Risk2: Measuring the Risk in Value at Risk

Philippe Jorion

The recent derivatives disasters have focused the attention of the finance industry on
the need to control financial risks better. This search has led to a uniform measure of
risk called value at risk (VAR), which is the expected worst loss over a given horizon
at a given confidence level. V A R numbers, however, are themselves affected by
sampling variation, or "estimation risk"--thus, the risk in value at risk itself.
Nevertheless, given these limitations, VAR is an indispensable tool to control
financial risks. This article lays out the statistical methodology for analyzing
estimation error in VAR and shows how to improve the accuracy of VAR estimates.

he need to improve control of financial risks In addition to financial reporting, VAR can be
T has led to a u n i f o r m measure of risk called
value at risk (VAR), which the private sector is
used for a variety of other purposes, such as setting
position limits for traders, measuring returns on a
increasingly adopting as a first line of defense risk-adjusted basis, and m o d e l evaluation. Institu-
against financial risks. Regulators and central tional investors are also embracing VAR as a dy-
banks also p r o v i d e d the impetus b e h i n d VAR. namic m e t h o d for controlling their exposure to risk
The Basle Committee on Banking Supervision factors, especially w h e n m a n y outside f u n d man-
a n n o u n c e d in April 1995 that capital adequacy agers are involved. Nonfinancial corporations, es-
requirements for commercial banks are to be pecially those involved with derivatives, are also
based on VAR. 1 In December 1995, the Securities considering risk-management systems centered
and Exchange Commission issued a proposal that a r o u n d VAR. VAR provides a consistent measure
requires publicly traded U.S. corporations to dis- of the effect of hedging on total risk, which is a
close information about derivatives activity, with significant i m p r o v e m e n t u p o n traditional h e d g i n g
a VAR measure as one ot! three possible m e t h o d s p r o g r a m s that typically focus only on individual
for making such disclosures. Thus, the unmistak- transactions. N o d o u b t these desirable features ex-
able trend is t o w a r d m o r e - t r a n s p a r e n t financial plain the wholesale trend t o w a r d VAR.
risk reporting based on VAR measures. Current implementations of VAR, however,
VAR summarizes the worst'expected loss over have not recognized the fact that VAR measures are
a target horizon within a given confidence interval. only estimates of risk. VAR should be considered
VAR s u m m a r i z e s in a single n u m b e r the global a first-order approximation to possible losses from
exposure to market risks and the probability of adverse financial risk. Although VAR is a vast im-
adverse m o v e s in financial variables. It measures p r o v e m e n t over no measure at all, VAR n u m b e r s
risk using the same units as the b o t t o m l i n e - - cannot be taken at face value. A VAR figure com-
dollars. Bankers Trust, for example, revealed in its bines existing positions with estimates of risk (in-
cluding correlations) over the target horizon. If
1994 annual report that its daily VAR was an aver-
these estimates are based on historical data, they
age of $35 million at the 99 percent confidence level
inevitably will be affected b y "estimation risk";
over one day; this n u m b e r can be readily c o m p a r e d
thus, value at risk also entails risk. 2
with its annual profit of $615 million or total equity
Recognizing the existence of estimation risk
of $4.7 billion. On the basis of such data, sharehold-
has several important consequences. For instance,
ers and managers can decide w h e t h e r they feel
users might w a n t to set the confidence level, usual-
comfortable with a level of risk. If the answer is no,
ly set arbitrarily, to a value that will minimize the
the process that led to the c o m p u t a t i o n of VAR can
error in VAR. Or, the statistical m e t h o d o l o g y might
be used to decide w h e r e to trim risk.
be g u i d e d b y the n e e d to minimize estimation error.
In addition, VAR should be reported with con-
fidence intervals. For instance, a b a n k m i g h t an-
Philippe Jorion is a professor of finance at the Graduate nounce that its VAR over the next day is $35 million
Schoolof Management at the University of California with a 95 percent confidence interval of $32 million
at Irvine. to $38 million. A tight interval indicates relative con-

Financial Analysts Journal • November/December 1996 47

© 1996, AIMR®
fidence in the $35 million estimate, particularly com- VAR = E(W) - W*
p a r e d with a hypothetical interval of $5 million to = W0(, - R*), (1)
$65 million. The latter would say that the VAR n u m -
where W* is the lowest portfolio value at given
ber is quite inaccurate--although not in the range of
confidence level c. Finding VAR is equivalent to
billions. The purpose of this article is to provide a for-
identifying the m i n i m u m value, W*, or the cutoff
mal f r a m e w o r k for a n a l y z i n g estimation e r r o r in
return, R*.
VAR and, more importantly, to discuss methods for
improving the accuracy of VAR measures.
VAR for General Distributions
MEASURING VAR In its most general form, VAR can be derived
from the probability distribution for the future
To formally define a portfolio's VAR, one first m u s t
portfolio value,f(w). At a given confidence level, c,
choose two quantitative factors: the length of the
we wish to find the worst possible realization, W*,
holding h o r i z o n and the confidence level. Both are
such that the probability of exceeding this value is
arbitrary. As an example, the latest proposal of the
c, w h e r e
Basle C o m m i t t e e defines a VAR m e a s u r e using a 99
percent confidence interval over 10 trading days.
The resulting VAR is then multiplied b y a safety c = f f(w)dw, (2)
factor of 3 to arrive at the m i n i m u m capital require- W*
m e n t for regulatory purposes.
Presumably, the 10-day period corresponds to or such that the probability of a value lower than
the time n e e d e d for regulators to detect problems W* is 1 - c, w h e r e
and take corrective action. P r e s u m a b l y also, the
W*
choice of a 99 percent confidence level reflects the
l-c= S f(w)dw" (3)
trade-off b e t w e e n the desire of regulators to ensure
a safe and s o u n d financial system and the adverse
effect of capital requirements on b a n k profits. Dif- In other words, the area from -oo to W* m u s t
ferent choices of horizon and confidence level will s u m to I - c, which might be, say, 5 percent. This
result in trivially different VAR numbers. specification is valid for any distribution, discrete
The significance of the quantitative factors de- or continuous, fat- or thin-tailed. As an example, in
p e n d s on h o w they are to be used. If the resulting its 1994 annual report, J.P. M o r g a n revealed that its
VARs are directly used for the choice of a capital daily trading VAR a v e r a g e d $15 million at the 95
cushion, then the choice of the confidence level is percent level over one day. This n u m b e r can be
crucial. This choice should reflect the c o m p a n y ' s derived from Figure 1, which reports the distribu-
degree of risk aversion and the cost of a loss exceed- tion of J.P. Morgan's daily revenues in 1994.
ing the VAR. H i g h e r risk aversion, or greater costs, From Figure 1, we find the average r e v e n u e is
implies that a larger a m o u n t of capital should be about $5 million. Next, we h a v e to find the obser-
available to cover possible losses, thus leading to a vation (also called a quantile) such that 5 percent of
higher confidence level. the distribution is on its left side. There are 254
In contrast, if VAR n u m b e r s are used only to observations, so we n e e d to find W* such that the
p r o v i d e a c o m p a n y w i d e yardstick to c o m p a r e risks n u m b e r of observations to its left is 254 x 0.05 = 13.
a m o n g different markets, then the choice of confi- This exercise yields W* equal to -$10 million and a
d e n c e level is not v e r y important. A s s u m i n g a nor- daily VAR of $15 million.
mal distribution, disparate VAR measures are easy
to convert into a c o m m o n number. VAR for Normal Distributions
To c o m p u t e the VAR of a portfolio, define W 0 If the distribution can be assumed to be normal,
as the initial investment and R as its rate of return. the computation can be simplified considerably. By
The portfolio value at the end of the target h o r i z o n using a multiplicative factor that is a function of the
is W = W0(1 + R). Define p and ry as the annual m e a n confidence level, VAR can be derived directly from
and s t a n d a r d deviation of R, respectively, and At as the portfolio standard deviation.
the time interval considered. If successive returns First, m a p the general distribution f ( w ) into a
are uncorrelated, the expected return and risk are standard normal distribution cI~(¢), in which the
then p a t and 0c,4r~over the holding horizon. r a n d o m variable ~ has a m e a n of zero and a standard
VAR is defined as the dollar loss relative to deviation of 1. The cutoff return, R*, can be associ-
w h a t was expected; that is, ated with a standard n o r m a l deviate o~ such that

48 @Association for Investment Management and Research

 Figure 1. Measuring Value at Risk

20
VAR=$15million ~
5% of Occurrences

15

10
E
Z

0
<-25 -20 -15 -10 -5 0 5 10 15 20 25
J.P. Morgan Daily Revenue ($ millions)

gAt-R* cffoW0) = 1.65 x $9.2 million

-ct-- tj4r~ (4) = $15.2 million.

Then, VAR m a y be f o u n d in terms of portfolio This n u m b e r is v e r y close to the VAR obtained

value W*, cutoff return R*, or n o r m a l deviate c~;that from the general distribution, s h o w i n g that the
is, n o r m a l a p p r o x i m a t i o n p r o v i d e s a good estimate
of VAR.
W*

l-c= J f(w)dw
Sigma-Based VAR
R*
Generally, this m e t h o d applies to any proba-
bility function besides the normal, a convenient
= Stir)dr attribute because m a n y financial variables have
fatter tails (i.e., m o r e extreme observations) than
the n o r m a l distribution. Most notably, the stock
(5) m a r k e t crash of October 1987 was a 20 standard
deviation e v e n t - - o n e that u n d e r a n o r m a l distribu-
tion should n e v e r have h a p p e n e d . This behavior is
To report VAR at the 95 percent confidence level,
particularly w o r r i s o m e because VAR attempts to
for example, the 5 percent left-tailed deviate from
describe tail behavior precisely.
a standard n o r m a l distribution can be f o u n d from
One possible explanation is that volatility
standard n o r m a l tables as 1.645. Once 0~ is identi-
changes t h r o u g h time, increasing in times of great-
fied, VAR can be recovered as
er than n o r m a l turbulence. A stationary m o d e l
VAR= W0 x c~a,f~. (6) might then erroneously view large observations as
outliers w h e n they are really d r a w n from a distri-
The key result is that VAR is associated with the b u t i o n with temporarily greater dispersion. In-
standard deviation only. deed, the recent literature on time variation in
For instance, for the J.P. M o r g a n example from second m o m e n t s provides o v e r w h e l m i n g evidence
Figure 1, the standard deviation of the distribution that variances on a variety of financial assets do
is $9.2 million. Therefore, the normal-distribution change over time. 3
VAR is Even controlling for time variation, however,

Financial Analysts Journal • November/December 1996 49

residual returns still appear to be fat tailed. One pling variation leads to material changes in VAR.
simple method to account for these tails is to model This possibility is why sensitivity analysis of
a Student t distribution, which is characterized by an VAR is a useful exercise. Beder (1995), for instance,
additional parameter called "degrees of freedom" compared VAR results from different models. Risk
(v) that controls the size of tails. As v grows large, was measured at a two-week (10-day) horizon at
the distribution converges to a normal distribution. the 5 percent level for a $1 million bond portfolio.
Table 1 provides estimates of the Student pa- Historical-simulation methods based on the previ-
rameter for a number of daily price returns over the ous 100 and 250 days yielded VARs of $2,000 and
1990-94 period. Typically, the parameter v is in the $17,000, respectively. The RiskMetrics method
range of 4 to 8, which confirms the existence of fat yielded a VAR of $18,200. These discrepancies ap-
tails. pear to be wide and unsettling.
U p o n f u r t h e r inspection, the historical-
Table 1. Estimates of the Student Parameter simulation method using 100 days appears to have
Estimated been inadequate because of the small number of
Asset Parameter effective observations: only 10 (100 historical obser-
U.S. stocks 6.8 vations divided by the horizon of 10 business days).
DM/US$ exchange rate 8.0 Such a small sample size makes the 5 percent left
DM/£ exchange rate 4.6
tail of a distribution difficult to, measure. Other-
U.S. long bond 4.4
U.S. three-month T-bill 4.5 wise, the remaining measures of VARs are in line
with each other.
To find the VAR under a Student distribution, This experiment demonstrates the need for a
Equation 5 still applies, but • is replaced by the good understanding of the methodology behind
standard Student distribution and cz by the appro- VAR. The question is whether discrepancies arise
priate (1 - c) deviate. For example, for a Student because of fundamental differences in methodolo-
with v equal to 6 at the 95 percent confidence level, gies or simply because of sampling variation.
oc equals 1.943.
Thus, for many distributions, the dispersion Estimation Error in Quantile-Based
can be summarized by one parameter, the standard VAR
deviation. This approach applies to most financial For arbitrary distributions, the cth quantile can
prices, stock prices, bond prices, exchange rates, be empirically determined from the historical dis-
and commodities. Of course, it is inappropriate for tribution as O(c). Of course, some sampling error is
strongly asymmetric distributions, such as posi- associated with the statistic. Kendall (1994), for
tions in options. With large portfolios, such as trad- instance, showed that the asymptotic standard er-
ing portfolios of commercial banks, however, the ror of the sample quantile, 0, is derived as
issue is one of fatness in tails, not asymmetry.
se(q) = 4 rf(q) 2 (7)
EVALUATING VAR
So far, much of the analysis has been standard fare. where T is the sample size and f(. ) is the probability
What is less recognized, however, is the effect of distribution function evaluated at the quantile q.
estimation error. Indeed, all VAR measures are Kupiec (1995) pointed out that this standard error
merely estimates. VAR measures are exact only can be quite large and argued that this method does
when the underlying distribution is measured with not provide "suitable benchmarks" for measuring
an infinite number of observations. In practice, data VAR. In particular, the standard error increases
are available for only a limited time period. markedly as the confidence level increases. In other
Various methods can be used to create VAR words, the estimate grows increasingly unreliable
measures. "Historical-simulation" methods repli- farther into the left tail; that is, the I percent left tail
cate the behavior of the current portfolio over a is less reliable than the 10 percent quantile.
sample of previous days. "Delta-normal" methods This phenomenon is illustrated in Figures 2
summarize risk factors by a variance-covariance and 3, where the expected quantile and two stan-
matrix estimated from historical data. J.P. Mor- dard error intervals are plotted for the normal and
gan's RiskMetrics system, for instance, provides an Student distributions, respectively.
application of the delta-normal method in which For the normal distribution, the 5 percent left-
risk measures are time varying. In each case, using tailed interval is centered aromad 1.645. With T
different time periods will invariably lead to differ- equal to 100, a two standard error confidence inter-
ent measures of VAR. The issue is whether sam- val is 1.24 to 2.04, which is quite large. With 250

50 @Association for Investment Management and Research

 tion with 6 degrees of freedom, typical of financial
Figure 2. Confidence Bands for Sample Quan-
data. The figure shows that the (x deviates are great-
tile: Normal Distribution
er for the Student distribution than for the normal
distribution; confidence bands are also much wid-
er. For instance, the 5 percent quantile has an ex-
pected value of 1.943 and a band of [1.33, 2.56] with
T equal to 100. This interval is much wider than for
the normal distribution.
These observations bring us back to an unre-
solved issue in the computation of VAR: The choice
£y of the confidence level is completely arbitrary, es-
pecially since the Basle Committee has decided to
multiply the VAR by another arbitrary factor, 3.
• . ~ ~--~.~-~ ~ .....
Commercial banks now report their VARs with
various incompatible parameters. 4 Given this arbi-
trariness, we might want to choose a level c that
10 20 30 40 50 provides the best sampling characteristics•
Left-Tail Probability (%)

Expected -- -- - T = 250 Estimation Error in Sigma-Based VAR

. . . . T = 1,000 ....... T = 100 Additional precision might be gained by di-
rectly measuring the standard deviation, which can
be multiplied by an appropriate scaling factor to
obtain the desired quantile.
observations, which correspond to one year of trad- For the normal distribution, for instance, the
ing days, the interval is still 1.38 to 1.91. With T equal VAR can be computed in two steps: First, compute
to 1,000, the interval shrinks to 1.51 to 1.78. The the sample standard deviation, s; then, multiply the
interval widens substantially as one moves to more number by a scaling factor, (x(c), to obtain the de-
extreme quantiles. For instance, the same interval sired confidence level--say, 1.645 for a 95 percent
for the I percent quantile is 2.09 to 2.56 with T equal confidence level for a normal distribution•
to 1,000. As expected, more imprecision is found in Using this method, the standard error of the
the extreme tails, which have fewer data points. estimated quantile is
Contrast these results with Figure 3, which
displays confidence bands for a Student t distribu- se(ocs) = a x s e ( s ) . (8)

This method leads to substantial efficiency gains

Figure 3. Confidence Bands for Sample Quan- relative to using the estimated quantile. In the case
tile: Student t Distribution of the normal distribution, for instance, we know
that the sample standard deviation is a sufficient
il: statistic for the dispersion and also is the most
efficient; that is, it has the lowest standard error.
\'. Intuitively, this efficiency is explained by the fact
that s uses information about the whole distribu-
"'•",.•..•.. tion (in terms of all squared deviations around the
mean) but a quantile uses only the ranking of ob-
£y servations and the two observations around the
estimated value.
For the normal distribution, we have an analyt-
ical formula for the standard error of s, which is

se(sl~) = C~T. (9)

1 10 20 30 40 50
Left-Tail Probability (%)
Figure 4 displays the standard error of the estimat-
Expected -- -- -- T = 250 ed quantile using the two methods applied to a
....... T = 1,000 ....... T = 100 normal distribution with T equal to 250. As theory
would suggest, the sample standard deviation
method has uniformly lower standard errors and

Financial Analysts Journal • November/December 1996 51

Figure 4. Standard Error of VAR for Different precision can be obtained by switching estimation
Estimation Methods technique. Bankers Trust, for example, reported a
daily VAR of $35 million at the 99 percent level.
0.6 Assuming this number came from the sample
quantile based on one year of data from a Student
0.5 t distribution, the associated confidence interval is
[$24, $46] million--not particularly tight. In con-
~©0 . 4 trast, if the number were based on the sample stan-
dard deviation method, the interval would shrink
to [$30, $40] million.
2 One remaining issue is whether the distribu-
"~ 0.2
tions underlying the simulations adequately repre-
I" sent empirical distributions Of financial returns.
0.1
/ / Table 4 presents efficiency comparisons for an ac-
I tual distribution: daily returns on the on-the-run
0
-3 -2 -1 0 1 2 3 30-year bond from 1990 to 1994. The distribution is
N u m b e r of Standard Deviations free to take any form--fat tailed, asymmetric, and
so on. I took this distribution as the true distribu-
- - Quantile Method Standard Deviation Method
tion, then bootstrapped it--that is, sampled with
- - - - - -

replacement--to generate random drawings. As

before, the table shows that the standard deviation
is, therefore, uniformly superior to the sample method is much more efficient than the sample
quantile method. quantile method, with standard errors typically
A similar adjustment can be made for the Stu- about one-third to two-thirds lower.
dent t distribution, for which the quantile can be These results clearly indicate that estimation
estimated from the sample standard deviation. error in the estimated quantiles can be substantially
Without analytical results for se(s), however, simu- reduced by using a multiple of the sample standard
lations must be used to determine the advantage of deviation. When the distribution is normal, the
the s-based estimator over the usual quantile. Sim- deviates can be obtained from a normal distribu-
ulations are also useful for assessing the small- tion table. In practice, when the tails appear to be
sample properties of these estimators. too fat, a Student t distribution, from which devi-
ates are readily available, can be fitted easily. 5
Comparisons of Methods
Tables 2 and 3 describe VAR statistics from CONCLUSIONS
drawings from normal and Student t distributions, The rapidly spreading use of VAR must be seen as
respectively, using simulations based on 10,000 a vast improvement over antiquated or nonexistent
replications. Two methods are compared: Method risk-management practices, some of which have
1 is based on the sample quantile, and Method 2 is caused financial disasters. One of the lessons of the
based on the sample standard deviation. Orange County, California, bankruptcy, for exam-
For Table 2 with one year of data (T = 250), for ple, is that municipalities investing in the pool
example, the standard errors of VAR estimated would have been more careful had the value at risk
using the two methods are generally in close accor- of their investment been clearly explained to them. 6
dance with the asymptotic numbers but the stan- In addition, investors would not have had the ex-
dard deviation method is about twice as efficient as cuse that they did not know what they were getting
the quantile method. For instance, the standard into, which would have limited the rash of lawsuits
error of Method I is 0.133, as compared with 0.074 against third parties. This experience demonstrates
for Method 2, an improvement of about 45 percent. w h y regulators now embrace VAR as a means of
The advantage is also substantial for the Stu- improving transparency and stability in financial
dent t distribution. For instance, using the same markets.
parameters as before, the standard error of Method The benefits of VAR should not, however,
1 is 0.200 as opposed to 0.132 for Method 2, an mask its shortcomings. Any VAR number is itself
improvement of about 35 percent. The advantage is measured with some error, or estimation risk.
even greater for quantiles farther in the tail, where Thus, understanding the statistical methodology is
the improvement is on the order of 60 percent. important in order to interpret VAR estimates. This
In practical terms, substantial improvement in interpretation would be made easier not only by

52 ©Association for Investment Management and Research

 to

~K
UJ

to
"O

<

oC

< "O
" ~Vq' V"T?

"O

o=

"O

<

to

toe
to
Z O tO

<
~ ff~ ff
2o

biD,.o

N ~.~
to
I-- b3 to

Financial Analysts Journal • November/December 1996 53

 t~

£
= 4 fff~ ~ffff

~ <<<
o~ ZZZ
ddd

b
<
g
o
~ 2 0
II "~
0 0 0 ~ 0 0 ~
N
o I ~ d ddd ddd

g~
m

~ m ZZZ
0 0 0
m

~ 2
m
0 0 0 0 0 0 0
o

~kg ~NN

m
~ zzz o
0 0 0

° ~O
g
. m
II

o o o o o o 0

ii
t~ £~
, io

t~
, m

I--

54 @Association for Investment Management and Research

 x

d ddd ddd

o %
H

<

111
g
d,

f-I
E
•

"MM'~
<

d ddd ddo ~
o o
No

<
W

.o
ffl

4 <

Financial Analysts Journal ° November/December 1996 55

reporting a single VAR n u m b e r but also b y report- In the end, the greatest benefit of VAR m a y lie
ing a confidence b a n d a r o u n d it. in the imposition of a structured m e t h o d o l o g y for
The main p u r p o s e of this article was to show thinking critically about risk. Financial institutions
the extent to which VAR is affected b y estimation that go t h r o u g h the process of c o m p u t i n g their
risk and also to make r e c o m m e n d a t i o n s on h o w to VARs are forced to confront their exposure to finan-
measure quantiles. For the distributions consid- cial risks and to set u p a risk-management function
ered here, estimating quantiles from a multiple of to supervise the front and back offices. Thus, the
the sample standard deviation is far superior to process of getting to VAR m a y be as i m p o r t a n t as
estimating them directly from sample quantiles. the n u m b e r itself. Nevertheless, VAR is u n d o u b t -
Therefore, recognizing estimation error can lead to edly here to stay.
better m e a s u r e m e n t methods.

NOTES
1. The Basle Committee consists of central bankers from a risk can also capture structural changes as long as the
group of 10 countries. This committee sets minimum changes do not occur too abruptly.
standards for capital requirements in member countries. 4. For instance, Bankers Trust uses a 99 percent level of confi-
2. In addition, for complex portfolios involving positions in dence; Chemical Bank and Chase Manhattan, a 97.5 percent
options that are difficult to price, VAR is also affected by level; Citibank, a 95.4 percent level; and BankAmerica and
"model risk," which results from differences in valuations J.P. Morgan, a 95 percent level.
that can be traced to different option valuation models.
5. These simulations assumed that the underlying distribution
Another conceptual problem is that, especially over long
was known. If the distribution is truly irregular because of,
horizons, VAR does not account for changing positions.
Sound risk-management practices typically reduce the size for example, heavy optionality in the portfolio, nonparamet-
of positions in response to losses or increasing volatility, ric methods such as kernel estimation can be used to provide
which decreases the worst loss relative to a static VAR estimates of the quantile and associated standard errors.
measure. For further analysis of limitations of VAR methods, These methods lead to improved precision by smoothing the
see Jorion (1996). distribution. See, for example, Sheather and Marron (1990).
3. For a review, see, for instance, Bollerslev, Chou, and Kroner 6. For a description of the Orange County disaster, see Jorion
(1992). Time-series models that allow for time variation in (1995).

REFERENCES
Beder, Tanya Styblo. 1995. "VAR: Seductive but Dangerous." Kendall, M. 1994. Kendall's Advanced Theory off Statistics. New
FinancialAnalysts Journal, vol. 51, no. 5 (September/October):12-24. York: Halsted Press.
Bollerslev, Tim, R. Chou, and K. Kroner. 1992. "ARCH Modeling Kupiec, P. 1995. "Techniques for Verifying the Accuracy of Risk
in Finance: A Review of the Theory and Empirical Evidence." Measurement Models." Journal of Derivatives, vol. 2 (Decem-
Journal of Econometrics, vol. 52, no. 1 (April/May):5-59. ber):73-84.
Jorion, Philippe. 1995. Big Bets Gone Bad: Derivatives and Sheather, S., and J. Marron. 1990. "Kernel Quantile Estimators."
Bankruptcy in Orange County. San Diego: Academic Press. American Statistical Association Journal, vol. 85 (June):410-16.
. 1996. Value at Risk: The New Benchmark for Controlling
Market Risk. Chicago: Irwin.

56 @Association for Investment Management and Research