends the Monte Carlo method through stratification of
the quadratic approximation. Section 7 presents numerical examples.
2. MULTIVARIATE HEAVY TAILS

The univariate t distribution with m degrees of freedom (abbreviated tm ) has density
ðmþ1Þ=2
C 12 ðm þ 1Þ x2
f ðxÞ ¼ pffiffiffi 1 1 þ ; 1 < x < 1;
pC 2 m m
with CðÞ denoting the gamma function. This distribution is symmetric about 0. It has
polynomial tails and the weight of the tails is controlled by the parameter m: if X has the
tm distribution then
P ðX > xÞ constant
xm ; as x ! 1:
In contrast, if Z has a standard normal distribution then
2
ex =2
P ðZ > xÞ constant
; as x ! 1;
x
so the tails are qualitatively different, especially for small values of m. If X tm , then E½X r
is finite for 0 < r < m and infinite for r m. We are mainly interested in values of m roughly
in the range of 3 to 7 since this seems to be the level of heaviness typical of market data.
As m ! 1, the tm distribution converges to the standard normal. Figure 2.1 compares the
t3 distribution with a normal distribution scaled to have the same variance. As the figure
illustrates, the t3 has a higher peak and its tails decay much more slowly.
For m > 2, the tm distribution has variance m=ðm 2Þ. One can scale a tm random
variable X by a constant to change its variance and translate it by a constant to change
its mean. A linear transformation of tm random variable is sometimes said to have a
Pearson Type VII distribution (Johnson, Kotz, and Balakrishnan 1994, p. 21).
Following Anderson (1994), Tong (1990), and others, we refer to
1
C 12 ðm þ mÞ 1 0 1 2ðmþmÞ
ð2:1Þ f m;R ðxÞ ¼ 1þ xR x ; x 2 Rm :
ðmpÞm=2 C 12 m jRj1=2 m
as a multivariate tm density. Here, R is a symmetric, positive definite matrix and jRj is

the determinant of R. If m > 2, then mR=ðm 2Þ is the covariance matrix of f m;R . Tong’s
definition requires that the diagonal entries of R be 1 (making R the distribution’s
correlation matrix if m > 2); in the more general case of (2.1), the marginals are actually
scalar multiples of univariate tm random variables. Anderson’s (1984) definition allows
a general R and also a nonzero mean vector. This makes each marginal a linear
transformation of a tm random variable (and thus a Pearson Type VII random
variable). As is customary in estimating risk measures over short horizons, we will
assume a mean of zero and thus use (2.1).
The density in (2.1) depends on the argument x only through the quadratic form
x0 R1 x; it is therefore elliptically contoured, meaning that the sets on which f is constant
are hyperellipsoids. Within the elliptically contoured family this density belongs to the
class of scale mixtures of normals and thus has a representation as the distribution of the
product of a multivariate normal random vector and a univariate random variable
independent of the normal (cf. Fang, Kotz, and Ng 1987). In the case of (2.1), this
representation can be expressed as follows: if ðX 1 ; . . . ; X m Þ has density f m;R , then
ðn1 ; . . . ; nm Þ
ð2:2Þ ðX 1 ; . . . ; X m Þ ¼ d pffiffiffiffiffiffiffiffi ;
Y =m
where ¼d denotes equality in distribution, n ¼ ðn1 ; . . . ; nm Þ has distribution N ð0; RÞ, Y
has distribution v2m (chi-square with m degrees of freedom), and n and Y are
Comparison of densities Magnified view of right tail

0.4 0.018
0.35 0.016
t 3
0.3 0.014
normal
0.012
0.25
0.01
0.2
normal 0.008
0.15
0.006
0.1
0.004
t 3
0.05 0.002
0 0
–8 –6 –4 –2 0 2 4 6 8 4 4.5 5 5.5 6 6.5 7 7.5 8
FIGURE 2.1. Comparison of t3 and normal distribution. The normal distribution has
been scaled to have variance 3, like the t3 distribution.
independent. This representation is central to our analysis and indeed several of our
results hold if Y is replaced with some other positive random variable having an
exponential tail. See Chapter 3 of Fang et al. for specific examples of multivariate
distributions of the form in (2.2).
From (2.2) we see that modeling changes in risk factors with a multivariate t is
similar to assuming a stochastic level of market volatility: given Y , the variance of X i is
mRii =Y . It is also clear from (2.2) that X i and X j are not independent even if they are
uncorrelated—that is, even if Rij ¼ 0. In (2.2), risk factors with little or no correlation
may occasionally make large moves together (because of a small outcome of Y ), a
phenomenon sometimes observed in extreme market conditions (see, e.g., Longin and
Solnik 1998).
A possible shortcoming of (2.1) and (2.2) is that they require all X i to share a parameter
m and thus have equally heavy tails. To extend the model to allow multiple degrees of
freedom, we use a copula. (For background on copulas see Nelsen 1999; for applications
in risk management see Embrechts et al. 2002 and Li 2000.) Let Gm denote the cumulative
distribution function for the univariate tm density. Let ðX 1 ; . . . ; X m Þ have the density in
(2.1), assuming for the moment that R has all diagonal entries equal to 1. Define
ð2:3Þ ðX~1 ; . . . ; X~m Þ ¼ ðG1 1
m1 ðGm ðX 1 ÞÞ; . . . ; Gmm ðGm ðX m ÞÞÞ:
Each X i has distribution tm so each Gm ðX i Þ is uniformly distributed on the unit interval

and each G1 mi ðGm ðX i Þ has distribution tmi ; thus, this transformation produces a
multivariate distribution whose marginals are t distributions with varying degrees of
freedom. This mechanism (as well as the algorithms and proofs in this paper) easily
extends to the case where each X i and X~i is a scalar multiple of a t random variable, but
for the sake of simplicity we restrict overselves to the unscaled t whenever we use this
copula mechanism.
A limiting special case of this approach takes m ¼ 1 in (2.1) and (2.3). This gives
ðX 1 ; . . . ; X m Þ a normal distribution and thus generates ðX~1 ; . . . ; X~m Þ through a ‘‘normal
copula.’’ In practice, we are most interested in values of mi in a relatively narrow range
(e.g., 3 to 7); graphical inspection of the joint distributions produced suggests that
using (2.3) with m close to the values mi of interest is preferable to using a normal copula.
For example, Figure 2.2 compares contours of bivariate distributions with
ðm1 ; m2 Þ ¼ ð7; 3Þ generated using m ¼ 5 (left) and m ¼ 1 (right). In both cases the
correlation in the copula (the correlation between the underlying X 1 and X 2 ) is 0.4.
We briefly describe how we envisage the use of (2.3) in modeling market risk factors;
see Hosking et al. (2000) for a more thorough discussion and empirical results along
these lines. (Hosking et al. use m ¼ 1 and call the resulting joint distribution
metagaussian; the same approach can be used with a finite m.) Let S denote an m-vector
of risk factors (market prices and rates or volatility factors) and let DS denote the
change in S over an interval of length Dt. Think of Dt as the horizon over which VAR is
to be calculated and thus typically either one day or two weeks. We model each
DS i , i ¼ 1; . . . ; m, as a scalar multiple of a tmi random variable; assuming mi > 2, we can
write
sffiffiffiffiffiffiffiffiffiffiffiffi
mi 2 ~
ð2:4Þ DS i ¼ r
~i Xi ; X~i tmi :
mi
This makes r~2i the variance of DS i . The parameter mi could be estimated using, for
example, the methods in Hosking et al. or Johnson, Kotz, and Balakrishnan (1995,
t Copula Density with ρ = 0.40, Reference d.o.f. = 5 Normal Copula Multivariate t Density with ρ = 0.40
10 10
8 8
6 6
3 degrees of freedom
3 degrees of freedom
4 4
2 2
0 0
2 2
4 4
6 6
8 8
10 10
10 8 6 4 2 0 2 4 6 8 10 10 8 6 4 2 0 2 4 6 8 10
7 degrees of freedom 7 degrees of freedom
FIGURE 2.2. Comparison of contours of bivariate t distributions with ðm1 ; m2 Þ ¼ ð7; 3Þ

generated using a t copula with m ¼ 5 (left) and a normal copula (m ¼ 1, right), both
with a correlation of 0.4.
Sec. 28.6). (Alternatively, one might fit a scaled t distribution to the return DS i =S i .

Since we are ultimately interested in the distribution of DS i conditional on the current
S i , the effect of this is to change the definition of r ~i in (2.4).)
Once we have chosen marginal distributions as in (2.4), we can define the
transformed variables X i ¼ G1 ~
m ðGmi ðXi ÞÞ, i ¼ 1; . . . ; m, for some choice of m. Applying
this transformation to historical data, we can then estimate the correlation matrix of
X ¼ ðX 1 ; . . . ; X m Þ. Letting R denote this correlation matrix and positing that X has the
density in (2.1) completes the specification of the model: applying (2.3) to X and then
(2.4) to the X~i we recover the DS i . We denote the combined transformation by
DS ¼ KðX Þ with
sffiffiffiffiffiffiffiffiffiffiffiffi
mi 2 1
ð2:5Þ DS i ¼ K i ðX i Þ; K i ðxÞ ¼ r
~i Gmi ðGm ðxÞÞ:
mi
This produces a joint distribution for DS that accommodates tails of different heaviness
for different marginals and captures some of the dependence among risk factors
observed in historical data. Note that R is not the correlation matrix of DS because
(2.3) does not in general preserve correlations. As a monotone transformation, KðX Þ
does however preserve rank correlations. For an extensive discussion of dependence
properties and the use of copula models in risk management applications, see
Embrechts et al. (2002).
3. QUADRATIC APPROXIMATION AND TRANSFORM ANALYSIS

This section develops a method for calculating the distribution of the change in
portfolio value over a fixed horizon assuming that the changes in underlying risk
factors over the horizon are described by a multivariate t distribution and that the
change in portfolio value is a quadratic function of the changes in the risk factors. As in
the previous section, let S denote an m vector of risk factors to which a portfolio is
exposed and let DS denote the change in S from the current time 0 to the end of the
horizon Dt. Fix a portfolio and let V ðt; SÞ denote its value at time t and risk factors S.
The delta-gamma approximation to the change in portfolio value is
@V 1
V ðDt; S þ DSÞ V ð0; DSÞ Dt þ d0 DS þ DS 0 CDS;
@t 2
with
@V @2
di ¼ ; Cij ¼ ; i; j ¼ 1; . . . ; m;
@S i @S i @S j
and all derivatives evaluated at the initial point ð0; SÞ.
An important feature of this approximation is that most of the required first- and
second-order sensitivites (the deltas, gammas, and time decay) are regularly calculated
by financial firms as part of their trading operations. It is therefore reasonable to
assume that these sensitivities are available ‘‘for free’’ in calculating VAR and related
portfolio risk measures. Although this is an important practical consideration, from a
mathematical perspective there is no need to restrict attention to this particular
approximation—we will simply assume the availability of some quadratic approxima-
tion. Also, we find it convenient to work with the loss L ¼ V ð0; DSÞ V ðDt; S þ DSÞ,
rather than the increase in value approximated above. Thus, we work with an
approximation of the form
ð3:1Þ L a0 þ a0 DS þ DS 0 ADS a0 þ Q;
with a0 a scalar, a an m vector and A an m
m symmetric matrix. The delta-gamma
approximation has a ¼ d and A ¼ 12 C. Our interest centers on calculating loss
probabilities P ðL > xÞ assuming equality in (3.1), and on the closely related problem of
calculating VAR—that is, of finding a quantile xp for which P ðL > xp Þ ¼ p with, for
example, p ¼ :01.
We model the changes in risk factors DS using a multivariate t distribution f m;R as in
(2.1) for some symmetric, positive definite matrix R and a degrees-of-freedom
parameter m. (We consider the more general model in (2.5) at the end of this section.)
pffiffiffiffiffiffiffiffi
From the ratio representation (2.2) we know that DS has the distribution of n= Y =m
with n Nð0; RÞ. If C is any matrix for which CC 0 ¼ R, then n has thepdistribution
ffiffiffiffiffiffiffiffi of
CZ with Z N ð0; IÞ. Thus, DS has the distribution of CX with X ¼ Z= Y =m (i.e., with
X having density f m;I ). It follows that
Q ¼d ða0 CÞX þ X 0 ðC 0 ACÞX ;
with X having uncorrelated components. We verify in the proof of Theorem 3.1 below
that among all C for which CC 0 ¼ R it is possible to choose one for which C 0 AC is
diagonal. Letting K denote this diagonal matrix, k1 ; . . . ; km its diagonal entries, and
b ¼ a0 C we conclude that
X
m
ð3:2Þ Q ¼d b0 X þ X 0 KX ¼ ðbj X j þ kj X 2j Þ:
j¼1
At this point we encounter major differences between the normal and t settings. In
the normal case (m ¼ 1) the X j are independent because they are uncorrelated. The
characteristic function of the sum in (3.2) thus factors as the product of the
characteristic functions of the summands. Moreover, each ðbj X j þ kj Xj2 Þ is a linear
transformation of a noncentral chi-square random variable and thus has a convenient
moment generating function and characteristic function (see p. 447 of Johnson et al.
1995). An explicit expression for the characteristic function of Q follows; this can be
inverted numerically to compute probabilities P ðQ > xÞ which can in turn be used to

approximate the loss distribution through (3.1). A similar analysis applies if X is a finite
mixture of normals.
These simplifying features do not extend to the multivariate t. Though uncorrelated,
the X j in (3.2) are no longer independent so the characteristic function does not factor
as a product over j. Even if it did, the one-dimensional transforms would be difficult to
work with: because they are heavy tailed, X j and X 2j do not have moment generating
functions; each X 2j has an F distribution, for which the characteristic function is a
doubly infinite series (see Johnson et al. 1995, eqn. 27.2). It therefore seems fair to
describe the characteristic function of Q in (3.2) as intractable.
Through an indirect approach, we are nevertheless
pffiffiffiffiffiffiffiffi able to calculate the distribution
P ðQ xÞ. Recall the representation X ¼ Z= Y =m, define
ð3:3Þ Qx ¼ ðY =mÞðQ xÞ;
and observe that P ðQ xÞ ¼ P ðQx 0Þ F x ð0Þ. To compute P ðQ xÞ we may
therefore find the characteristic function of Qx and invert it to find P ðQx 0Þ. Note
that Qx is not heavy-tailed and so, unlike Q, its moment generating function exists and
consequently its characteristic function is more tractable. The necessary steps, starting
from (3.1), are provided by the followingpresult.
ffiffiffiffiffiffiffiffi We formulate a more general result by
letting Y in the representation DS ¼ Z= Y =m be fairly arbitrary (but positive). Define
the moment generating function

/Y ðhÞ ¼ E ehY
and suppose /Y ðhÞ < 1 for 0 < h <

hY . We specialize to the multivariate t by taking
Y v2m .
Theorem 3.1. Let k1 k2 km be the eigenvalues of RA and let K be the

diagonal matrix with these eigenvalues on the diagonal. There is a matrix C satisfying
CC 0 ¼ R and C 0 AC ¼ K. Let b ¼ a0 C. Then P ðQ xÞ ¼ F x ð0Þ, where the distribution F x
has moment generating function (mgf)
Y
m
1
ð3:4Þ /x ðhÞ ¼ /Y ðaðhÞÞ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
j¼1 1 2hkj
with
hx 1 X m h2 b2j
ð3:5Þ aðhÞ ¼ þ ;
m 2m j¼1 1 2hkj
provided k1 h < 1=2 and aðhÞ <

hY . In the case of the multivariate tm ,
!m=2
2hx X m
h2 b2 =m Ym
1
ð3:6Þ /x ðhÞ ¼ 1 þ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi :
m j¼1
1 2hk j j¼1 1 2hkj
pffiffiffiffiffiffiffi
The characteristic function of Qx is given by E½expðixQx Þ ¼ /x ðixÞ with i ¼ 1.
Proof. The existence of the required matrix C is the same here as in the normal case
(Glasserman et al. 2000); we include the proof because it is constructive and thus useful
in implementation. Let B be any matrix for which BB0 ¼ R (e.g., the Cholesky factor
of R). As a symmetric matrix, B0 AB has real eigenvalues; these eigenvalues are the same
as those of BB0 A ¼ RA, namely k1 ; . . . ; km . Moreover, B0 AB ¼ U KU 0 where U is an
orthogonal matrix whose columns are eigenvectors of B0 AB. It follows that
U 0 B0 ABU ¼ K and ðBU ÞðBU Þ0 ¼ BB0 ¼ R, so setting C ¼ BU produces the required
matrix.
Given C, we can assume Q has the diagonalized form in (3.2) with X having density
f m;I . To calculate the mgf of Qx , we first condition on Y :
ð3:7Þ E½expðhQx ÞjY ¼ E½expðhðY =mÞðQ xÞÞjY
" " #! #
Xm pffiffiffiffiffiffiffiffi
¼ E exp h ðbj Y =m Z j þ Z 2j jÞ xY =m jY
j¼1
Y
m h pffiffiffiffiffiffiffiffi i
¼ exhY =m E exp hðbj Y =m Z j þ kj Zj2 Þ jY ;
j¼1
because the uncorrelated standard normal random variables Z j are in fact independent.
As in equation (29.6) of Johnson et al. (1995) we have, for Z j Nð0; 1Þ,
2
au
E½expðuðZ j þ aÞ2 Þ ¼ ð1 2uÞ1=2 exp ; u < 1=2;
1 2u
this is the mgf of a noncentral v21 random variable. Each factor in (3.7) for which kj 6¼ 0
can be evaluated using this identity by writing
pffiffiffiffiffiffiffiffi!2
pffiffiffiffiffiffiffiffi 2 bj Y =m b2j ðY =mÞ
bj Y =mZ j þ kj Zj ¼ kj Z j þ ;
2kj 4k2j
If kj ¼ 0, use E½expðuZ j Þ ¼ expðu2 =2Þ. The product in (3.7) thus evaluates to

!
m h2 b2 Y =m Y
xhY =m 1X j
m
1
e exp pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 j¼1 1 2hkj j¼1 1 2hkj
which is
Y
m
1
ð3:8Þ eaðhÞY pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi :
j¼1 1 2hkj
Taking the expectation over Y yields (3.4). If Y v2m , then
/Y ðhÞ ¼ ð1 2hÞm=2 ; h < 1=2;

so the expectation of (3.8) becomes
Y
m
1
ð1 2aðhÞÞm=2 pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ;
j¼1 1 2hkj
which is (3.6).
Finally, from Lukacs (1970, p.11) we may conclude that if the moment generating
function is finite in a neighborhood of the origin, then the characteristic function is
equal to mgf evaluated at purely imaginary arguments. h
In the specific case of the delta-gamma approximation, the matrix A in (3.1) is 12 C

and the parameters k1 ; . . . ; km are the eigenvalues of 12 RC. The constant a0 is
Dtð@V =@tÞ. The delta-gamma approximation to a loss probability is P ðL > xÞ

P ðQ > x a0 Þ. We evaluate this approximation using P ðQ > x a0 Þ ¼ 1
P ðQxa0 0Þ ¼ 1 F xa0 ð0Þ. Values of F xa0 can be found through the inversion integral
Z 1 iuy
1 e 1 iut pffiffiffiffiffiffiffi
ð3:9Þ F xa0 ðtÞ F xa0 ðt yÞ ¼ Re /xa0 ðiuÞ e du; i ¼ 1
p 0 iu
which is obtained from the standard inversion formula (see, e.g., (3.11) on p. 511 of
Feller 1971). This integral can be evaluated numerically; see Abate and Whitt (1992) for
a discussion of numerical issues involved in transform inversion. In implementing this
method we choose a large value of y for which F xa0 ðt yÞ can be assumed to be
approximately zero. As the mean and variance of Qxa0 can be easily computed,
Chebychev’s Inequality may be used to find a value of y for which F xa0 ðt yÞ is
appropriately small.
It should be noted that this transform inversion procedure is not quite as
computationally efficient as the corresponding approach based on normally distributed
risk factors. In the normal case one can evaluate the transform of Q explicitly; a single
Fast Fourier Transform inversion then computes N points on the distribution of Q
based on an N-term approximation to the inversion integral in OðN log N Þ time. In our
setting, each value x at which the distribution of Q is to be evaluated relies on a separate
inversion integral so computing M points of the distribution, each based on an N -term
approximation to the corresponding integral, requires OðMN Þ time. Nevertheless, the
total computing time of this approach remains modest and the additional effort makes
possible the inclusion of realistically heavy tails.
The transform analysis provided by Theorem 3.1 can accommodate different degrees
of heaviness in the tails of different risk factors. The result extends to this case through
the copula mechanism in (2.3) and the chain rule for differentiation. Suppose we model
the changes in risk factors DS as KðX 1 ; . . . ; X m Þ using (3.5) with ðX 1 ; . . . ; X m Þ having
density f m;R (with the diagonal elements of R being 1). Then
@V dK i
¼ di ;
@X i dX i
2
@2V dK i dK j @2V dK i d 2K i
¼ Cij ; i 6¼ j; ¼ Cii þ di ;
@X i @X j dX i dX j @X 2i dX i dX 2i
with all derivatives of K evaluated at 0. In this way, the delta-gamma approximation
generalizes to a quadratic approximation
ð3:10Þ a0 X þ X 0 A
L a0 þ ~ ~X
in ðX 1 ; . . . ; X m Þ, the parameters ~ ~ now depending on the derivatives of K as well

a and A
as the usual portfolio deltas and gammas. The derivatives of K do not depend on the
portfolio and are therefore easily computed.
Figure 3.1 shows the accuracy of the delta-gamma approximation for one of the
portfolios considered in Section 7, portfolio (a.3) (0.1 yr ATM) consisting of European
puts and calls. The exact loss distribution was estimated to a high degree of precision
using Monte Carlo simulation. Although the absolute errors (delta-gamma approx-
imation minus exact) are all within 0.01, the relative errors are large (up to 90%) and
this translates into a large error in the VAR. This illustrates the need for high-accuracy
Monte Carlo techniques.
0.06
Delta gamma approximation
Exact (simulation)
0.05
0.04
P (L > x )
0.03
Error in 0.01 VAR

0.02
0.01
0
200 250 300 350 400 450 500 550 600
x
FIGURE 3.1. Comparison of the delta-gamma approximate and exact loss distributions
for (a.3), the 0.1 yr ATM portfolio. The exact loss distribution was estimated by Monte
Carlo simulation to a high degree of precision.
4. IMPORTANCE SAMPLING
The quadratic approximation of the previous section is fast but may not be sufficiently
accurate for all applications. In contrast, Monte Carlo simulation does not require any
approximation to the relation between changes in risk factors and changes in portfolio
value, but it is much more time consuming. The rest of this paper develops methods for
accelerating Monte Carlo by exploiting the delta-gamma approximation; these
methods thus combine some of the best features of the two approaches.
A generic Monte Carlo simulation to estimate a loss probability P ðL > xÞ consists of
the following main steps. Samples of risk-factor changes DS are drawn from a
distribution; for each sample, the portfolio is revalued to compute V ðDt; S þ DSÞ and
the loss L ¼ V ð0; SÞ V ðDt; S þ DSÞ; the fraction of samples resulting in a loss greater
than x is used to estimate P ðL > xÞ. For large portfolios of complex derivative
securities, the bottleneck in this procedure is portfolio revaluation. The objective of
using a variance reduction technique is to reduce the number of samples (and thus the
number of revaluations) needed to achieve a desired precision. We use importance
sampling based on the delta-gamma (or other quadratic) approximation to reduce
variance, particularly at large loss thresholds x.
4.1. Exponential Change of Measure

We begin by reviewing an importance sampling method developed in Glasserman
et al. (2000) in the case of normally distributed risk factors. Suppose that Q has been
diagonalized as in (3.1) but with the X j replaced (for the moment) with independent
standard normals Z j . As in Theorem 3.1, let k1 be the largest of the eigenvalues kj ;
suppose now that k1 > 0 (otherwise, Q is bounded above). In Glasserman et al. (2000)
we introduced an exponential change of measure by setting, for 0 < h < 1=ð2k1 Þ,
ð4:1Þ dP h ¼ ehQwðhÞ dP ;
with P the original measure under which Z N ð0; IÞ and wðhÞ ¼ log E½expðhQÞ .
Moreover, we showed that under P h the components of Z remain independent but with

hbi 1
Zj N ; :
1 2hki 1 2hki
It is thus a simple matter to sample Z under P h and then to sample DS by setting
DS ¼ CZ.
By (4.1), the likelihood ratio is given by ehQþwðhÞ , so the key identity for importance
sampling in the normal setting is
h i
P ðL > xÞ ¼ Eh ehQþwðhÞ IðL > xÞ ;
with IðÞ denoting the indicator of the event in parentheses. We may thus generate
samples under P h and estimate P ðL > xÞ using the expression inside the expectation on
the right. This estimator is unbiased and its second moment is
h i h i
Eh e2hQþ2wðhÞ IðL > xÞ ¼ E ehQþwðhÞ IðL > xÞ :
If L a0 þ Q, then we can expect ehQ to be small when L > x, resulting in a reduced

second moment, especially for large x. In Glasserman et al. (2000) we showed that if
L ¼ a0 þ Q holds exactly, then this estimator (with suitable h) is asymptotically optimal
in the sense that its second moment decreases exponentially (as x ! 1) at twice the
rate of exponential decrease of P ðL > xÞ itself. This is the best possible rate for the
second moment of any unbiased estimator. Asymptotic optimality holds with, for
example, h ¼ hxa0 , where hxa0 solves
d
wðhÞ ¼ x a0 :
dh
This choice makes Eh ½Q ¼ x a0 and thus makes losses L close to x typical rather than
rare. As shown in Glasserman et al. (2000), the method is not very sensitive to the exact
choice of h.
An attempt to apply similar ideas with a multivariate t seems doomed by the failure
of (4.1) to generalize to the heavy-tailed setting. In any model in which the X i (and
hence also Q) are heavy tailed, one cannot define an exponential change of measure
based on Q because E½expðhQÞ is infinite for any h > 0. Most successful applications of
importance sampling are based on an exponential change of measure; and the
extension of light-tailed methods to heavy-tailed models can often give disastrous
results, as demonstrated in Asmussen et al. (2000).
As in the transform analysis of Section 3, we circumvent this difficulty by working
with Qx in (3.3) rather than Q itself. We use the notation of Theorem 3.1. Let
wx ¼ log /x and wY ¼ log /Y . Recall that k1 ¼ maxj kj .
Theorem 4.1. If hk1 < 1=2 and aðhÞ <

hY , then
dP h ¼ exp ðhQx wx ðhÞÞdP
defines a probability measure and

h i h i
ð4:2Þ P ðL > yÞ ¼ Eh ehQx þwx ðhÞ IðL > yÞ ¼ Eh ehðY =mÞðQxÞþwx ðhÞ IðL > yÞ :
Under P h , X has the distribution of Z= Y =m where
h i
ð4:3Þ P h ðY uÞ ¼ E eaðhÞY wY ðaðhÞÞ IðY uÞ ;
and conditional on Y , the components of Z are independent with

2 hbj Y =m 1
ð4:4Þ Z j N ðlj ðhÞ; rj ðhÞÞ; lj ðhÞ ¼ ; r2j ðhÞ ¼ :
1 2hkj 1 2hkj
In the specific case that the distribution of X under P is multivariate tm (i.e., the
P -distribution of Y is v2m ), the distribution of Y under P h is Gammaðm=2; 2=ð1 2aðhÞÞÞ,
the gamma distribution with shape parameter m=2 and scale parameter 2=ð1 2aðhÞÞ.
Proof. The first assertion follows from the fact that wx ðhÞ is finite under the
conditions imposed on h, and (4.2) then follows from the fact that the likelihood ratio
dP =dP h is expðhQx þ wx ðhÞÞ.
Fix constants h and a satisfying hk1 < 1=2 and a < hY . The probability measure P a;0
obtained by exponentially twisting Y by a is defined by the likelihood ratio
dP a;0
¼ eaY wY ðaÞ :
dP
pffiffiffiffiffiffi
Let hðzÞ denote the standard normal density expðz2 =2Þ= 2p; the density of the
Nðl; r2 Þ distribution is hð½z l =rÞ=r. The probability measure P a;h obtained by
exponentially twisting Y by a and, conditional on Y , letting the Z j be independent with
the distributions in (4.4) is defined by the likelihood ratio
dP a;h Ym
hð½Z j lj ðhÞ =rj ðhÞ Þ=rj ðhÞ
ð4:5Þ ¼ eaY wY ðaÞ
:
dP j¼1
hj ðZ j Þ
The jth factor in the product in (4.5) evaluates to

! !
hð½Z j lj ðhÞ =rj ðhÞ Þ=rj ðhÞ 1 1 2 1 lj ðhÞ 1
¼ exp Zj 1 2 þ Zj 2 l ðhÞ2
hj ðZ j Þ rj ðhÞ 2 rj ðhÞ rj ðhÞ 2 j
!
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2
pffiffiffiffiffiffiffiffi h2j b2j Y =m
¼ 1 2hkj exp hkj Zj þ Z j hbj Y =m :
2ð1 2hkj Þ
The likelihood ratio (4.5) is thus

! !
Y m
X pffiffiffiffiffiffiffiffi m h2 b2 Y =m
aY wY ðaÞ
m pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1X j
e
1 2hkj exp h kj Zj2 þ Z j bj Y =m
exp ;
j¼1 j¼1
2 j¼1 1 2hkj
which we can write as

! !
Y m h2 b2 =m
aY þhðY =mÞQ
1X j
e 1 2hkj ewY ðaÞ exp Y :
j¼1
2 j¼1 1 2hkj
If we choose
m h2 b2 =m
hx 1 X j
a ¼ aðhÞ þ ;
m 2 j¼1 1 2hkj
the likelihood ratio simplifies to

!
Y
hxðY =mÞþhðY =mÞQ
e 1 2hkj ewY ðaðhÞÞ ¼ ehQx wx ðhÞ ;
j¼1
in light of the definition of Qx in (3.3), the definition of wx as log /x , and the expression
for /x in (3.4). Since this coincides with the likelihood ratio dP h =dP , we conclude that
the P h -distribution of ðY ; Z 1 ; . . . ; Z m Þ is as claimed.
Consider now the multivariate tm case. To find the density of Y under P h , we multiply
the v2m density by the likelihood ratio to get
m=2 ðm2Þ=2
y ðm2Þ=2 ey=2 2 y y
eaywY ðaÞ m=2 ¼ exp ;
2 Cðm=2Þ 1 2a Cðm=2Þ 2=ð1 2aÞ
using expðwY ðaÞÞ ¼ ð1 2aÞm=2 . This is the gamma density with shape parameter m=2
and scale parameter 2=ð1 2aÞ.
4.2. Importance Sampling Algorithm

Embedded in the proof of Theorem 4.1 are the essential steps of a simulation
procedure that uses the quadratic approximation to guide the sampling of risk factors.
We now make this explicit, detailing the steps involved in estimating a loss probability
P ðL > yÞ. We assume for now that a value of h > 0 consistent with the conditions of
Theorem 4.1 has been selected. Later we address the question of how h should be chosen.
Algorithm 4.1. Importance Sampling Estimate of Loss Probability

For each of n independent replications:
1. Generate Y from the distribution in (4.3). (In the multivariate t model, this
means generating Y from the gamma distribution in the theorem.)
2. Given Y , generate
pffiffiffiffiffiffiffiffiffi independent normals Z 1 ; . . . ; Z m with parameters as in (4.4).
3. Set X ¼ Z= Y = .
4. Set S ¼ CX and calculate the resulting portfolio loss L and the quadratic
approximation Q. Set Qx ¼ ðY = ÞðQ xÞ.
5. Multiply the loss indicator by the likelihood ratio to get
ð4:6Þ ehQx þwx ðhÞ IðL > yÞ
Average (4.6) over the n independent replications.
Applying this algorithm with the copula model (2.5) requires changing only the first
part of step 4: to sample the change in risk factors, set DS ¼ KðC ~ Z= Y =mÞ, where C ~
~ ~ 0 ~ 0~ ~ ~
satisfies C C ¼ R and C AC is diagonal, with A as in (3.10). (Recall that in the setting of
(2.5) R is assumed to have diagonal entries equal to 1 and represents the correlation matrix
of the copula variables ðX 1 ; . . . ; X m Þ and not of DS.) The matrixP C~ can be constructed by
P
following the steps in the proof of Theorem 3.1. Also, Q ¼ j bj X j þ j k2j Xj2 with
ðb1 ; . . . ; bm Þ ¼ ~
aC~ and k1 ; . . . ; km the diagonal elements of C ~C
~ 0A ~.
Notice that in Algorithm 4.1 we have not applied an exponential change of measure
to the heavy-tailed random variables X i . Instead, we have applied an exponential
change of measure to Y and then, conditional on Y , applied an exponential change of
measure to Z.
To develop some intuition for this procedure, consider again the diagonalized
quadratic approximation in (3.2) and the representation X ¼ Z= Y =m of the trans-
formed (and uncorrelated) risk factors X . An objective of any importance sampling
procedure for estimating P ðL > yÞ is to make the event fL > yg more likely under the
importance sampling measure than under the original measure. Achieving this is made
difficult by the fact that the actual loss function may be quite complicated and may be
known only implicitly through the procedures used to value individual components of a
portfolio. As a substitute we use an approximation to L (in particular, the quadratic
approximation Q) and design the change of measure to make large values of Q more likely.
Consider the change of measure in Theorem 4.1 and Algorithm 4.1 from this
perspective. The parameter aðhÞ will typically be negative because h is positive and
typically small (so h2 h). In exponentially twisting Y by a < 0, we are giving greater
probability to smaller values of Y and thus increasing the overall variability of the risk
factors, since Y appears in the denominator of X . Given Y , the mean lj ðhÞ in (4.4) is
positive if bj > 0 and negative if bj < 0. This has the effect of increasing the probability
of positive values of Z j if bj > 0 and negative values of Z j if bj < 0; in both cases, the
combined effect is to increase the probability of large values of bj Z j and thus of Q.
Similarly, rj ðhÞ > 1 if kj > 0 and by increasing the standard deviation of Z j we make
large values of kj Zj2 more likely.
This discussion should help motivate the importance sampling approach of Theorem
4.1 and Algorithm 4.1, but it must be stressed that the validity of the algorithm (as
provided by (4.2)) is not based on assuming that the quadratic approximation holds
exactly. In fact, the procedure above produces unbiased estimates even if the bj and kj bear
no relation to the true portfolio. Of course, the effectiveness of the procedure in reducing
variance will in part be determined by the accuracy of the quadratic approximation.
We close this section by addressing the choice of parameter h. In fact we also have
flexibility in choosing the value x used to define Qx . The choice of x is driven by the
approximation P ðL > yÞ P ðQ > xÞ; in light of (3.1), it is natural to take x ¼ y a0 .
Ideally, we would like to choose h to minimize the second moment
h i h i
ð4:7Þ Eh e2hQx þ2wx ðhÞ IðL > x þ a0 Þ ¼ E ehQx þwx ðhÞ IðL > x þ a0 Þ :
Since this is ordinarily intractable, we apply the quadratic approximation and choose a
value of h that would be attractive with L replaced by a0 þ Q. After this substitution,
noting that Qx > 0 when Q > x, we can bound the second moment using
h i
E ehQx þwx ðhÞ IðQ > xÞ ewx ðhÞ :
The function wx is convex and wx ðhÞ ! 1 as h " 1=ð2k1 Þ (assuming k1 > 0) and as h
decreases to a point at which aðhÞ ¼
hY . Hence, this upper bound is minimized by the
point hx solving
d
ð4:8Þ w ðhÞ ¼ 0:
dh x
The root of this equation is easily found numerically.
The value hx determined by (4.8) has a second interpretation that makes it appealing.
The function wx is the cumulant generating function (the logarithm of the moment
generating function) of the random variable Qx . A standard property of exponential
families implies that at any h for which wx ðhÞ < 1, we have Eh ½Qx ¼ dwx ðhÞ=dh. By
choosing hx as the root of (4.8) we are choosing Ehx ½Qx ¼ 0. This may be viewed as
centering the distribution of Qx near 0, which is equivalent to centering Q near x. Thus,
by sampling under P hx we are making values of Q near x typical, whereas they may have
been rare under the original probability measure.
Equation (4.8) provides a convenient and effective means of choosing h. In our
numerical experiments we find that the performance of the importance sampling
method is not very sensitive to the exact choice of h and a single parameter can safely
be used for estimating P ðL > yÞ at multiple values of y. These observations are
consistent with the theoretical properties established in the next section.
5. ANALYSIS OF THE ESTIMATOR

In this section we provide theoretical support for the effectiveness of the importance
sampling method of the previous section. We do this by analyzing the second moment
of the estimator at large loss thresholds (and thus small loss probabilities). We carry
out this analysis under the hypothesis that the quadratic approximation (3.1) holds
exactly and interpret the results as ensuring the effectiveness of the method whenever
the quadratic approximation is sufficiently informative. The results of this section are
specific to the multivariate tm distribution though similar results should hold under
appropriate conditions on the distribution of Y .
5.1. Bounded Relative Error and Asymptotic Optimality

p of the probability P ðQ > xÞ and let m2 ð^pÞ denote its
Consider any unbiased estimator ^
second moment. The variance of the estimator is m2 ð^pÞ ðP ðQ > xÞÞ2 . Since the variance
must be nonnegative, the second moment can never be smaller than ðP ðQ > xÞÞ2 . As
in Shahabuddin (1994), we say that an estimator has bounded relative error if
m2 ð^pÞ
ð5:1Þ lim sup < 1:
x!1 ðP ðQ > xÞÞ2
In Lemma 5.1 we provide conditions under which P ðQ > xÞ is of the order of xm=2 .
When this holds, we must also have
pÞ constant
xm :
m2 ð^
In this case, bounded relative error becomes equivalent to the requirement that there be
a constant c such that
ð5:2Þ pÞ cxm ;
m2 ð^ for all sufficiently large x:
The bounded relative error property ensures that the number of replications required
to achieve a fixed relative accuracy (confidence interval halfwidth of the estimator
divided by the quantity that is being estimated) remains bounded as x ! 1, unlike
standard simulation where this can be shown to tend to infinity. This property is
stronger than the standard notion of asymptotic optimality used in much of the rare
event simulation literature (see, e.g., the discussion in Glasserman et al. 1999a) where
the number of replications may also tend to infinity but at a much slower rate than in
standard simulation. It is also worth noting that (5.1) and (5.2) apply to the degenerate
(and best possible) estimator ^p P ðQ > xÞ, which corresponds to knowing the quantity
being estimated. The second moment of this estimator is simply ðP ðQ > xÞÞ2 , and from
Lemma 5.1, below, we know that this decays at rate xm . Conditions (5.1) and (5.2) may
thus be interpreted as stating that an estimator with bounded relative error is, up to a
constant factor, as good as knowing the answer, at large values of x.
As indicated by this discussion, the first step in analyzing the relative error of our
estimator is analyzing the tail of Q in (3.1). As explained in the discussion leading to
(3.2) we may assume that the X i are uncorrelated, with density fm;I . We begin by noting
that the quadratic form Q is bounded above by a constant if ki < 0 for all i; that is,
P ðQ > xÞ ¼ 0 for large enough x in this case. To avoid such trivial cases, we assume
k1 > 0 (recall that k1 is the largest of the ki ’s).
Lemma 5.1. If k1 > 0, there are positive constants c1 ; c2 such that for all sufficiently
large x
ð5:3Þ P ðQ > xÞ c1 xm=2

and if kj > 0, j ¼ 1; . . . ; m,
ð5:4Þ P ðQ > xÞ c2 xm=2 :
Proof. Recall from the definition of Qx in (3.3) that P ðQ > xÞ ¼ P ðQx > 0Þ. For any
h > 0 at which wx ðhÞ < 1 we have

P ðQx > 0Þ E ehQx IðQx > 0Þ E ehQx ¼ ewx ðhÞ ¼ /x ðhÞ:
From (3.6) we see that

a1
ð5:5Þ /x ðhÞ ¼ c1 xm=2 ;
ða2 þ a3 xÞm=2
for some a1 , a2 , a3 , and c1 > 0 (the ai depending on h). This proves (5.3).
For the second claim, let d j ¼ bj =ð2kj Þ, j ¼ 1; . . . ; m, so
!
X m
2
ð5:6Þ P ðQ > xÞ ¼ P ðkj Xj þ bj X j Þ > x
j¼1
!
X
m
2
¼P kj ðX j þ d j Þ d 2j =kj >x
j¼1
!
X
m
2
P k1 ðX 1 þ d 1 Þ > x þ ðd 2j =kj Þ
j¼1
0 v"
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
# 1
u
u X m
P @X 1 > d 1 þ t x þ ðd 2j =kj Þ =k1 A
j¼1
pffiffiffi
P ðX 1 > c3 xÞ
for some constant c3 and all sufficiently large x. But because X 1 tm , we have
P ðX 1 > uÞ c4 um for some c4 and all sufficiently large u. Applying this to (5.6) proves
(5.4). h
We now use this result and the ideas surrounding (5.1) and (5.2) to examine our
importance sampling estimator applied to P ðQ > xÞ, namely
ð5:7Þ ehQx þwx ðhÞ IðQ > xÞ
sampled under P h . This coincides with our estimate of P ðL > x þ a0 Þ under the
hypothesis that the quadratic approximation is exact. Let
h i h i
ð5:8Þ m2 ðh; xÞ ¼ Eh e2hQx þ2wx ðhÞ IðQ > xÞ ¼ E ehQx þwx ðhÞ IðQ > xÞ
denote the second moment at parameter h.
Theorem 5.1. If k1 > 0, for any fixed h > 0 at which wx ðhÞ < 0 there is a constant cðhÞ
for which
ð5:9Þ m2 ðh; xÞ cðhÞP ðQ > xÞxm=2
for all sufficiently large x; if kj > 0, j ¼ 1; . . . ; m, the estimator (5.7) of P ðQ > xÞ has
bounded relative error. If hx denotes the solution to (4.8) and k1 > 0, then there is a
constant d for which
ð5:10Þ m2 ðhx ; xÞ P ðQ > xÞxm=2 d
for all sufficiently large x; if kj > 0, j ¼ 1; . . . ; m, the estimator based on hx also has
bounded relative error.
Proof. From (5.8) and the fact that h is positive we get

h i
m2 ðh; xÞ E ewx ðhÞ IðQ > xÞ ¼ /x ðhÞP ðQ > xÞ:
From (5.5) we get (5.9). If all kj are positive then (5.4) holds and
m2 ðh; xÞ /x ðhÞ c1 xm=2 c1

¼ ;
ðP ðQ > xÞÞ2 P ðQ > xÞ c2 xm=2 c2
for some positive constants c1 , c2 and all sufficiently large x. This establishes the
bounded relative error property.
For (5.10), we claim that
ð5:11Þ 0 < lim hx < 1=ð2k1 Þ

x!1
with k1 the largest of the kj . Once we establish that (5.11) holds, it follows from (3.6)
that /x ðhx Þxm=2 is bounded by a constant d as x ! 1. This implies (5.10) by the same
argument used for (5.9). Similarly, bounded relative error using hx again follows once
(5.4) holds.
It remains to verify (5.11). We can write the derivative of wx as
d Xm
kj m daðhÞ=dh
w0x ðhÞ wx ðhÞ ¼
dh j¼1
1 2hkj 2 1 þ aðhÞ
with
2
d 2x 2 Xm hb ð1 þ k hÞ
j j
aðhÞ ¼ :
dh m m j¼1 1 2hkj
From this we see that the limit gðhÞ ¼ limx!1 w0x ðhÞ exists for all 0 < h < 1=ð2k1 Þ and is
given by
m Xm
kj
gðhÞ ¼ þ :
2h j¼1 1 2hkj
The function g is increasing in h with gðhÞ ! 1 as h # 0 and gðhÞ ! 1 as h " 1=ð2k1 Þ.

It follows that there is a unique point b, in ð0; 1=ð2k1 ÞÞ at which gðbÞ ¼ 0. The claim (5.11)
holds if we can show that hx ! b. For this, choose > 0 sufficiently small that b > 0
and b þ < 1=ð2k1 Þ. Then gðb Þ < 0 and gðb þ Þ > 0. For all sufficiently large x we
therefore also have w0x ðb Þ < 0 and w0x ðb þ Þ > 0, which implies b < hx < b þ
for all sufficiently large x. Since is arbitrary, we conclude that hx ! b. h
This result indicates that we can expect the importance sampling procedure to be
effective at large loss thresholds x if Q provides a good approximation to L (more
precisely, to L a0 ). It also indicates that we should have quite a bit of freedom in
choosing the parameter h. In our numerical experiments, we choose h ¼ hx . In fact, the
constant d in the upper bound (5.10) on the second moment when using hx can be made
as small as the best constant cðhÞ in the upper bound (5.9) when using a fixed value of h.
This follows since, in the notation of (5.5), xm=2 /x ðhÞ ! a1 ðhÞ=a3 ðhÞm=2 and hx ! b;
simple algebra shows that b also minimizes the function a1 ðhÞ=a3 ðhÞm=2 .
Since we may be interested in estimating multiple points on the loss distribution from
a single set of simulations, it is worth considering whether importance sampling using
hx remains effective in estimating P ðQ > yÞ with y 6¼ x. Let m2 ðh; x; yÞ be the second
moment in (5.8) but with the indicator IðQ > xÞ replaced by IðQ > yÞ. Arguing as in the
proof of Theorem 5.1, we find that if y > x and y=x ! c, then
y m=2 m2 ðhx ; x; yÞ
lim sup cm=2 d
x!1 P ðQ > yÞ
with d the same constant as in (5.10). In particular, if c ¼ 1 (i.e., y ¼ x þ oðxÞ) then we

get the same upper bound using hx as we would using the ‘‘optimal’’ value hy . This
suggests that we can optimize the parameter for some loss level x and still obtain good
estimates for a moderate range of values y, y > x.
5.2. Estimating the Conditional Excess

A common criticism of VAR as a measure of risk is that it is insensitive to the
magnitude of losses beyond a certain percentile. An alternative type of measure
sometimes proposed (Artzner et al. 1999; Bassi, Embrechts, and Kafetzaki 1998;
Uryasev and Rockafellar 2000) is the conditional excess
ð5:12Þ g ¼ gðyÞ ¼ E½LjL > y :
Unlike VAR, the conditional excess weights large losses by their magnitudes. The
threshold y in the definition (5.12) may be a fixed loss level or else the VAR itself.
We examine the effectiveness of our importance sampling procedure in estimating g.

Using ordinary Monte Carlo, one generates independent replications L1 ; . . . ; Ln , all
having the distribution of L, and estimates gðyÞ using
Pn
Lk IðLk > yÞ
ð5:13Þ gn ¼ Pk¼1
^ n :
k¼1 IðLk > yÞ
Applying the law of large numbers to both numerator and denominator shows that this
estimator is consistent—though, being a ratio estimator, it is biased for finite n. Under
importance sampling, the estimator is
Pn
k¼1 ‘k Lk IðLk > yÞ
ð5:14Þ gh;n ¼ P
^ n ;
k¼1 ‘k ðLk > yÞ
where ‘k denotes the likelihood ratio on the kth replication.

The following result compares these estimators based on their asymptotic (as n ! 1)
variances. Let p ¼ P ðL > yÞ and Rk ¼ Lk IðLk > yÞ gIðLk > yÞ. Define r2 ¼ E½R2k
and r2h ¼ Eh ½ð‘k Rk Þ2 Þ . In the following, ) denotes convergence in distribution.
Proposition 5.1. If 0 < r2 < 1, then as n ! 1,

pffiffiffi
nð^ gn gÞ
) Nð0; 1Þ;
r=p
and if 0 < r2 ðhÞ < 1
pffiffiffi
nð^ gh;n gÞ
) N ð0; 1Þ:
rh =p
If there is a constant such that ‘k whenever Lk > y, then r2h r2 .
The first two statements in this proposition are instances of the usual central limit
theorem for ratio estimators (see, e.g., Serfling 1980) and the last statement follows
from the definition of r2h . An upper bound on the likelihood ratio, as required for this
result, holds if Q > x whenever L > y (e.g., if the quadratic approximation (3.1) is exact
and x ¼ y a0 ). In this case, as in Lemma 5.1, the likelihood ratio is bounded by /x ðhÞ
and, also as in Lemma 5.1, this bound becomes smaller than a constant times xm=2 for
sufficiently large x. Thus, in this case, the in Proposition 5.1 can be made quite small if
x is large. This suggests that our importance sampling method should be similarly
effective for estimating the expected excess as for estimating a loss probability.
6. STRATIFYING THE LIKELIHOOD RATIO

In this section we further exploit the delta-gamma (or other quadratic) approximation
and the structure of the multivariate t distribution to further reduce variance in Monte
Carlo estimates of portfolio loss probabilities. Inspection of the importance sampling
estimator (4.6) suggests that to achieve greater precision we should reduce the sampling
variability associated with the likelihood ratio expðhQx þ wx ðhÞÞ. This general
approach to improving importance sampling estimators proved effective in the
multivariate normal settings treated in Glasserman et al. (1999a, 1999b, 2000).
For the estimator in (4.6), reducing sampling variability in the likelihood ratio is
equivalent to reducing it in Qx as defined in (3.3). We accomplish this through stratified
sampling of Qx : we partition the real line into intervals (these are the strata) and
generate samples of Qx so that the desired number of samples falls in each stratum.
Two issues need to be addressed in developing this method. First, we need to have a
way of defining strata with known probabilities, and this requires being able to
compute the distribution of Qx under P h . Second, we need a way of generating samples
within strata which ensures that the ðY ; Z 1 ; . . . ; Z m Þ generated have the correct
conditional distribution given the stratum in which Qx falls.
To find the distribution of Qx under P h we extend the transform analysis of Section 3.
In particular, we find the characteristic function of Qx under P h through the following
simple observation.
pffiffiffiffiffiffiffi
Lemma 6.1. The characteristic function of Qx under P h is given by /h;x ð 1xÞ, where
/h;x ðsÞ ¼ /x ðh þ sÞ=/x ðhÞ and /x is as in (3.4).
Proof. The moment generating function of Qx under P h is

/x;h ðsÞ ¼ Eh esQx

h i
¼ E ehQx wx ðhÞ esQx
h i
¼ E eðhþsÞQx =ewx ðhÞ ¼ /x ðh þ sÞ=/x ðhÞ:
As in the proof of Theorem 3.1, the characteristic function is the moment generating
function evaluated at a purely imaginary argument. h
Using this result and the inversion integral (3.9) applied to /h;x , we can compute
P h ðQx aÞ for arbitrary a. Given a set of probabilities p1 ; . . . ; pN summing to 1, we can
use the transform inversion iteratively to find points 1 ¼ a0 < a1 < <
aN < aN þ1 ¼ 1 such that P h ðQx 2 ðai1 ; ai ÞÞ ¼ pi , i ¼ 1; . . . ; N. The intervals
ðai1 ; ai Þ form the strata for stratified sampling. We often use equiprobable strata
ðpi 1=N Þ but this is by no means necessary. Alternatively, if the ai ’s are given, then
the pi ’s can be found via transform inversion.
Given N strata and a budget of n samples, we allocate ni samples to stratum i, with
n1 þ þ nN ¼ n. For example, we may choose a proportional allocation with
ni npi ; this choice guarantees a reduction in variance. Let QðijÞ x denote the jth sample
from stratum i, j ¼ 1; . . . ; ni and let LðijÞ denote the corresponding portfolio loss for
that scenario. The combined importance sampling and stratified sampling estimator of
the loss probability P ðL > yÞ is
X
N
p X
ni
ðijÞ
ð6:1Þ i
ehQx þwx ðhÞ
IðLðijÞ > yÞ:
i¼1
ni j¼1
This estimator is unbiased for any set of positive stratum probabilities and
positive allocations. This is true for any h at which wx ðhÞ < 1 (e.g., h ¼ hx Þ. With
the loss threshold y specified we would typically use x ¼ y a0 as suggested by
(3.1).
It remains to specify the mechanism for sampling the QðijÞ x so that ni samples fall in
stratum i. Recall from Algorithm 4.1 that we do not sample Qx directly. Rather, we
generate Y from its exponentially twisted distribution and then generate ðZ 1 ; . p ZmÞ
. . ;ffiffiffiffiffiffiffiffi
according to (4.4). Given ðY ; ZÞ ðY ; Z 1 ; . . . ; Z m Þ, we can then calculate X ¼ Z= Y =m,
DS, L, and Qx .
To implement stratified sampling, we apply a ‘‘bin-tossing’’ method developed in

Glasserman et al. (2000). Keeping count of how many samples have produced values of
Qx in each stratum, we repeatedly generate ðY ; ZÞ as in Algorithm 4.1. For each ðY ; ZÞ
we compute Qx and check which stratum it falls in. If Qx falls in stratum i and we have
previously generated j < ni samples with Qx in stratum i, then the newly generated
ðY ; ZÞ becomes the ðj þ 1Þth sample for the stratum. If we already have ni
samples for stratum i, we discard ðY ; ZÞ and generate a new sample. We repeat this
procedure until the number of samples for each stratum reaches the allocation for the
stratum. This method is somewhat crude, but it is fast and easy to implement; see
Glasserman et al. (2000) for an analysis of its computational requirements.
The combined simulation algorithm using both importance sampling and stratified
sampling follows. We formulate the algorithm to estimate a specific loss probability
P ðL > yÞ, though multiple y’s could be considered simultaneously.
Algorithm 6.1. Importance Sampling and Stratified Sampling Estimate of Loss

Probability
1. Set x ¼ y a0 and find x solving 0x ðx Þ ¼ 0 as in (4.8). Set ¼ x .

2. Numerically invert the characteristic function of Qx under P to find stratum
boundaries a1 ; . . . ; aN for which P ðai1 < Qx < ai Þ ¼ pi , for given p1 ; . . . ; pN .
3. Use the bin-tossing method to generate ðY ðijÞ ; Z ðijÞ Þ, j ¼ 1; . . . ; ni , i ¼ 1; . . . ; N,
so that each QðijÞ
x q
ðijÞ ðijÞ
ffiffiffiffiffiffiffiffiffiffiffiffiffiffi from ðY ; Z Þ falls in stratum i.
calculated
4. Set X ðijÞ ¼ Z ðijÞ = Y ðijÞ = and S ¼ CXðijÞ with C as in Algorithm 4.1.
ðijÞ
Compute the portfolio loss LðijÞ ¼ V ð0; SÞ V ðt; S þ S ðijÞ Þ.

5. Return the estimator in (6.1).
This is also applicable with the copula specification in (2.5). As in Algorithm 4.1,
only the sampling of the values of DS changes. The required modification of step 4 of
Algorithm 6.1 is exactly as described immediately following Algorithm 4.1.
7. NUMERICAL EXAMPLES
We perform experiments with the transform inversion routine of Section 3, the
importance sampling procedure of Section 4, and the combination with stratified
sampling in Section 6. We use a subset of the portfolios that were considered in
Glasserman et al. (2000), but with the light-tailed Gaussian assumptions of that paper
replaced by the heavy-tailed assumptions of this paper. The portfolios in Glasserman et
al. (2000) were chosen so as to incorporate a wide variety of characteristics, such as
portfolios that have all eigenvalues ki positive, portfolios that have some negative ki ’s,
portfolios that have all ki ’s negative, portfolios with discontinuous payoffs (e.g., cash-
or-nothing puts and barrier options), and portfolios with block diagonal correlation
matrices. In the subset of those portfolios that we consider in this paper, we have tried
to give sufficient representation to most of these characteristics. We have, in particular,
included both the best and worst performing cases of Glasserman et al. (2000), where
we experimented with diagonal and nondiagonal correlation matrices and found that
this had little effect on performance. To limit the number of cases, here we mainly
consider uncorrelated risk factors. Also, we focus on estimating loss probabilities and
the conditional excess; issues specific to estimating a quantile rather than a loss
probability were addressed in Glasserman et al. (2000).
In our numerical experiments we value the options in a portfolio using the Black– pffiffiffiffiffi
Scholes formula and its extensions. For the implied volatility of S i we use r ~i =S i Dt
with r~i as in (2.5); in other words, we make the implied volatility consistent with the
standard deviation of DS i over the VAR horizon Dt. There is still an evident
inconsistency in applying Black–Scholes formulas when price changes follow a t
distribution, but option pricing formulas are commonly used this way in practice.
Moreover, it seems reasonable to expect that this simple approach to portfolio
revaluation gives a good indication of the variance reduction that would be obtained
from our Monte Carlo method even if more complex revaluation procedures were used.
The greatest computational gains from reducing the number of Monte Carlo samples
required would in fact be obtained in cases where revaluation is most time consuming,
such as when revaulation requires finite-difference methods, lattices and trees, and
possibly even simulation.
As in Glasserman et al. (2000), we assume 250 trading days in a year and a
continuously compounded risk-free rate of interest of 5%. For each case we investigate
losses over 10 days (Dt ¼ 0:04 years). Most of the test porfolios we consider are based
on 10 uncorrelated underlyingpassets ffiffiffiffiffi having an initial value of 100 and an annual
~i ¼ 0:3S i Dt). In three cases we also consider correlated assets
volatility of 0.3 (i.e., r
and in one of these the portfolio involves 100 assets with different volatilities. Detailed
descriptions are given in Table 7.1. For comparison purposes, in each case we adjust
the loss threshold x so that the probability to be estimated is close to 0.01.
In the first set of experiments, we assume all the marginals to be t distributions
with degree of freedom (d.o.f) 5. Results are given in Table 7.2, which lists the
quadratic approximation and the estimated variance ratios using importance
sampling (IS) and IS combined with stratified sampling (ISS-Q)—that is, the
estimated variance using standard Monte Carlo divided by the estimated variance
using IS (or ISS-Q). This variance ratio indicates how many times as many samples
would be required using ordinary Monte Carlo to achieve the same precision
achieved with the corresponding variance reduction technique; it is thus an estimate
of the computational speedup that can be obtained using a method. In all
experiments, unless otherwise mentioned, the variance ratios are estimated from a
total of 40,000 replications; the stratified estimator uses 40 (approximately)
equiprobable strata with 1,000 samples per stratum. In practice, fewer replications
are usually used; the high number we use is to get accurate estimates of the variances
and thus the computational speedups.
We achieve at least double-digit variance reduction in all cases. It is also encouraging
that the variance ratios obtained for the 100 asset example (a.9) are comparable to the
best variance ratios obtained for the other much smaller 10 asset examples. The
effectiveness of the method is further illustrated in Figure 7.1, which compares standard
simulation to importance sampling with stratification for the 0.1 yr ATM portfolio.
The figure plots point estimates and 99% confidence for P ðL > xÞ over a range of
x values; a total of 4,000 replications were used for each method to simultaneously
estimate P ðL > xÞ for the set of x values indicated in the figure. The importance
sampling uses a single value of the parameter h, chosen to be hx for an x in the middle of
the range. Notice how much narrower the confidence intervals are for the ISS-Q
method over the entire range of x’s.
In the next set of experiments, for (a.1) to (a.11), we assume that the first five
marginals have d.o.f 3 and the next five have d.o.f. 7. The ‘‘reference d.o.f.’’ was taken
to be 5 (i.e., we use the copula method described earlier with m ¼ 5 in (2.5)).
TABLE 7.1
Test Portfolios for Numerical Results
Portfolio Description
(a.1) 0.5 yr ATM Short 10 at-the-money calls and 5 at-the-money puts on
each asset, all options having a half-year maturity. All
eigenvalues are positive.
(a.2) 0.5 yr ATM, k Long 10 at-the-money calls and 5 at-the-money puts on
each asset, all options having half a year maturity. All
eigenvalues are negative.
(a.3) 0.1 yr ATM Same as (a.1) but with a maturity of 0.10 years.
(a.4) 0.1 yr ATM, k Same as (a.2) but with a maturity of 0.10 years.
(a.5) Delta hedged Same as (a.3) but with number of puts increased so that
d ¼ 0.
(a.6) Delta hedged, k Short 10 at-the-money calls on first five assets. Long 5
at-the-money calls on the remaining assets. Long or short
puts on each asset so that d ¼ 0.
This has both negative and positive eigenvalues.
(a.7) DAO-C Short 10 down-and-out calls on each asset with barrier at
95.
(a.8) DAO-C & CON-P Short 10 down-and-out calls with barrier at 95, and short
5 cash-or-nothing puts on each asset. The cash payoff is
equal to the strike price.
(a.9) DAO-C & CON-P, Same as (a.8) but the number of puts is adjusted so that
Hedged d ¼ 0.
(a.10) Index Short 50 at-the-money calls and 50 at-the-money puts
on 10 underlying assets, all options expiring in 0.5 years.
The covariance matrix was downloaded from the
RiskMetricsTM web site for international equity indices.
The initial asset prices are (100, 50, 30, 100, 80, 20, 50,
200, 150, 10).
(a.11) Index, km < k1 Same as (a.10) but short 50 at-the-money calls and 50
at-the-money puts on the first three assets, long 50 at-the-
money calls and 50 at-the-money puts on the next seven
assets. This has both negative and positive eigenvalues
with the absolute value of the minimum eigenvalue
greater than that of the maximum.
(a.12) 100, Short 10 at-the-money calls and 10 at-the-money puts on
Block-diagonal 100 underlying assets, all expiring in 0.10 years. Assets are
divided into 10 groups of 10. The correlation is 0.2
between assets in the same group and 0 across groups.
Assets in the first three groups have volatility 0.5, those in
the next four have volatility 0.3, and those in the last three
groups have volatility 0.1.
TABLE 7.2
Comparison of Methods for Estimating P ðL > xÞ for Test Portfolios. All the Marginals
are t with 5 Degrees of Freedom
Variance ratios
Portfolio x P fL > xg P fQ þ c > xg IS ISS-Q

(a.1) 0.5 yr ATM 311 1.02% 1.17% 53 333
(a.2) 0.5 yr ATM, k 145 1.02% 1.33% 35 209
(a.3) 0.1 yr ATM 469 0.97% 1.56% 46 134
(a.4) 0.1 yr ATM, k 149 0.97% 0.86% 21 28
(a.5) Delta hedged 617 1.07% 1.69% 42 112
(a.6) Delta hedged, k 262 1.02% 1.70% 27 60
(a.7) DAO-C 482 0.91% 0.52% 58 105
(a.8) DAO-C & CON-P 835 0.97% 1.19% 18 20
(a.9) DAO-C & CON-P, 345 1.09% 0.36% 17 25
Hedged
(a.10) Index 2019 1.04% 1.22% 26 93
(a.11) Index, km < k1 426 1.02% 1.16% 18 48
(a.12) 100, Block-diagonal 5287 0.95% 1.58% 61 287
For (a.12), we assume that the marginals in the first three groups have d.o.f. 3, the
marginals in the second four groups have d.o.f. 5, and the marginals in the last three
groups have d.o.f. 7; the reference d.o.f. was again taken to be 5. Results for all these
cases are given in Table 7.3. Note that the performance of IS remains roughly the same
(except for (a.9)), but the performance of ISS-Q decreases substantially. This is to be
expected as the transformation from the t distribution with the reference d.o.f. to
the t distribution of the marginal introduces further nonlinearity in the relation bet-
ween the underlying variables and the portfolio value. Case (a.9) was also the
worst performing case in Glasserman et al. (2000; case (b.6) in that paper); in this
particular case, the delta-gamma approximation gives a poor approximation to the
true loss.
Finally, we estimate the conditional excess for all the portfolios described above.
Table 7.4 gives results using IS and ISS-Q. We again compare each case with standard
simulation, where by standard simulation we mean the estimator given by (5.13). In
particular, for IS we estimate the ratio r2 =r2h where r2 and r2h have been defined in
Proposition 5.1; expressions that may be used to estimate these quantities are given
in Serfling (1980). One can similarly estimate the variance ratios for the ISS-Q.
8. CONCLUDING REMARKS
This paper develops efficient computational procedures for approximating or estimat-
ing portfolio loss probabilities in a model that captures heavy tails in the joint
distribution of market risk factors. The first method is based on transform inversion of
a quadratic approximation to portfolio value. The second method uses the first to
develop Monte Carlo sampling procedures that can greatly reduce variance compared
with ordinary Monte Carlo.
0.06
Standard simulation
Confidence interval
0.05 IS and stratification
Confidence interval
0.04
P (L > x )
0.03
0.02
0.01
0
200 300 400 500 600
x
Figure 7.1. Point estimates and 99% confidence intervals for the 0.1 yr ATM portfolio
using standard simulation and importance sampling with stratification. The estimates
are from a total of 4,000 replications and 40 strata.
TABLE 7.3
Comparison of Methods for Estimating P ðL > xÞ for Different Portfolios Where All
the Marginals Are t’s with Different Degrees of Freedom
Variance ratios
Portfolio x P fL > xg P fQ þ c > xg IS ISS-Q

(a.1) 0.5 yr ATM 322 1.05% 0.82% 37 48
(a.2) 0.5 yr ATM, k 143 0.99% 1.25% 35 136
(a.3) 0.1 yr ATM 475 1.01% 1.16% 38 55
(a.4) 0.1 yr ATM, k 149 1.08% 0.91% 17 20
(a.5) Delta hedged 671 0.98% 1.11% 39 57
(a.6) Delta hedged, k 346 0.95% 0.18% 27 34
(a.7) DAO-C 447 1.16% 0.50% 28 32
(a.8) DAO-C & CON-P 777 1.23% 1.15% 16 18

(a.9) DAO-C & CON-P, 333 1.26% 0.29% 1:1 3:1
Hedged
(a.10) Index 1979 1.12% 1.02% 23 39

(a.11) Index, km < k1 442 0.99% 0.27% 3:7 4:0
(a.12) 100, Block-diagonal 5690 0.96% 0.86% 79 199
** indicates that we used 400,000 replications (instead of 40,000) for these cases but the
variance estimates and thus the variance ratio estimates still did not stabilize.
TABLE 7.4
Comparison of Methods for Estimating E½LjL > x for Different Portfolios Where All
the Marginals Are t’s with Varying Degrees of Freedom
Variance ratios
Portfolio x EðLjL > xÞ IS ISS-Q

(a.1) 0.5yr ATM 322 564 33 35
(a.2) 0.5yr ATM, k 143 160 39 59
(a.3) 0.1yr ATM 475 785 50 53
(a.4) 0.1yr ATM, k 149 156 20 21
(a.5) Delta hedged 671 638 40 63
(a.6) Delta hedged, k 346 1106 44 55
(a.7) DAO-C 447 680 46 50
(a.8) DAO-C & CON-P 777 969 17 20

(a.9) DAO-C & CON-P, Hedged 333 548 8:4 10
(a.10) Index 1979 3583 23 24

(a.11) Index, km < k1 442 841 2:2 4:6
(a.12) 100, Block-diagonal 5690 9300 93 236
** indicates that we used 400,000 replications (instead of 40,000) for these cases but the
variance estimates and thus the variance ratio estimates still did not stabilize.
Our results are based on modeling the joint distribution of risk factors using a
multivariate t and some extensions of it. This may be viewed as a reduced-form
approach to modeling changes in risk factors, in the sense that we have not
specified a continuous-time process for the evolution of the risk factors. Though it
is possible to construct a process with a multivariate t distribution at a fixed time
Dt, we know of no process having this distribution at all times. So the model used
here requires fixing a time interval Dt. (The same is true of most time-series
models, including GARCH, for example.) This is in contrast to Lévy process
models considered in Barndorff-Nielsen (1998), Eberlein et al. (1998), and Madan
and Seneta (1990); but the distributions in those models have exponential tails and
are thus not as heavy as the distributions considered here. Some of the
distributions that arise in these models admit representations through the normal
distribution similar to (2.2) so the methods developed here may be applicable to
them as well.
REFERENCES
Abate, J., and W. Whitt (1992): The Fourier-series Method for Inverting Transforms of
Probability Distributions, Queueing Sys. 10, 5–88.
Anderson, T. W. (1984): An Introduction to Multivariate Statistical Analysis, 2nd ed. New
York: Wiley.
Artzner, P., F. Delbaen, J.-M. Eber, and D. Heath (1999): Coherent Measures of Risk, Math.
Finance 9, 203–228.
Asmussen, S., and K. Binswanger (1997): Simulation of Ruin Probabilities for Subexponential
Claims, ASTIN Bull. 27, 297–318.
Asmussen, S., K. Binswanger, and B. Højgaard (2000): Rare Events Simulation for Heavy Tailed
Distributions, Bernovlli 6, 303–322.
Barndorff-Nielsen, O. E. (1998): Processes of Normal Inverse Gaussian Type, Finance Stoch. 2,
41–68.
Bassi, F., P. Embrechts, and M. Kafetzaki (1998): Risk Management and Quantile Estimation;
in A Practical Guide to Heavy Tails. eds. R. J. Adler, R. Feldman and M. Taqqu. Boston:
Birkhäuser, 111–130.
Blattberg, R. C., and N. J. Gonedes (1974): A Comparison of the Stable and Student Distri-
butions as Statistical Models for Stock Prices, J. Business 47, 244–280.
Bouchaud, J. P., D. Sornette, and M. Potters (1997): Option Pricing in the Presence of Extreme
Fluctuations, in Mathematics of Derivative Securities, eds. M. A. H. Dempster and S. Pliska.
New York: Cambridge University Press.
Britten-Jones, M., and S. M. Schaefer (1999): Non-linear Value-at-Risk, European Finance Rev.
2, 161–187.
Cardenas, J., E. Fruchard, J.-F., Picron, C. Reyes, K. Walters, and W. Yang (1999): Monte
Carlo Within a Day, Risk 12(2), 55–59.
Danielsson, J., and C. G. de Vries (1997): Tail Index and Quantile Estimation with Very High
Frequency Data, Working paper, Tinbergen Insitute, Rotterdam, The Netherlands.
Duffie, D., and J. Pan (2001): Analytical Value-at-Risk with Jumps and Credit Risk, Finance
and Stochastics 5(2), 155–180.
Eberlein, E., and U. Keller (1995): Hyperbolic Distributions in Finance, Bernoulli 1,
281–299.
Eberlein, E., U. Keller, and K. Prause (1998): New Insights Into Smile, Mispricing, and Value-
at-Risk: The Hyperbolic Model, J. Business 71, 371–406.
Embrechts, P., A. McNeil, and D. Straumann (2002): Correlation and Dependence Properties in
Risk Management: Properties and Pitfalls, Extremes and Integrated Risk Management
London: Risk Book.
Fama, E. F. (1965): The Behavior of Stock Market Prices, J. Business 38, 34–105.
Fang, K.-T., S. Kotz, and K. W. Ng (1987): Symmetric Multivariate and Related Distributions.
London: Chapman & Hall.
Feller, W. (1971): An Introduction to Probability Theory and its Applications, Vol. II, 2nd ed.
New York: Wiley.
Glasserman, P., P. Heidelberger, and P. Shahabuddin (1999a): Asymptotically Optimal
Importance Sampling and Stratification for Pricing Path-Dependent Options, Math.
Finance 9, 117–152.
Glasserman, P., P. Heidelberger, and P. Shahabuddin (1999b): Importance Sampling in the
Heath-Jarrow-Morton Framework, J. Derivatives 7(1), 32–50.
Glasserman, P., P. Heidelberger, and P. Shahabuddin (2000): Variance Reduction Techniques
for Estimating Value-at-Risk, Mgmt. Sci. 46, 1349–1364.
Heyde, C. C. (1999): A Risky Asset Model with Strong Dependence Through Fractal Activity
Time, J. Appl. Probab. 36, 1234–1239.
Hosking, J. R. M., G. Bonti, and D. Siegel (2000): Beyond the Lognormal, Risk 13(5),
59–62.
Huisman, R., K. Koedijk, C. Kool, and F. Palm (1998): The Fat-Tailedness of FX Returns.
Working paper, Limburg Institute of Financial Economics, Maastricht University, The
Netherlands. Available at www.ssrn.com.
Hurst, S. R., and E. Platen (1997): The Marginal Distributions of Returns and Volatility, in
L1 -Statistical Procedures and Related Topics, IMS Lecture Notes, Monograph Series Vol.
31. Hayward, CA: IMS, 301–314.
Jamshidian, F., and Y. Zhu (1997): Scenario Simulation: Theory and Methodology, Finance
Stoch. 1, 43–67.
Johnson, N., S. Kotz, and N. Balakrishnan (1994): Continuous Univariate Distributions, Vol. 1.
New York: Wiley.
Johnson, N., S. Kotz, and N. Balakrishnan (1995): Continuous Univariate Distributions, Vol. 2.
New York: Wiley.
Juneja, S., and P. Shahabuddin (1999): Simulating Heavy Tailed Processes Using Delayed
Hazard Rate Twisting. To appear in ACM TOMACS.
Koedijk, K., R. Huisman, and R. Pownall (1998): VaR-x: Fat Tails in Financial Risk Man-
agement, J. Risk 1, 47–62.
Li, D. (2000): On Default Correlation: A Copula Function Approach, J. Fixed Income 9, 43–54.
Longin, F., and B. Solnik (2001): Correlation Structure of International Equity Markets
During Extremely Volatile Periods, J. Finance 56, 649–676.
Lukacs, E. (1970): Characteristic Functions, 2nd ed. New York: Hafner Publishing.
Madan, D., and E. Seneta (1990): The Variance Gamma Model for Share Market Returns,
J. Business 63, 511–524.
Mandelbrot, B. (1963): The Variation of Certain Speculative Prices, J. Business 36, 394–419.
Matacz, A. (2000): Financial Modeling and Option Theory with the Truncated Levy Pro-
cesses, Inter. J. Theo. Appl. Fin. 3, 143–160.
McNeil, A. J., and R. Frey (1999): Estimation of Tail-Related Risk Measures for Hetero-
scedastic Financial Time Series: An Extreme Value Approach. Working paper, RiskLab,
ETH Zurich.
Nelsen, R. B. (1999): An Introduction to Copulas. New York: Springer.
Owen, A., and D. Tavella (1999): Scrambled Nets for Value-at-Risk Calculations; in Monte
Carlo: Methodologies and Applications for Pricing and Risk Management, ed. B. Dupire.
London: Risk Publications, 255–263.
Picoult, E. (1999): Calculating Value-at-Risk with Monte Carlo Simulation; in Monte Carlo:
Methodologies and Applications for Pricing and Risk Management, ed., B. Dupire. London:
Risk Publications, 209–229.
Praetz, P. D. (1972): The Distribution of Share Price Changes, J. Business 45, 49–55.
RiskMetrics Group (1996): RiskMetrics Technical Document, www.riskmetrics.com/research/
techdocs.
Rouvinez, C. (1997): Going Greek with VAR, Risk 10(2), 57–65.
Rydberg, T. H. (1999): Generalized Hyperbolic Diffusion Processes with Applications in
Finance, Math. Finance 9, 183–201.
Serfling, R. J. (1980): Approximation Theorems of Mathematical Statistics. New York: Wiley.
Shahabuddin, P. (1994): Importance Sampling for the Simulation of Highly Reliable Marko-
vian Systems, Mgmt. Sci. 40, 333–352.
Shaw, J. (1999): Beyond VAR and Stress Testing; in Monte Carlo: Methodologies and Appli-
cations for Pricing and Risk Management, ed., B. Dupire. London: Risk Publications,
231–244.
Tong, Y. L. (1990): The Multivariate Normal Distribution. New York: Springer.

Uryase, S., and R. T. Rockafellar (2000): Optimization of Conditional Value-At-Risk, J. Risk
2, 21–41.
Wilson, T. (1999): Value at Risk; in Risk Management and Analysis, Vol 1, ed., C. Alexander.
Chichester, England: Wiley, 61–124.

Mathematical Finance, Vol. 12, No. 3 (July 2002), 239-269: EY Ords

Uploaded by

Copyright:

Available Formats

Mathematical Finance, Vol. 12, No. 3 (July 2002), 239-269: EY Ords

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Mathematical Finance, Vol. 12, No. 3 (July 2002), 239-269: EY Ords

Uploaded by

Copyright:

Available Formats

Mathematical Finance, Vol. 12, No.

3 (July 2002), 239–269

PORTFOLIO VALUE-AT-RISK WITH HEAVY-TAILED RISK

All methods for estimating or approximating the distribution of changes in portfolio

2. MULTIVARIATE HEAVY TAILS

as a multivariate tm density. Here, R is a symmetric, positive deﬁnite matrix and jRj is

Comparison of densities Magnified view of right tail

Each X i has distribution tm so each Gm ðX i Þ is uniformly distributed on the unit interval

FIGURE 2.2. Comparison of contours of bivariate t distributions with ðm1 ; m2 Þ ¼ ð7; 3Þ

Sec. 28.6). (Alternatively, one might ﬁt a scaled t distribution to the return DS i =S i .

3. QUADRATIC APPROXIMATION AND TRANSFORM ANALYSIS

inverted numerically to compute probabilities P ðQ > xÞ which can in turn be used to

and suppose /Y ðhÞ < 1 for 0 < h <

Theorem 3.1. Let k1  k2      km be the eigenvalues of RA and let K be the

provided k1 h < 1=2 and aðhÞ <

If kj ¼ 0, use E½expðuZ j Þ ¼ expðu2 =2Þ. The product in (3.7) thus evaluates to

Taking the expectation over Y yields (3.4). If Y v2m , then

/Y ðhÞ ¼ ð1  2hÞm=2 ; h < 1=2;

In the speciﬁc case of the delta-gamma approximation, the matrix A in (3.1) is  12 C

Dtð@V =@tÞ. The delta-gamma approximation to a loss probability is P ðL > xÞ 

in ðX 1 ; . . . ; X m Þ, the parameters ~ ~ now depending on the derivatives of K as well

Error in 0.01 VAR

4.1. Exponential Change of Measure

If L  a0 þ Q, then we can expect ehQ to be small when L > x, resulting in a reduced

Theorem 4.1. If hk1 < 1=2 and aðhÞ <

deﬁnes a probability measure and

and conditional on Y , the components of Z are independent with

The jth factor in the product in (4.5) evaluates to

The likelihood ratio (4.5) is thus

which we can write as

the likelihood ratio simpliﬁes to

4.2. Importance Sampling Algorithm

Algorithm 4.1. Importance Sampling Estimate of Loss Probability

5. ANALYSIS OF THE ESTIMATOR

5.1. Bounded Relative Error and Asymptotic Optimality

ð5:3Þ P ðQ > xÞ  c1 xm=2

ð5:4Þ P ðQ > xÞ  c2 xm=2 :

P ðQx > 0Þ  E ehQx IðQx > 0Þ  E ehQx ¼ ewx ðhÞ ¼ /x ðhÞ:

From (3.6) we see that

denote the second moment at parameter h.

Proof. From (5.8) and the fact that h is positive we get

m2 ðh; xÞ /x ðhÞ c1 xm=2 c1

ð5:11Þ 0 < lim hx < 1=ð2k1 Þ

The function g is increasing in h with gðhÞ ! 1 as h # 0 and gðhÞ ! 1 as h " 1=ð2k1 Þ.

with d the same constant as in (5.10). In particular, if c ¼ 1 (i.e., y ¼ x þ oðxÞ) then we

5.2. Estimating the Conditional Excess

We examine the eﬀectiveness of our importance sampling procedure in estimating g.

where ‘k denotes the likelihood ratio on the kth replication.

Proposition 5.1. If 0 < r2 < 1, then as n ! 1,

6. STRATIFYING THE LIKELIHOOD RATIO

Proof. The moment generating function of Qx under P h is

/x;h ðsÞ ¼ Eh esQx

To implement stratiﬁed sampling, we apply a ‘‘bin-tossing’’ method developed in

Algorithm 6.1. Importance Sampling and Stratiﬁed Sampling Estimate of Loss

1. Set x ¼ y  a0 and ﬁnd x solving 0x ðx Þ ¼ 0 as in (4.8). Set  ¼ x .

Compute the portfolio loss LðijÞ ¼ V ð0; SÞ  V ðt; S þ S ðijÞ Þ.

Portfolio x P fL > xg P fQ þ c > xg IS ISS-Q

Portfolio x P fL > xg P fQ þ c > xg IS ISS-Q

Theorem 3.1. Let k1 k2 km be the eigenvalues of RA and let K be the

/Y ðhÞ ¼ ð1 2hÞm=2 ; h < 1=2;

In the speciﬁc case of the delta-gamma approximation, the matrix A in (3.1) is 12 C

Dtð@V =@tÞ. The delta-gamma approximation to a loss probability is P ðL > xÞ

If L a0 þ Q, then we can expect ehQ to be small when L > x, resulting in a reduced

ð5:3Þ P ðQ > xÞ c1 xm=2

ð5:4Þ P ðQ > xÞ c2 xm=2 :

P ðQx > 0Þ E ehQx IðQx > 0Þ E ehQx ¼ ewx ðhÞ ¼ /x ðhÞ:

m2 ðh; xÞ /x ðhÞ c1 xm=2 c1

The function g is increasing in h with gðhÞ ! 1 as h # 0 and gðhÞ ! 1 as h " 1=ð2k1 Þ.

1. Set x ¼ y a0 and ﬁnd x solving 0x ðx Þ ¼ 0 as in (4.8). Set ¼ x .

Compute the portfolio loss LðijÞ ¼ V ð0; SÞ V ðt; S þ S ðijÞ Þ.