You are on page 1of 13

MAR: Seductive but Dangerous

Tanya Styblo Beder

Value at risk (VAR) has gained rapid acceptance as a valuable approach to risk management.
Not all VARs are equal, however. A study of VAR techniques used by dealers and end-users
reveals that VAR calculations differ significantly for the same portfolio. VARs are extremely
dependent on parameters, data, assumptions, and methodology. Calculation of eight common
VARs for three hypothetical portfolios demonstrates the potentially seductive but dangerous
nature of any single approach to risk management. In sum, although VAR and other
quantitative techniques are necessary aspects of an effective risk-management program, they
are not sufficient to control risk.

ralue at risk is Wall Street's latest advancement dealers use either an internal methodology or a
V in risk measurement. Simply defined, VAR is Bank for International Settlements (BIS) standard
an estimate of maximum potential loss to be ex- methodology to compute VAR and that the results
pected over a given period a certain percentage of be multiplied by a factor of three to determine the
the time. Its simplicity is seductive. Used to the amount of capital to be set aside for market risk.
extreme, in a single statistic, a firm can measure its Our research indicates that this amount may be too
exposure to markets worldwide. VAR enables a high or too low, depending upon the method used.
firm to determine which businesses offer the great- The need for a uniform VAR methodology or for
est expected returns at the least expense of risk. differing multiplication factors according to the
When one considers that risk management in the type of VAR is paramount to establish a common
early 1970s consisted almost entirely of the evalua- ground for comparative purposes.
tion of credit risk, VAR's power in the context of the In our analysis, historical simulations present
galaxy of risks we track, analyze, and manage quite different views of risk relative to Monte Carlo
today is breathtaking to consider. simulations. This difference is attributable to the
VAR can be dangerous, however. A review of extreme dependence of historical simulations on
dozens of dealers' and end-users' VARs revealed the underlying data set and the value of the relative
radically different approaches to the calculation. In randomness of key variables in Monte Carlo simu-
this study, eight common VAR methodologies were lations compared with sample-specific values. The
applied to three hypothetical portfolios. As illus- results also reveal the exceptional time sensitivity of
trated in Figure 1, the magnitude of the discrepancy certain portfolio risks and highlight the potential
among these methods is shocking, with VAR re- failure of VAR, even when bolstered by stress
sults varying by more than 14 times for the same testing. In sum, although VAR and stress testing are
portfolio. These results illustrate the VAR's extreme necessary, they are not sufficient to contain risk.
dependence on parameters, data, assumptions, and The differences in common VARs emphasize
methodology. the fact that no single set of parameters, data,
The implications of these discrepancies for assumptions, and methodology is accepted as the
capital adequacy standards are significant, espe- "correct" approach. Even if two firms use the same
cially given the Basle Committee on Banking Su-
quantitative technique, they often apply different
pervision's treatment of VAR in its proposed
assumptions in implementing the technique. For
amendment to The 1988 Basle Capital Accord, "The
example, some firms calculate the global VAR of
Supervisory Treatment of Market Risks," published
the firm over a one-day time horizon, using histor-
on April 12, 1995. This amendment proposes that
ical data series on markets and a specific set of
mathematical models. Others calculate regional
Tanya Styblo Beder is a principal of Capital Market Risk Advisors, VAR of product areas over a monthly or annual
Inc., in New York. time horizon, using random or implied data series

12 Financial Analysts Journal / September-October 1995

© 1995, AIMR®
Figure 1, Range of VARS: All Simulations ical observations. The BIS/Basle proposal permits
correlation only within asset classes, not across,
16 effectively forcing the correlation between asset
14 classes to be plus or minus 1, whichever produces
12 the higher estimate of VAR.
4 10 The three portfolios were constructed to haw~

~3 8 increasing complexity in terms of optionality

6 ~ a n d / o r asset class composition and possess prop-
4 erties sought frequently by dealers and end-users.
2 The eight VAR calculations performed for each
0 portfolio are summarized in Table 1.
Portfolio 1 Portfolio 2 Portfolio 3
Portfolio 1
• Low [ ] High [ ] Multiple Portfolio I consists exclusively of U.S. Treasurv
strips. It was designed to satisfy three conditions at
construction: (1) The duration of the portfolio
equals that of the ten-year strip, 2 (2) the portfolio
on markets and multiple mathematical models. has greater convexity than the ten-year strip, and
Depending on the selection of time horizon, data (3) the portfolio performs at least as well as the
base, and correlation assumptions across instru- ten-year strip under a 100 basis point parallel
ment/asset classes, the same model may produce increase or decrease in the Treasury yield curve or
widely divergent VAR views for the same portfolio under an inversion of the Treasury yield curve.
and, therefore, different capital requirements. Table 2 describes the composition of and con-
straints on Portfolio 1, which consists of a long
THE PORTFOLIOS A N D VAR CALCULATIONS position in 2-year and 30-year U.S. Treasury strips.
In the remainder of this article, we describe the The ten-year U.S. Treasury strip, the benchmark, is
common VAR calculations and apply them to three included in the table for reference. The net invest-
hypothetical portfolios. For each methodology pre- ment in Portfolio 1 is $1 million.
sented, VAR is calculated for both one-day (ld) and The traditional risk measures show that for
two-week (2w) time horizons. The first methodol- very small parallel shifts in the Treasury yield
ogy, historical simulation, is performed twice, curve, Portfolio 1 performs similarly to the ten-year
changing the data base used from the past 100 strip. Moreover, for the plus or minus 100 basis
trading days (Prl00d) to the past 250 trading days point parallel yield curve shifts and inversion, the
(Pr250d). The second methodology, Monte Carlo portfolio's $1 million investment performs slightly
simulation, also is performed twice, changing the better than if it been invested in the ten-year strip.
correlation estimates from the JP Morgan RiskMet- The VAR analyses of Portfolio 1 reveal quite
rics data set to those from the BIS/Basle Committee different risk profiles than older risk measures such
proposal. 1 Differences in correlation estimates be- as duration, convexity, and scenario analysis. Fig-
tween RiskMetrics and BIS/Basle are significant. ure 2 displays the eight common VAR calculations
RiskMetrics permits correlation across all asset for this portfolio. The VAR results place signifi-
classes, using exponentially weighted daily histor- cantly different degrees of capital at risk both

Table 1. Eight Common VAR Calculations

VAR Type of Data Base/Correlation Holding
Approach Simulation Assumption Period

1 Historical Prior 100 trading days One day

2 Historical Prior 250 trading days One day
3 Monte Carlo Historical, RiskMetrics correlations One day
4 Monte Carlo Historical, BIS/Basle correlations One day
5 Historical Prior 100 trading days Two weeks
6 Historical Prior 250 trading days Two weeks
7 Monte Carlo Historical, RiskMetrics correlations Two weeks
8 Monte Carlo Historical, BIS/Basle correlations Two weeks

Financial Analysts Journal/September-October 1995 13

Table 2. Portfolio 1: Composition and Constraints
10-Year Strip
Characteristic 2-Year Strip 30-Year Strip Total Portfolio (Benchmark)
Yielda 5.91% 6.85% 6.58%
Priceb 89.12 14.94 52.42
Face amount $779,778 $2,041,424 $2,281,202 $1,907,670
Purchase amount $694,964 $305,036 $1,000,000 $1,000,000
Duration and convexity
Duration 1.712 4.078 5.063
Duration contribution 1.335 8.325 9.660
Convexity 0.041 1.133 0.514
Convexity contribution 0.032 2.312 2.344
Scenario analysis
Yield + 100 bp 6.91% 7.85% 7.58%
Price/yield + 100 bp 87.43 11.38 47.60
Position/yield + 100 bp $681,775 $232,333 $914,108 $908,051
Yield - 100 bp 4.91% 5.85% 5.58%
Price/yield - 100 bp 90.86 19.64 57.75
Position/yield - 100 bp $708,743 $401,017 $1,109,490 $1,101,679
Yield curve inversion 7.20% 6.30% 6.58%
Price/inversion 86.95 17.37 52.42
Position/inversion $678,009 $354,508 $1,032,517 $1,000,000
Note: Prices and yields as of 5/25/95. The maturities of the 2-year, 10-year, and 30-year strips are 5/15/97, 5/15/05, and 8/15/23,
"The market yield for each strip is stated on an actual/365 basis with semiannual compounding.
b The price of each strip is stated as a percentage of face amount.

w i t h i n a n d b e t w e e n m e t h o d o l o g i e s . T h e V A R sta- s u m p t i o n s m a d e to p e r f o r m h i s t o r i c a l s i m u l a t i o n
tistics to the r i g h t of each b a r m a y b e i n t e r p r e t e d as o v e r the p r i o r 2 5 0 - d a y p e r i o d , the p r o b a b i l i t y is 1
follows: U n d e r the a s s u m p t i o n s specific to the p e r c e n t t h a t a loss e q u a l to or e x c e e d i n g 1.29
p a r t i c u l a r V A R c a l c u l a t i o n , the p r o b a b i l i t y is 5 p e r c e n t of the $1 m i l l i o n p o r t f o l i o i n v e s t m e n t w i l l
p e r c e n t (1 p e r c e n t ) t h a t the p o r t f o l i o w i l l suffer a occur over a one-day time horizon.
loss g r e a t e r t h a n or e q u a l to the statistic s h o w n . For F i g u r e 3 c o m p a r e s the r e s u l t s of the h i s t o r i c a l
the t h i r d set of bars, for e x a m p l e , u n d e r the as- s i m u l a t i o n for Portfolio 1 w i t h the r e s u l t s of the

Figure 2. MAR Calculations: Portfolio 1

Hist. I Pr 100 d ld
~ 0 . 6 7
Hist. I Pr 100 d 2w 0 ' 2 0 ~ 0 . 8 2

Hist. I Pr 250 d ld
o 1.29
o Hist. I Pr 250 d 2w ] _ ~ / .7 2.14

MontC I RiskMetrics ld ~
0 . 8 8
MontC I RiskMetrics 2w , _ 1!82
~ 2 . 6 1
MontC I BIS/Basle ld • 10.68 0.97

MontC I BIS/Basle 2w
0.5 1.0 1.5 2.0 2.5 3.0
VAR (%)
[] 5% Probability [] 1% Probability

14 Financial Analysts Journal September-October 1995

Figure 3. Distributions of One-Day and Two-Week Returns: Portfolio 1

One-Da~' Returns

g 14

12 -
,.Q 10 -
2 8 -
,4 Z
6 -
4 -

2 -
-2.0 -1.6 -1.2 -0.8 -0.4 0 0.4 0.8 1.2 1.6 2.0

Return (%)

Two-Week Returns



g2 8

4 -

2 -

-4.0 -3.2 -2.4 -1.6 -0.8 0 0.8 1.6 2.4 3.2 4.0

Return (%)

-- -- -- RiskMetrics 100 D a y s

BIS/Basle ....... 250 Days

Monte Carlo simulations for one-day and two-week rate in a trending market, they will be less accurate
returns. 3 As illustrated by the graphs, the historical when the trend changes•
simulations present a different view relative to the As summarized in the bar charts in Figure 2,
Monte Carlo simulations• This result is attributable the result for the prior 100 trading days, 5 percent
to their extreme dependence on the underlying data probability VAR equals 0.49 percent over a one-day
set. During the 100-day and 250-day periods in- horizon but then drops to 0.20 percent over a
cluded in the historical simulations, the value of two-week period• For all other VAR types, the VAR
Treasury strips largely appreciated• Had a period of result increases with the time horizon, as would be
rising interest rates been selected, the result would expected• This surprising result is explained by the
have been the opposite. The danger in basing VAR pattern of results during the specific historical pe-
estimates on relatively short periods of direct his- riods. Although the average return is positive dur-
torical observations is apparent--history must re- ing the first 100 trading days (the left side of Figure
peat itself for the results to predict the future• 4), negative returns are more common over the
Although historical estimates may be fairly accu- one-day time horizon than over the two-week time

Financial Analysts Journal / September-October 1995 "15

Figure 4. Historical One-Day Returns and Two-Week Returns: Portfolio 1
One-Day Returns
.2 0

Two-Week Returns
1 14 27 40 53 66 79 92 105 118 131 144 157 170 183 196 209 222 235 248

horizon. Thus, VAR is higher for the one-day time potential capital exposure and expected profit over
horizon than for the two-week time horizon. a short-term or a long-term horizon? In terms of
Several other conclusions follow from this set risk-reward appetite, will the firm be satisfied if
of VAR results, as expected given interest rate losses m o u n t for two years but huge profits make
trends at the time. Monte Carlo simulations indicate u p for the losses in the third year? Or, is less erratic
higher expected losses than does the 100-day his- performance desired? Often, more-even perfor-
torical simulation but lower expected losses than mance is desired b y firms that report public finan-
for the 250-day historical simulation. Historical cial information, as well as by funds that must
simulations indicate that increasing the holding publish daily net asset values. Thus, two firms
period from one day to two weeks decreases the performing identical VAR calculations, other than
expectation of the highest expected profits and selection of time horizon, m a y have different but
increases the m a g n i t u d e of expected losses. Note not necessarily inconsistent VAR results.
that although Monte Carlo simulations indicate a Although a model m a y p r o d u c e adequate
similar change in the expectation of the highest views of capital at risk on an overnight or weekly
profits, they predict a larger increase in the magni- basis, it m a y p r o d u c e inadequate risk views over
tude of expected losses (three times versus two time horizons of several months, a year, or longer.
times). The difference in VAR driven b y the relative For example, the calculation of short-horizon VAR
r a n d o m n e s s of key variables in Monte Carlo versus m a y be misleading for customized or exotic prod-
sample-specific historical simulations is clear. ucts that cannot be liquidated u n d e r the assumed
Time horizon is clearly a crucial parameter in time horizon. To the degree that multiple horizons
VAR. Firms select quite different time horizons to are required, risk systems rarely are capable of
view their risk. Does the firm wish to analyze its incorporating them in aggregating portfolio risks,

16 Financial Analysts Journal/September-October 1995

creating a limitation in sizing true risk exposures. distinguished by inclusion of outlier events, are
Note that some firms address this problem by likely to produce different VAR calculations.
adjusting midmarket valuations. Yet another challenge in data set selection is
Yet another challenge is that although longer determining whether an outlier event is an indica-
time horizons may be appropriate for instruments tion of structural change in the market. For exam-
such as illiquid, path-dependent options, some ple, the prepayment patterns for mortgage-based
mathematical functions are inaccurate beyond securities in the United States have changed funda-
small market moves. For example, many mathe- mentally during the past few years, driven by
matical models are incapable of handling disconti- mortgage broker activity. Prior to the change, a
nuities such as market gapping, or they require drop in interest rates had to prevail for several
strict assumptions such as linearity to produce months before home owners refinanced their morb
accurate information. The April 12, 1995, proposed gages. Subsequently, the refinancing lag shortened
amendment to the 1988 Basle Capital Accord sug- from months to weeks during the rally that ended
gests that firms use a single time horizon of two with the Federal Reserve's interest rate hike in
weeks (ten business days) for VAR calculations. February 1994. Use of historical prepayment data
The selection of data sets is another critical could thus be misleading in determining the ex-
component of VAR. As Portfolio 1 illustrates, alter- pected life of many mortgage securities.
nate data sets may produce vastly different risk To reduce dependence on historical data, given
views. In our experience, intraday versus end-of- that history may not repeat itself, some firms use
day data often produce contrary views of risk data sets based on implied market information.
Sensitivity of the VAR calculation not only to
during periods of high volatility. Different risk
exclusion of any data points but also to use of
views will also be created by the use of historical as
implied versus historical data and to the specific
opposed to market-implied data. Although histori-
time period covered by the data set should be tested
cal data are most often used to calculate VAR, the
to reveal a particular VAR's dependence on such
length of the historical period selected varies sig-
nificantly from firm to firm. Several firms and
software packages use a 90-day historical time
Portfolio 2
horizon, but many market participants believe that,
Portfolio 2 consists of outright and options
at a minimum, a one-year data set should be used.
positions on the S&P 500 equity index contract. This
The proposed Basle amendment suggests that firms
portfolio was designed to satisfy several conditions
use a one-year-minimum data set for VAR calcula-
at construction: (1) The delta, or price change of the
portfolio, equals that of the S&P 500; (2) the gamma,
Length of time is not the sole criterion to
or convexity of the portfolio, is non-negative; and
establish regarding the data set. Sampling fre-
(3) the portfolio significantly outperforms the S&P
quency must be set high enough to ensure that the
500 equity index contract under downward shocks.
data set is statistically significant. For example, a Table 3 describes the composition of and con-
one-year data base composed of 12 end-of-month straints on Portfolio 2, which consists of a long
data points may be no more relevant than a data set position in the S&P 500 equity index contract plus
of 12 points selected through random chance. Of- long and short options on the same index. As with
ten, sampling frequency is also sensitive to the time Portfolio 1, the net investment in Portfolio 2 is $1
the data are collected. For example, end-of-day data million.
points are likely to produce a different VAR picture The traditional risk measures show that the
than daily high/low/close data points. portfolio outperforms the S&P 500 equity index
After type, sampling frequency, and length of contract by an approximate factor of five times
data base are selected, the VAR user must deter- under a negative 20 percent shock to the S&P 500.
mine whether to exclude certain data points. For This outperformance in a bearish scenario is accom-
example, should the data set include "outliers" plished at the price of underperformance during a
caused by one-time events, market gapping, or comparable rise in the S&P 500. The portfolio's $1
other dislocations? Such occurrences are often char- million investment is preserved in the value of the
acterized as extreme but low-probability events. S&P 500 equity index contract.
Recent examples are the devaluation of the Mexican Figure 5 summarizes the VAR results for Port-
peso, the 1987 stock market crashes, and commod- folio 2. Over a one-day horizon, VAR ranges be-
ity volatility during the Gulf War. Two data bases, tween 0.69 percent to 0.91 percent with a 5 percent

Financial Analysts Journal / September-October 1995 17

T a b l e 3. P o r t f o l i o 2: C o m p o s i t i o n and Constraints
Instrument Jun 520 Jun 545 Sep 530 Dec 540 S&P 500 Portfolio
Type Put Call Call Put Long
Strike versus market +20 +45 +30 +40 0
Price 1.95 0.60 14.90 18.45 528.59
Number 4,157.40 -28,723.80 19,784.80 11,617.00 945.90
Purchase amount $8,107 ($17,234) $294,793 $214,335 $499,999 $1,000,000
Delta and gamma
Unit delta -0.239 0.105 0.545 -0.503 1.000
Delta contribution -0.001 -0.003 0.011 -0.006 0.001 0.002
Unit gamma 0.023 0.015 0.012 0.009 0.000
Gamma contribution 0.000 0.000 0.000 0.000 0.000 0.000
Scenario analysis
S&P + 20% 634.31 634.31 634.31 634.31 634.31
Price/S&P + 20 0.00 90.02 107.91 0.25 634.31
Position/S&P + 20 $0 ($2,585,725) $2,134,983 $2,897 $600,001 $152,156
S&P - 20% 422.87 422.87 422.87 422.87 422.87
Price/S&P - 20% 96.12 0.00 0.00 106.91 422.87
Position/S&P - 20% $399,623 $0 $49 $1,241,941 $399,997 $2,041,610

e x p e c t a t i o n a n d b e t w e e n 1.07 p e r c e n t to 1.30 per- F i g u r e 6 c o m P a r e s the r e s u l t s of the h i s t o r i c a l

cent with a 1 percent expectation. Over a two-week s i m u l a t i o n for Portfolio 2 w i t h the r e s u l t s of the
horizon, VAR increases significantly a n d ranges M o n t e C a r l o s i m u l a t i o n s for o n e - d a y a n d t w o - w e e k
b e t w e e n 2.31 p e r c e n t to 3.93 p e r c e n t w i t h a 5 r e t u r n s , r e s p e c t i v e l y . O f n o t e is the s i g n i f i c a n t
p e r c e n t e x p e c t a t i o n . T h e t w o - w e e k , 1 p e r c e n t ex- c h a n g e i n r e t u r n p a t t e r n s b a s e d o n i n c r e a s i n g the
p e c t a t i o n V A R s are fairly c o n s i s t e n t across m e t h o d - h o l d i n g p e r i o d f r o m o n e d a y to t w o w e e k s . I n the
ologies, r a n g i n g b e t w e e n 3.48 p e r c e n t a n d 3.95 case of o n e - d a y r e t u r n s , all s i m u l a t i o n s d i s p l a y
p e r c e n t as i l l u s t r a t e d b y the b a r charts. low-probability h i g h - r e t u r n / l a r g e - l o s s expecta-

Figure 5. M A R Calculations: Portfolio 2

Hist. I Pr100d I l d ~!~%~~ % 0.69

Hist. I Pr 100d 12w

e~ Hist. I Pr 250 d ld
a= Hist. I Pr 250 d 2w

MontC I RiskMetrics ld
MontC I RiskMetrics 2w

MontC I BIS/Basle ld

MontC I BIS/Basle 2w
0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5
VAR (%)
[] 5% Probability [] 1% Probability

18 Financial Analysts Journal/September-October 1995

Figure 6. Distributions of One-Day and Two-Week Returns: Portfolio 2

One-Day Returns

16 -

14 -

12 -

10 -

8 -

6 -

4 -

2 -

0 I I

-2.0 -1.6 -1.2 -0.8 -0.4 0 0.4 0.8 1.2 1.6 2.0
Return (%)

Two-Week Returns


"~ 6

-g 4
_ .': '.. . ... .. /

M.0 -3.2 -2.4 -1.6 -0.8 0 0.8 1.6 2.4 3.2 4.0


---- RiskMetrics 100 D a y s

BIS/Basle ....... 250 Days

tions. In the case of two-week returns, the distribu- time horizons. The top panel of Figure 7 shows tile
tion changes to display bimodal behavior. This return on Portfolio 2 on the day of construction as a
behavior is apparent in the lines that appear to be function of the underlying asset price (the S&P 500).
upside down "normal" distributions. A further Given the starting S&P 500 level of approximately
observation is that the historical simulations pro- 529, the portfolio reflects a small positive return
duce high-probability high-return expectations and (less than 1 percent) if the S&P 500 rises to 544, but
low-probability large-loss expectations relative to its greatest returns occur if the S&P 500 drops
the Monte Carlo simulations. below 510. Small downward moves of the S&P 500
The VAR calculations for Portfolio 2 expose (between 510 and 529), and larger upward moves
several weaknesses of VAR, which can be managed (above 544) produce losses.
with the addition of stress testing and limit policies. The center panel of Figure 7 shows the return
These weaknesses are illustrated by viewing the on Portfolio 2 at the end of the first two-week
differences in the portfolio's return over various holding period, again as a function of the underly-

Financial Analysts Journal/September-October 1995 "19

Figure 7. Portfolio Performance at One Day, at a sole measure of risk because it does not reveal the
Two Weeks, and after June 545 Call true exposure a firm faces. VAR produces a small,
Expires: Portfolio 2 finite number in all eight cases (less than 4 percent).
Another common risk control measure, a prohibi-
tion on portfolios with negative gamma, may also
One Day fail. Portfolio 2 displays slightly positive gamma at
6 - -
construction and as of the time horizons in all
panels of Figure 7.
2 Most users combine VAR with stress testing to
address questions such as, " H o w much do I expect
to lose the other I percent of the time?" Note that in
-2 the case of Portfolio 2, stress tests may fail to reveal
-4 I I I I I I I I I I I the true nature of the firm's risk, producing finite
pictures of profits and losses that depend upon the
level of the S&P 500 assumed. As with VAR, the
Two Weeks quality of the answer depends on the inputs, in-
cluding the financial engineer's ability to select
2 appropriate scenarios.
As experienced during the European currency


, ,q crisis, the Gulf War, and the Mexican peso crisis,
not only are key factors such as "maximum" vola-
tility difficult to predict but also correlation rela-
tionships often change substantially during ex-
treme market moves. The increasing complexity
and optionality of many derivatives makes relevant
After June 545 Call Expires scenario selection even harder. Given these chal-
lenges, many firms design stress tests to analyze the
impact of large historical market moves. In our
0 experience, portfolios do not necessarily produce
~-5 their greatest losses during extreme market moves.
Whether asset based or asset plus liability based,
-15 I I__ I I I I I I I I I portfolios often possess Achilles' heels that require
501 505 509 513 517 521 525 529 533 537 541 545 549 only small moves or changes between instruments
S&P Price
or markets to produce significant losses. Stress
testing extreme market moves will do little to reveal
the greatest risk of loss for such portfolios. Further-
more, a review of a portfolio's expected behavior
over time often reveals that the same stress test that
ing asset price (the S&P 500). At about the S&P indicates a small impact today indicates embedded
value of 541, the magnitude of Portfolio 2's perfor- land mines with a large impact during future
mance changes by a factor of four from the first periods. This trait is particularly true of options-
one-day horizon to the first two-week horizon, as based portfolios that change characteristics because
illustrated by the amplitude of the graphs (1 per- of time rather than because of changes in the
cent versus 4 percent, respectively). The price inter- components of the portfolio. The need for other risk
vals under which Portfolio 2 loses and makes measures--for example, limits that restrict writing
money change as well. uncovered call options--is clear.
Portfolio 2 poses significantly different risks at
different points in time. For example, as shown in Portfolio 3
the bottom panel of Figure 7, prior to the expiration Portfolio 3 consists of the combination of Port-
of the short position in the June call (strike of 545), folio 1 and Portfolio 2. The portfolios are equally
the portfolio presents the possibility of huge loss. weighted, with the net investment in Portfolio 3
After expiration of the June 545 call, however, totaling $1 million.
Portfolio 2 no longer presents this possibility. Again, the VAR analyses, shown in Figure 8,
From management's perspective, VAR fails as reveal a wide range of risk profiles for the portfolio.

20 Financial Analysts Journal / September-October 1995

Figure 8. MAR Calculations: Portfolio 3

Hist. I Pr 100 d ld :~0::i~ 0.73

Hist. I Pr 100 d 2w 1.24

:~ 0.74 ]
Hist. I Pr 250 d ld
e~xO ~:~::~::ii~iiii::i!!: iii:!iii!!i!:i!ii:iii 1.08
© Hist. I Pr 250 d 2w :i::: :i ili !: ! ii :iii:ii:ii::iiiii:ii:iiiiii:iiii:iiiiiiiiii:ii!l
M o n t C I RiskMetrics ld

<> M o n t C I RiskMetrics 2w
] 2.51
M o n t C I BIS/Basle l d ::::~;:~;:~:~::;i;;:i!:;i:J!ii!;!i:10.80
M o n t C I BIS/Basle 2w ii ~i: iji i:ii:~i::~i:i~:i!~i:i!:~:!i~i~i!:i~i:~i~i:i~:ii:~:j~::i:~!ii:j~i:i:~i~3.4
0.5 1.0 1.5 2.0 2.5 3.0 3.5

VAR (%)
[ ] 5% Probability [ ] 1% Probability

As in the case of Portfolios 1 and 2, historical correlated with movements in the Italian lira or tile
simulations present a different view of risk than do Mexican peso? Is the price of Saudi Light correlated
the Monte Carlo simulations, and Portfolio 3's VAR with movements in the price of natural gas? If so,
differences are magnified over the two-week hori- by how much? VAR requires that the user deter-
zon. One-day returns for this multiasset class port- mine correlations not only within markets (for
folio display more consistency under the VAR example, currency underlyings or commodity un-
methodologies than one-day returns for the single- derlyings) but also across markets (for example,
asset class portfolios that compose it. how do changes in the bond market in the United
The sensitivity to correlation assumptions is States affect the Australian equity market?). Given a
demonstrated by the difference in results between portfolio with multiple instruments within and
the Monte Carlo simulations under RiskMetrics and across markets, VAR varies significantly under al-
BIS/Basle factors. Under the RiskMetrics model, ternate correlation assumptions. Pension funds
positive correlation is assumed between the Trea- have addressed the issues of correlation for decades
sury strips in Portfolio 1 and the S&P 500 equity
in studying strategic versus tactical allocation of
index positions in Portfolio 2. Under the BIS/Basle
assets. Correlation issues are also a crucial compo-
method, the correlation is assumed to be 1 between
nent of performance measurement across asset
long positions and -1 between long and short
classes. A single approach to assessing correlation
positions. Not surprisingly, the BIS/Basle VAR
does not exist, and opposite views are common. For
factors are higher than RiskMetrics factors in all
example, what happens when a market breaks
Figure 9 compares the results of the historical through its historical or implied trading pattern
simulation for Portfolio 3 with the results of the and violates the correlation assumption in place'.,'
Monte Carlo simulations for one-day and two-week Recently, many currencies that previously had
returns. As with Portfolio 2, return patterns change displayed little or no correlation with the Mexican
significantly when the holding period is increased peso made sympathy moves during the devalua-
from one day to two weeks, but high-probability tion of the Peso. In some cases, the increased
extreme events no longer occur. volume of barrier options on spreads (also known
Correlation assumptions are an important as- as knock-out or knock-in options) has been blamed
pect of VAR. Firms select quite different answers to for unexpected high correlations during periods
which exposures are allowed to offset each other when market levels approach strike levels, with
and by how much. For example, is the Japanese yen both the writers and the buyers of the barriers

Financial Analysts Journal/September-October 1995 21

Fkjure 9. Distributions of One-Day and Two-Week Returns: Portfolio 3

One-Day Returns
10 --
8 -
6 -
4 - ,/' ." '\
2 - ..'.Td ' ..
0 I

-2.0 -1.6 -1.2 -0.8 -0.4 0 0.4 0.8 1.2 1.6 2.0
Return (%)

Two-Week Returns

5 - / .." ~ ' % , 1.
(... ~",,,:
/, ..'. -,_ ",~ .'.
4 -
• .'Y.i .'". : "< "
:'....~..'. ..': :
3 -
o "-)P - . E"
2 - J/: '.." ::

1 -
0 ~'
-4. -3.2 -2.4 -1.6 -0.8 0 0.8 ] .6 2.4 3.2 4.0
Return (%)

-- -- -- RiskMetrics 100 Days

........ BIS/Basle ....... 250 Days

s u s p e c t e d of t r a d i n g in l a r g e v o l u m e to i n f l u e n c e I n o u r r e v i e w of d i f f e r e n t a p p r o a c h e s to V A R ,
the outcome. s o m e f i r m s a s s u m e d t h a t all c a s h f l o w s w e r e c o r r e -
Correlation assumptions also can mask risks l a t e d a c r o s s all m a r k e t s a n d o t h e r s a s s u m e d a l o w e r
t h a t m a y b e s i g n i f i c a n t for m a n y firms. F o r e x a m - d e g r e e of c o r r e l a t i o n . S o p h i s t i c a t e d m e a n - v a r i a n c e
ple, m a n y p o r t f o l i o s d i s p l a y e m b e d d e d r o l l o v e r m o d e l s - - f o r e x a m p l e , t h e o n e u s e d to c o m p u t e t h e
risk created through timing mismatches. An exam- R i s k M e t r i c s d a t a s e t - - a l l o w c o r r e l a t i o n for all in-
p l e is t h e c o m m o n s t r a t e g y t h a t f u n d s u s e to h e d g e s t r u m e n t s a c r o s s all m a r k e t s t h a t a r e c o v e r e d . A t
long-dated foreign currency positions by rolling the other extreme are models that allow correlation
over short-dated forward foreign exchange con- o n l y w i t h i n a s s e t c l a s s e s (e.g., f i x e d i n c o m e , f o r e i g n
tracts. U n d e r c o m m o n t i m e h o r i z o n s for V A R a n d e x c h a n g e , e q u i t y ) a n d r e q u i r e p e r f e c t p o s i t i v e cor-
its c o r r e l a t i o n a s s u m p t i o n s , a flat c u r r e n c y r i s k r e l a t i o n a c r o s s r i s k - f a c t o r g r o u p s . A n e x a m p l e of
p o s i t i o n o f t e n a p p e a r s . This p a t t e r n c a n m a s k t h e this a p p r o a c h is t h e p r o p o s e d a m e n d m e n t to t h e
l o n g - t e r m r o l l o v e r r i s k i n t r i n s i c in h e d g i n g t h e 1988 Basle C a p i t a l A c c o r d o n m a r k e t risks.
c u r r e n c y r i s k of 10- o r 2 0 - y e a r s e c u r i t i e s w i t h o n e - V A R r e q u i r e s t h e u s e of m a t h e m a t i c a l m o d e l s
to t h r e e - m o n t h c u r r e n c y c o n t r a c t s . to v a l u e i n d i v i d u a l i n s t r u m e n t s , as w e l l as to v a l u e

22 Financial Analysts Journal / September-October 1995

the aggregate portfolio. Variance in the valuations tions. Furthermore, many risk variables such as
produced by widely accepted models (termed political risk, liquidity risk, persbnnel risk, regula-
"mark-to-model" risk) are Well-documented and tory risk, phantom liquidity risk, and others cannot
the subject of many research articles. 4 For example, be captured through quantitative techniques. Yet,
the Black-Scholes versus Hull and White options as demonstrated by recent, well-publicized losses,
models can produce differences of 5 percent or such Variables can cause significant risk. For this
more in pricing, even when all input data are reason, VAR must be supplemented not only with
identical. In addition, the selection of probability stress testing but also with prudent checks and
distribution(s) (an assumption of anticipated or balances, procedures, policies, controls, limits, ran-
experienced market behavior) in one VAR model dom audits, and appropriate reserves.
versus another is a topic of great debate among The BIS, the Group of Thirty, the Derivatives
theoreticians and practitioners.
Product Group, the International Swaps and Deriv-
atives Association, and many national regulators
have declared VAR fundamental to current best
practices in risk management. But models and
N e w studies of the differences in VAR are contem- math, although necessary to manage risk, are not
plated. Such studies will use additional computa-
sufficient to do so. The inability to capture manly
tional techniques, alternate assumptions, and a
qualitative factors and exogenous risk variables
broader range of portfolios in terms Of number of
points to the need to combine VAR with stress tests,
positions, type of positions, and number of asset
Checks and balances, procedures, policies, controls,
classes. As highlighted by the three portfolios, the
picture of expected capital at risk is wildly depen- limits, and reserves. Perhaps the limitations of
dent upon the VAR methodology and the assump- quantitative techniques explain the recent art-
tions behind the specific calculation. Not only do nouncement by Moody's that although 25 percent
the eight VAR results for the individual portfolios of its volatility rating for funds will be based on
differ significantly, but the magnitude of the differ- VAR, the remaining 75 percent will be based on
ence does not follow a clear pattern with increasing qualitative factors. In sum, mathematics is integral
complexity of the portfolio. Thus, dealers and end- to finance, but finance does not always follow
usei~s are in a precarious position. The dependence mathematics.
on technology and skilled professionals is gr6ater Some regulators propose to allow firms to use
thari ever before. Although this dependence has their own internal VAR models plus assumptions,
produced invaluable advances in financial engi- but others do not. Although the Basle amendment
neering a n d risk management, some firms have allows banks to select their own internal models to
been lulled into a false sense of security. Often, calculate VAR (subject to the proviso that certain
firms forget the degree to which the output of assumptions are required for VAR's key factors and
models depends upon the modeler's perspective to oversight by national regulators), the National
and assumptions. A firm's senior management and. Association of Insurance Commissioners' proposed
directors or trustees are shocked to learn that their risk-based capital standard-(RBC) requires insurers
firm's risk reports can change dramatically under to use a single, rigid approach for their VAR-type
alternate assumptions. This fact is even more sur- calculations. 5 The use of a single, rigid approach
prising when the "alternate" assumptions are those
may have a deleterious effect. Namely, because of a
they consider to be likely or reasonable.
rigid correlation assumption, RBC may penalize the
Some firms make the mistake of equating VAR
insurer for a successful asset allocation strategy.
under a 99 percent expectation to the certainty or
From a regulatory standpoint, the choice of stan-
confidence that the firm will not lose more than the
stated amount more than I percent of the time (i.e., dardized versus internally selected VAR ap-
fewer than three business days a year). As demon- proaches presents difficult trade-offs. The use of a
strated by the sample portfolios, the 99 percent single, rigid approach may stymie the development
VAR changes significantly based on the time hori- of new and improved risk measurement, creating
zon, data base, correlation assumptions, mathemat- expense for regulators, shareholders, employees,
ical models, and quantitative techniques that are and taxpayers alike. A proper understanding of the
used. Accordingly, VAR does not provide certainty assumptions behind VAR, as well as its limitations
or confidence of outcomes, but rather an expecta- and pitfalls, will help all to benefit from this pow-
tion of outcomes based on a specific set of assump- erful technique. 6

Financial Analysts Journal/September-October 1995 23


1. Note that selection of key statistical parameters such as the 4. See Tanya Styblo Beder, "The Realities of Marking to Model,"
mean and variance can significantly affect distributions and, Bank Accounting & Finance, vol. 7, no. 4 (Summer 1994):4-12.
therefore, the results of simulations. 5. Note that a single RBC approach exists for life insurance
2. Duration is defined here as the change in price with respect to companies, and a single alternate approach exists for proper-
yield. ty/casualty companies.
3. The Monte Carlo simulations are based on the assumptions 6. The author wishes to thank Frank Iacono, Maarten Nederlof,
that the Treasury strip yields are lognormally distri- Tom Riesing, Anil Suri, and Charles Taylor for their invalu-
buted and that the average change, or "drift," in each yield able assistance and input during the preparation of this
is zero. article.

24 Financial Analysts Journal/September-October 1995