Jamshid I An
Jamshid I An
Farshid Jamshidian
1
, Yu Zhu
2
1
New Products and Strategic Trading, Sakura Global Capital, 42 New Broad Street,
London EC2M 1JX, United Kingdom
2
Risk Assessment and Control, Sakura Global Capital, 65 East 55 Street, New York,
NY 10022, USA
Abstract. This paper presents a new simulation methodology for quantitative
risk analysis of large multi-currency portfolios. The model discretizes the mul-
tivariate distribution of market variables into a limited number of scenarios.
This results in a high degree of computational eciency when there are many
sources of risk and numerical accuracy dictates a large Monte Carlo sample.
Both market and credit risk are incorporated. The model has broad applications
in nancial risk management, including value at risk. Numerical examples are
provided to illustrate some of its practical applications.
Key words: Risk analysis, Monte Carlo studies, approximations to distributions
JEL classication: C15 D81 D84
Mathematics Subject Classication (1991): 62E17 62E25
1. Introduction
In recent years, value-at-risk (VaR) has been considered as one of the best
market risk measures by banks and other nancial institutions. VaR is dened
as the expected loss from an adverse market movement with a specied prob-
ability over a period of time. For each nancial institution, a certain amount
of capital must be set aside so that the probability that the institution will not
The Scenario Simulation model has been successfully developed and applied to monitor
risk exposures and calculate VaR and credit reserves at Sakura Global Capital and its AAA
derivative vehicle, Sakura Prime. We would like to thank Reza Bahar and Mark Gold from
S&P, Jeremy Gluck from Moodys, Joseph Snaller from CapMac, and the referee for their
valuable comments on the model. We are grateful to many of our colleagues at SGC, espe-
cially Mahadeo Nadkarni and Siyun Wang, for their contributions and assistance.
Manuscript received: February 1996; nal revision received: June 1996
44 F. Jamshidian, Yu Zhu
survive adverse market conditions remains very small. In December 1995, the
Bank for International Settlement (BIS) nalized its rules for allocating the
capital to market risk, allowing banks to use their internal model to measure
the market risk exposure, expressed in terms of value-at-risk.
1
There are several commonly applied methods to calculate VaR. One sim-
plest method is delta approximation. It uses variance-covariance matrix of
market variables and the portfolios sensitivities with each of the market vari-
ables (delta) to approximate the potential loss of the portfolio value. This
method critically depends on two dubious assumptions: the normality assump-
tion of portfolio value, and the linearity assumption of the relationship between
transactions prices and market variables.
In general, however, nancial transactions, derivatives in particular, have
non-linear price characteristics. Thus, Monte Carlo simulation is a more appro-
priate method to estimate their market exposures. A common implementation
is based on the joint lognormality assumption of market variables to generate
a large number of market scenarios using their historical variance-covariance
matrix. Then, transactions values for each scenario are calculated and aggre-
gated. From the obtained distribution of the portfolio value, VaR can be easily
estimated. The diculty with the Monte Carlo approach is its computational
burden. In order to obtain a reliable estimation, the sample size has to be large.
In the case of large multi-currency portfolios, the required huge sample size
often makes the approach impractical. Choosing a smaller sample size would
result in a distorted distribution, defeating the purpose of adopting the Monte
Carlo approach.
The Scenario Simulation model described in this paper is a new approach
to estimate VaR. The model approximates a multidimensional lognormal distri-
bution of interest rates and exchange rates by a multinomial distribution of key
factors. While it allows very large samples, the number of portfolio evalua-
tions is limited. As a result, a great computational eciency has been obtained
in comparison with conventional Monte Carlo methods. In addition to the ef-
ciency, it is our view that the Scenario Simulation model ts the needs of
risk management better than conventional Monte Carlo method. In the next
two sections, we describe the basic assumptions and mathematical models of
the scenario simulation model under the single currency yield curve setting. In
Sect. 4, the model is extended into a multi-currency setting. We also describe
how to incorporate other market risk variables such as volatility and basis
risk.
The concept of value-at-risk is not restricted to the market risk. For the
risk management of a nancial institution, it is important to examine not only
the market risk of the portfolio, but also its overall risk exposures. Section 5
discusses the models application in estimating a portfolios overall risk prole
of joint market risk, counterparty default and country risk events. Section 6 ap-
plies the model to analyze the risk prole of a multi-currency swap portfolio.
Section 7 concludes the paper.
Scenario Simulation 45
2. Yield curve modelling
We make the following assumptions:
a1. A yield curve is described by a vector of key zero coupon rates:
r
1
; : : : ; r
i
; : : : ; r
n
(1)
a2. Each key rate is lognormally distributed:
dr
i
r
i
= \
i
(t) dt + A
i
dz
i
(2)
where z
i
(t) N(0;
1
2
A
2
i
t
(4b)
so that the expected value of r
i
(t) equals the forward rate.
2
a3. z
i
(t); i = 1; : : : ; n, are correlated in the following way:
dz
i
; dz
j
) = @
ij
dt (5)
i; j =1; : : : ; n. Thus we shall have a nn correlation matrix R=@
ij
; i; j =1; : : : ; n.
Based on assumptions (a1) (a3), the distribution of yield curves at a
future time t is determined. To simulate the distribution, one method is to apply
the Monte Carlo method, which can be briey described as follows. First, a set
of random numbers z
1
; : : : ; z
n
is selected according to their joint distribution.
By applying Eq. (3), we can generate one future yield curve r
1
(t); : : : ; r
n
(t).
In order to simulate the distribution, a large number of future yield curve have
to be generated.
Suppose that 12 key rates are used to describe the yield curve. In other
words, the yield curve model has 12 factors. For each random variable z
i
to
have three possible values, the total number of future curves we need would
Table 1. Cumulative percentage of variations explained by principal components
By correlation matrix By covariance matrix
No. of factors 1 2 3 4 5 1 2 3 4 5
DEM 67.7 83.9 93.7 97.1 98.5 71.5 84.8 92.9 96.7 98.4
JPY 75.8 91.2 94.3 96.6 98.1 69.7 90.7 95.6 97.7 99.2
USD 75.8 85.1 93.0 97.0 98.9 84.4 90.9 94.4 97.3 98.9
46 F. Jamshidian, Yu Zhu
be 3
12
= 531; 441, about half a million. If each random variable takes ve
outcomes, the total number would be increased to above 200 million! Of course,
Monte Carlo simulation does not require all the permutations. A smaller number
of trials would be sucient in order to obtain reasonable value of moments
and expectations such as mean and variance. Even then, the required number of
trials will still be very large, especially in the case of multi-currency portfolio
where the correlations among currencies and currency yield curves are usually
low. Obviously, even with a large sample, the coverage of all extreme cases
is not guaranteed in the Monte Carlo method.
Fortunately, to simulate the future yield curve movements, we do not have
to use the brute-force Monte Carlo method. It has been documented that
yield curve movements can be largely explained by two or three factors.
3
These
studies show that the three factors can be interpreted as shift, twist, and buttery
(bend) of yield curves. The principal component analysis is one of the common
methods to extract these key factors from yield curve movement data. Table
1 is an example of principal component analysis which shows the cumulative
percentage of variations explained by the rst ve principal components of
yield curves of Germany, Japan and United States.
4
It can be seen that the rst three principal components can explain about
9395% of the variations. The principal component analysis can be preformed
to either the covariance matrix or the correlation matrix. When the volatilities
of all key interest rates are of the same order of magnitude, the results from the
two methods are very similar. For simplicity, we shall use the second method
in our discussion.
Suppose R = @
ij
i; j=1;:::;n
is the correlation matrix of the n key rates
r
1
; : : : ; r
n
. Let the j-th eigenvector of R be dened as
R
j
= (R
ij
; : : : ; R
nj
)
T
(6)
From the denition of the eigenvector, we have
RR
j
= [
j
R
j
j = 1; : : : ; n (7)
where [
j
is the j-th eigenvalue. Since R is a symmetric non-negative denite
matrix [
j
=0, j = 1; : : : ; n; and the eigenvectors are orthogonal to each other.
We normalize all R
j
such that
[R
j
[
2
=
n
i=1
R
2
ij
= [
j
Assume that
[
1
=[
2
= =[
n
; (8)
then R
1
is called the rst principal component, R
2
is the second principal com-
ponent, and R
j
is the j-th principal component. As we discussed above, the
empirical analysis of the historical yield curve data demonstrates that all but
the rst three principal components are small in magnitude. Now, dene the
Scenario Simulation 47
principal factors dw
j
by
dw
j
=
1
[
j
n
k=1
R
kj
dz
k
j = 1; : : : ; n (9a)
Then, it is easy to see that
dz
i
=
n
j=1
R
ij
dw
j
i = 1; : : : ; n (9b)
and dw
k
; dw
j
) = 0 (k ,= j); and dw
k
dw
k
= dt:
5
If we assume that
[
4
; : : : ; [
n
[
1
; [
2
; [
3
then Eq. (9b) can be approximated by
dz
i
R
i1
dw
1
+ R
i2
dw
2
+ R
i3
dw
3
The residual should be very small. Based on this approximation, Eq. (2) can
be rewritten as
dr
i
(t)
r
i
(t)
= \
i
(t)dt + T
i1
dw
1
+ T
i2
dw
2
+ T
i3
dw
3
(10)
where
T
ij
= A
i
R
ij
; i = 1; : : : ; n; j = 1; 2; 3:
Similarly, in the integrated form, Eq. (3) becomes
r
i
(t) = u
i
(t)e
Ti1w1(t)+Ti2w2(t)+Ti3w3(t)
(11)
The dimensionality of the problem is thus reduced from n to 3. Note also,
w
i
; i = 1; 2; 3, are now independent random variables.
3. Scenario simulation model
Equation (11) is a three-factor yield curve model. To simulate the future yield
curve movements, we can apply the Monte Carlo method. Say we select 100
possible outcome for each random variable w
i
; i = 1; 2; 3. The total number of
possible yield curve at one future point in time will be 10
6
. This generates an
adequate distribution of the future yield curves, but it requires a large number of
portfolio evaluations. The question is how to reduce the computational burden
further, while achieving the objective of obtaining a satisfactory representation
of the distribution of a portfolios risk exposure.
Suppose W is a random variable with distribution P(w). In the Monte Carlo
method, hundreds of possible w are selected to simulate the random variable
W, and each w has the same probability. The number of ws which fall in range
between w
A
and w
B
is proportional to the probability P(w
A
w 5w
B
). For the
risk simulation, we can select a state w
k
which represents the region (w
A
; w
B
],
and assign the probability of the region to that state (scenario). Thus a limited
48 F. Jamshidian, Yu Zhu
number of states can be selected. With appropriately assigned probabilities,
these states (scenarios) provide a good representation of the distribution.
It is well known that the multinomial distribution constitutes a good ap-
proximation to a multinormal distribution.
6
If m + 1 states (ordered from 0 to
m) are selected, the probability of state i of a binomial distribution can be
expressed as
Probability(i) = 2
m
m!
i ! (m i)!
i = 0; : : : ; m (12)
The distance between two adjacent states is 2=
t). Rather,
its distribution is
N
_
0;
_
1 e
2kt
2k
_
(14)
It can be seen that at a given time t; z
i
is normally distributed, equivalently,
r
i
(t) is lognormally distributed. Therefore, all previous analyses of the Scenario
Simulation model hold without modication. In particular, the principal factors
w
i
(t) dened in Eq. (9a) are also OresteinUhlenbeck processes with the same
mean reversion parameter k. Over time, the mean reversion factor k reduces
the variance of w
i
(t). But, since w
i
(t) is still normal, it can be discretized by
the same method as before.
Table 3 uses the aforementioned example of Japanese interest rates to
demonstrate the eect of the mean reversion factor k. It seems that a mean
reversion factor of 0:30 0:50 is warranted for this very volatile case. If k is
0.50, the standard deviation of 3-year interest rate would be reduced by about
half for a 5-year horizon, and by 2/3 for a 10-year horizon. This results in much
more reasonable simulated interest rates, even though they are still quite high.
An interesting issue is how to select the mean reversion factor k for each
currency yield curve. Empirical estimation based on historical data is one pos-
sible method. Another more practical approach is to select k such that the
52 F. Jamshidian, Yu Zhu
mean reversion will be kicked in only if volatilities are much higher than
the historical norms for that currency yield curve.
4. Joint probability model of Scenario Simulation
For a single currency, the principal factors of the yield curve are independent,
and their joint probability density is simply the product of the individual den-
sities. This furnishes the mean, standard deviation, histogram, and percentiles
of future portfolio value for each currency. However, in order to aggregate
the portfolios and obtain similar statistics for the U.S. dollar value of the to-
tal portfolio, we need a discrete joint probability model for the totality of the
principal factors of all countries as well as all the exchange rates.
One approach would be to perform the principal component factorization
of the international economy, i.e., of the combined correlation matrix of all
key rates and exchange rates, and choose a suciently high number of factors
(say 10) that explain most of the total variation. Then, the principal factors are
independent and the joint probability is again a simple product. However, these
principal factors are dicult to interpret intuitively. Moreover, this approach
breaks the symmetry and stratication of the single currency models, which
carry important information for risk management, and dramatically increases
computational time.
We seek an approach that is consistent with the single currency model in the
sense that the marginal distribution for each currencys yield curve has the same
scenarios (with the same probabilities) as those constructed for that currency
separately. To this end, we dene a discretization of the multivariate normal
distribution such that each coordinate is discretized into a binomial distribution
exactly as in the one-dimensional case. To our knowledge, this construction
has not appeared in the literature before. We will show that the multivariate
discretization converges with probability one to the original multivariate normal
variate. Further, we will describe a procedure for Monte Carlo simulation of
the discrete distribution.
To proceed, for each integer m 0, we dene an increasing sequence
numbers of a
i
by a
0
= , a
m+1
= , and the relations,
1
2
_
ai+1
ai
exp
_
x
2
2
_
dx =
2
m
m!
i!(m i)!
; i = 0; : : : ; m: (15)
Thus, a
i
are the points on the real line such that the area under the normal
curve on the segment [a
i
; a
i+1
] equals the binomial probability of state i. For
example, for 3, 5, and 7 states, one computes:
m = 2 : a
1
= 0:67449; a
2
= 0:67449:
m = 4 : a
1
= 1:53412; a
2
= 0:48878; a
3
= 0:48878; a
4
= 1:53412:
m = 6 : a
1
= 2:15387; a
2
= 1:22986; a
3
= 0:40225; a
4
= 0:40225;
a
5
= 1:22986; a
6
= 2:15387:
Scenario Simulation 53
Now dene the function B
(m)
on the real line with values in the set 0; : : : ; m
as follows:
B
(m)
(z) = i if a
i
5z a
i+1
:
By construction the function B
(m)
is so designed that if Z is a normally dis-
tributed random variable with mean zero and variance 1, then B
(m)
(Z) is bino-
mially distributed with m + 1 state.
If X is a k-dimensional normal variate with correlation matrix Q,
X = (X
1
; : : : ; X
k
) N(0; Q);
each X
i
having mean zero and variance 1, we dene its discretization to be the
multivariate discrete variate
B = B
(m)
= (B
(m)
(X
1
); : : : ; B
(m)
(X
k
)): (16)
Note, each coordinate of B is binomially distributed and equals the discretiza-
tion of the corresponding component of X. It is in this sense that we say this
discretization preserves the stratication.
From this denition, we see that it is easy to generate random samples
of the multidimensional discrete variate B. We simply generate a multivariate
N(0; Q) normal deviate x = (x
1
; : : : ; x
k
), then simply nd between which a
i
and a
i+1
each x
j
lies, and thus form (B
(m)
(x
1
); : : : ; B
(m)
(x
k
)).
The joint probability density of B is clearly given by
prob[B = (i
1
; : : : ; i
k
)] =
_
ai
1
+1
ai
1
_
ai
k
+1
ai
k
p(x
1
; : : : ; x
k
)dx
1
dx
k
; (17)
where
p(x) = p(x
1
; : : : ; x
k
) = (2)
k=2
det(Q)
1=2
exp
_
1
2
x
t
Q
1
x
_
is the multivariate normal density function of X.
We now show that the standardized B
(m)
converges to X with probability
one. In particular, it follows that the correlation matrix of B
(m)
approaches Q
for large m. But even for small values m, it is close enough to Q for practical
purposes.
Dene R
(m)
(z) = (2B
(m)
(z) m)=
i=1
k
j=1
X
i
X
j
V
i
V
j
;
E[V
2
] =
k
i=1
k
j=1
E[X
i
X
j
V
i
V
j
]:
The above equation indicates that joint probabilities of X
i
; X
j
; V
i
; V
j
are re-
quired to calculate E[V
2
]. Generally, this is a non-trivial computation. However,
if it is assumed that exchange rates and interest are uncorrelated and, yields
of dierent countries are correlated only through, say, their rst principal fac-
tors, then the above expression factors into expectations of several bivariate
distributions, which can be easily computed using Eq. (17).
In some applications, it is plausible to assume that, for a diversied port-
folio, V is normally distributed, so that the risk exposure can be determined
just by the expected value E[V] and the variance Var [V]. In general, however,
the normality assumption of portfolio value V is not plausible. For example,
the portfolio may contain a disproportionate number of options, or, skewness
may result from truncations brought by considerations of counterparty default.
In this case, we need the entire distribution which requires all possible states.
If the number of currencies k equals 2 or 3, this may well be computationally
manageable. But, for k = 6 for instance, the number of scenarios (105735
k1
)
is more than 10
16
, i.e., 10,000 trillion. We therefore use Eq. (16) to Monte
Carlo simulate the discrete distribution.
In this regard, it cannot be over-emphasized that our approach to simulation
fundamentally diers from the usual Monte Carlo approach applied in risk
management. The usual approach does not employ stratied discretization as
in our theory. If N is the number of Monte Carlo trials, it generates N yield
curves and exchange rates for each currency. This means that each transaction
must be evaluated in N states. For the computation to be feasible with a large
portfolio, then of necessity N must be rather small. On the other hand, the high
dimensionality of the problem dictates a large N for an adequately accurate
sample.
By contrast, in our approach each transaction (like an interest-rate swap or a
swaption) is only calculated in the 105 scenarios of the corresponding currency
yield curve. Transactions in each currency i are then simply aggregated for each
of the 105 scenarios, yielding V
i
. At this stage, N joint scenarios are randomly
generated using Eq. (16). Then, all that needs be done is the addition in Eq.
(19) for each of the N samples.
In short, our approach requires only 105 evaluation of each transaction
regardless of the number trials N. For this reason, even for very large portfolios,
56 F. Jamshidian, Yu Zhu
Table 4. Correlation matrices of yield curve principal
factors (Based on 9/15/1995 JPM RiskMetrics monthly
data)
CHF DEM GBP JPY USD
First principal factor
CHF 1.000 0.809 0.666 0.101 0.322
DEM 0.809 1.000 0.721 0.133 0.379
GBP 0.666 0.721 1.000 0.020 0.412
JPY 0.101 0.133 0.020 1.000 0.017
USD 0.322 0.379 0.412 0.017 1.000
Second principal factor
CHF 1.000 0.286 0.121 0.074 0.067
DEM 0.286 1.000 0.273 0.228 0.190
GBP 0.121 0.273 1.000 0.062 0.050
JPY 0.074 0.228 0.062 1.000 0.023
USD 0.067 0.190 0.050 0.023 1.000
Third principal factor
CHF 1.000 0.130 0.081 0.048 0.220
DEM 0.130 1.000 0.136 0.033 0.003
GBP 0.081 0.136 1.000 0.123 0.228
JPY 0.048 0.033 0.123 1.000 0.144
USD 0.220 0.003 0.228 0.144 1.000
we can comfortably take N to be as large as, say, a million, and thereby
achieve stable and accurate results.
We should also mention that the scenario simulation methodology does take
account of rare events and tails of the distribution. For example, although the
single currency yield curve has only 105 scenarios, some of them occur with
a probability of 0.024%, well below the 1% threshold commonly considered
as the probability of a rare event. These low probabilities compound for joint
scenarios.
4.2 Further details of model implementation
Although by no means required, certain reasonable assumptions can be made
to further exploit the symmetries of the model and increase its computational
eciency. Namely, we assume:
a4. Exchange rates and interest rates are uncorrelated. In fact, the historical
correlations of interest rates and exchange rates are generally low, and empirical
evidence suggests that their signs are often unstable and depend on the time
period and the length of the sample period.
a5. The rst principal factor of one countrys yield curve can be correlated
with the rst principal factor of another country, and the second principal
factors can be correlated, and so on, but the rst principal factors are not
Scenario Simulation 57
Table 5. Correlation matrix and volatilities of ex-
change rates (Based on 9/15/1995 JPM RiskMetrics
monthly data)
CHF DEM GBP JPY Volatility%
CHF 1.000 0.978 0.768 0.771 6.995
DEM 0.978 1.000 0.783 0.777 6.215
GBP 0.768 0.783 1.000 0.552 3.884
JPY 0.771 0.777 0.552 1.000 7.830
correlated with the second or third principal factors, etc. We argue that it
seems unlikely that, say, parallel shift and yield curve twist within the same
country be independent, yet across dierent countries be dependent. The data
for the ve major currency yield curves in Table 4 show that the correlation
between interest rate levels (rst principal factor) accounts for most of correla-
tion between yield curve of dierent countries.
11
Coupled with the fact that the
rst principal components have the largest magnitude and contribute greatest to
the variance, it seems that assumption (a5) is very innocuous. Table 5 shows
the correlation matrix and volatilities of the four exchange rates.
By enriching the structure of the model, these assumptions actually increase
the complexity of programming. But the result is well worth it in terms of
gained eciency. The primary benet is that the problem basically reduces from
one of dimensionality 4k, as it were, to 4 problems, each with dimensionality
k. Thus, we deal with the k k correlation matrix of the rst factors, and those
of the second and third factors, and the (k 1) (k 1) correlation matrix
of exchange rates. This reduces by a factor of 4 certain calculations needed
to form a 4k-dimensional random vector. It also enables use of antithetical
symmetry for each of the four groups, eectively reducing the total number of
random vector generations by a factor of 16.
The Scenario Simulation model can also be extended to cover other market
risks, such as volatility or basis spread. Consider for example the risk associ-
ated to caps, oors and swaptions as a result of a change in the overall level of
implied volatilities employed for the valuation of these options.
12
Let us denote
dv
i
as the change in the level of implied volatility for currency i; i = 1; : : : ; k:
For currency i, we can introduce dv
i
as the fourth independent factor, assuming
for example it takes three values, 0; +1% and 1% with probabilities 1/2, 1/4
and 1/4 respectively. Then, there will be a total of 1053 =405 joint yield-
curve/implied-volatility scenarios for each currency, with given probabilities.
Thus, securities that depend on volatility have to be evaluated 405 times, while
105 evaluations still suces for swaps or cross-currency swaps. The joint inter-
national distribution of yield-curve/implied-volatility is discretized and Monte-
Carlo sampled as before, one million times. Moreover, we can incorporate given
correlations between implied volatility factors dv
i
of dierent currencies i, in
the same manner as we have allowed for correlation between exchange rates,
58 F. Jamshidian, Yu Zhu
or between the rst principal factors, etc. In short, implied volatility simply
enters as another factor for each currency.
For the Monte Carlo sampling of the scenarios, we use LEcuyer algorithm
for uniform (0,1) deviates, supplemented by shuing. By combining two
linear congruential generators, it results in a very long period, as desired for
this problem, and induces practically no serial correlation. From two or more
uniform deviates, we generate a single Gaussian deviate, using the Box-Muller
method combined with acceptance/ rejection within the unit disk. To generate a
k-dimensional N(0; Q) deviate, we rst generate k i.i.d Gaussian deviates, put
them in a k-vector Y, and pre-multiply Y by some matrix A satisfying AA
t
=
Q:
13
For the choice of A, we can take the principal component factorization of
Q, the symmetric square root of Q, or the Cholesky factor of Q. We choose
the latter, because A is then a triangular matrix and multiplying it by a vector
can be made more ecient. Finally, we invoke the discretization function B
(m)
;
i.e., form B
(m)
(AY). The entire sequence is generated by a single seed.
14
It is instructive to point out what happens if, instead of rst multiplying
Y by the Cholesky factor A and then descretizing, we rst discretize and then
multiply by A. In the latter case, the resulting deviate A(B
(m)
(Y)) will have cor-
relation matrix exactly Q. However, the stratication is destroyed: whereas each
component of B
(m)
(AY) has only m+ 1 states, each component of A(B
(m)
(Y))
has in general m
k
states. So, we revert to the case of the usual Monte Carlo
where an extortionate number of portfolio valuations is required. Another tech-
nique known as rank correlation also preserves the stratication. But, we regard
our approach to be more appropriate and theoretically complete.
4.3 Example of simulation on a ve currency portfolio
As an example, we construct a portfolio with ve interest rate swaps in ve
dierent currencies. For simplicity, the tensor is 10 years, and the pay and
receive frequencies are semi-annual for all swaps in the portfolio. The total
portfolio value as of the valuation date (9/15/1995) is $315; 637: The details
of these swaps are listed below:
Currency Notional Receive Pay Market value
CHF sfr 90 mm 4.8% Libor sfr 502; 430
DEM DM 70 mm Libor 6.90% DM 719; 400
GBP $ 50 mm Libor 8.42% $ 386; 582
JPY Y = 5 bil Libor 3.10% Y = 23; 863; 598
USD $ 100 mm 6.5% Libor $ 256; 980
We ran the Scenario Simulation model for this ve currency portfolio to
estimate the portfolio value distribution for a 30-day horizon. The historical
volatility and correlation information of the yield curve and exchange rates are
based on 9/15/1995 JPM RiskMetrics monthly data.
The simulation takes two steps. First, we simulate each transaction in its
own currency within the framework of single currency Scenario Simulation
Scenario Simulation 59
Table 6. 30-day portfolio horizon value
Portfolio statistics and percentiles ($1,000)
No. of Mean STD 1% 2% 98% 99% Maxi- Minimum
trials mum
1,000,000 319 5,013 11; 918 10; 630 9; 809 11,055 20,784 21; 366
2,000,000 319 5,028 11; 971 10; 665 9; 838 11,096 22,148 23; 156
3,000,000 319 5,027 11; 974 10; 665 9; 836 11,096 22,185 23; 156
4,000,000 319 5,028 11; 972 10; 668 9; 838 11,092 22,185 23; 156
Fig. 1. Five currency portfolio horizon value (Histogram of 1,000,000 trials)
model described in Sect. 3. As before, we apply three principal factors and
7 5 3 = 105 scenarios for each currency yield curve. For this portfolio, the
single currency simulation generates 105 prices at the horizon date for each
swap. That is, the swap valuation model would be called 105 5 = 525 times.
In the second step, we perform Monte Carlo to simulate the distribution of the
portfolios horizon value. The simulation results are shown in Table 6. Clearly,
the sample statistics become quite stable with one million trials. Comparing the
histogram of the simulated potential horizon values to the normal distribution
with the same mean and standard deviation, it can be seen that the sample
distribution is well behaved and has a fatter tail (Fig. 1).
Even though the number of trials in our example is signicantly larger
than that in the usual Monte Carlo simulation applied in risk management, our
approach does not take too much computation time due to the eciency of the
methodology as well as its implementations. In fact, running on a Sun SPARC
Station 20, the second step takes 54; 96; 140 and 180 seconds to complete one,
two, three and four million trials, respectively.
60 F. Jamshidian, Yu Zhu
5. Estimating credit exposures with Scenario Simulation model
The credit exposure reects the potential loss one transacting party may incur if
its counterparty defaults. The credit risk is related to but dierent from the mar-
ket risk. When a counterparty defaults, all transactions with that counterparty
will be terminated, and the transactions value by the end of the termination
period is called the termination value (or replacement value).
For a given market scenario s, let V(s) be the net potential termination value
of a counterpartys portfolio at the end of one period. If p(s) is the probability
of the market scenario, then the distribution of the termination value can be
described by
(V(s); p(s); s = 1; : : : ; n)
where n is the total number of market scenarios. If the counterparty defaults, the
transacting party will suer a loss only if the transactions with the counterparty
have net gains. Otherwise, the credit loss would be zero. Thus, the distribution
of the potential credit loss amount with respect to the counterparty is
(max[V(s); 0]; p(s); s = 1; : : : ; n)
In order to apply the Scenario Simulation model to estimate the credit exposure,
additional assumptions are required. We assume:
a6. The default incident of a counterparty is independent of the market vari-
ables by which a market scenario is dened, and it is also independent of
other counterparties. This is a strong assumption as adverse market conditions
and nancial situations of related counterparties inevitably aect counterparties
economic well being. However, it is a convenient assumption because not only
it simplies the analysis considerably, but there is hardly any data available to
deduct their correlations.
a7. The default probability of a counterparty is solely determined by its credit
ratings.
a8. The loss ratio (0 L 51) of a defaulting party is assumed to be one.
Though this assumption is for simplicity and by no means necessary, it results
in a more conservative estimation of credit risk as empirical evidence shows
that, depending on the seniority of a transaction, the actual loss amount can be
a fraction of total default amount.
15
Assume q is a counterpartys marginal default probability for one period.
Several one period credit risk measures can be readily calculated using the
Scenario Simulation model. For example, the expected credit loss for the period
can be dened as
E[D] =
n
s=1
max[V(s); 0] p(s) q (20)
Scenario Simulation 61
One can also dene the maximum average credit loss with probability as
follows:
D
= (i
1
; i
2
; : : : ; i
k
; d
1
; d
2
; : : : ; d
J
)
For counterparty j, if netting is enforceable, the scenario value with respect
to the counterparty can be obtained by aggregating all transactions with the
counterparty. Assume there are g = 1; : : : ; G transactions with counterparty j
and let V
j
(s) be the scenario value without default. Then,
V
j
(s) =
G
g=1
V
j; g
(s)
where V
j; g
is the market value of transaction g with counterparty j. The joint
scenario value becomes
V
j
(s
) = min[V
j
(s); V
j
(s)(1 d
j
)] (22)
Equation (22) means that if d
j
= 0, i.e., no default, then the joint scenario value
is equal to the market value at scenario s : V
j
(s
) = V
j
(s). If the counterparty
j defaults, d
j
= 1, then V
j
(s
) = min[V
j
(s); 0].
If netting is not enforceable, then
V
j
(s
) =
G
g=1
min[V
j; g
(s); V
j; g
(s)(1 d
j
)] (23)
62 F. Jamshidian, Yu Zhu
It is easy to see that if the counterparty does not default, the joint scenario
value V
j
(s
) =
g
min[V
j; g
(s); 0].
The portfolio value V
p
at the joint scenario s
is
V
p
(s
) =
J
j=1
V
j
(s
) (24)
Using the Monte Carlo method with a large number of joint scenarios, we
can simulate the distribution of the termination values of each counterparty as
well as the whole portfolio, taking into account both market risk and credit
risk.
We can also incorporate the country risk event (e.g., a country defaults or it
imposes foreign exchange control) into the Scenario Simulation model. When
such a risk event happens in a country, we assume that all counterparties in
that country default simultaneously and the transacting partys exposure is the
netted amount of these counterparties.
17
Let c
1
; : : : ; c
H
, be H independent Bernoulli random variables, represent-
ing country risk events for H countries we are dealing with. Assume that c
h
equals one (zero) when the country does (not) default. The joint scenario of
market variables, counterparty default event and country risk event now be-
comes
s
= (i
1
; i
2
; : : : ; i
k
; d
1
; d
2
; : : : ; d
J
; c
1
; : : : ; c
H
)
The portfolio value with respect to all counterparties in country h under sce-
nario s
is
V
h
(s
) = min
_
j
V
j; h
(s
);
j
V
j; h
(s
) (1 c
h
)
_
(25)
where V
j; h
(s
is
V
p
(s
) =
H
h=1
V
h
(s
) (26)
6. Joint exposure distribution of a swap portfolio
We construct a swap sample portfolio to illustrate the methodology. The port-
folio consists of 305 interest rate swaps and cross currency swaps with 71
counterparties. The credit ratings of all counterparties are investment grade.
Scenario Simulation 63
These transactions are in ve dierent currencies and the total notional amount
is $11.8 billion, among which 16% is cross currency swaps. The average re-
maining tenor is 2.5 years. The basis statistics of the portfolio are listed below:
Ratings Number of counterparties Currency Proportion (%)
AAA 6 ( 8.4%) CHF 11.2
AA 19 (26.8%) DEM 5.4
A 40 (56.4%) GBP 4.1
BBB 6 ( 8.4%) JPY 43.3
USD 36.0
The Scenario Simulation model uses JPM RiskMetrics monthly data as
of 10/20/1995. The simulation is to calculate the potential exposure 30 days
from 11/10/1995, the valuation date. The total number of trials is one mil-
lion.
For illustration purposes, we apply the following extremely stressed one-
month default probabilities: for AAA, 1.3524%; Aa/AA, 1.7294%; A/A,
2.6866%; and Baa/BBB, 5.3596%.
18
The simulation results are shown in Table
7 and Figs. 24.
It is interesting to observe that while the market exposure of the swap port-
folio (Fig. 2) looks like bell-shaped or approximately normally distributed,
the joint distribution of market risk and credit risk shown in Fig. 3 is noticeably
skewed because of the possibilities of counterparty defaults. Figure 4 shows
that there is a possibility of very large losses due to counterparty default. But
for diversied portfolios and counterparties, the probability of these losses is
very small.
In this example, if the market risk is the only concern, the portfolios
30-day VaR would be $34 mm at 99% condence level. On the other hand,
the potential loss due to counterparty default is $55 mm at the same con-
dence level. Such a large potential credit loss is a result of using exagger-
ated default probabilities. However, the probability that in 30-days, the true
potential loss due to adverse market movements and/or counterparty default
exceeds $66 mm is less than 1%. In other words, VaR calculated by the
joint distribution is not the simple summation of the potential market loss
and the default loss, and it is perhaps a better measure of portfolios overall
risk.
Table 7. Portfolio 30-day exposure distribution ($1,000) using stressed default probabilities
Portfolio exposure Maxi- Mini- Potential loss percentile
Type Mean Std mum mum 0.1% 0.5% 1.0% 2.5% 5.0%
Market 30,825 13,139 86,262 21; 905 42; 322 36; 842 34; 120 29; 766 26; 055
Credit 8;756 12,537 0 143; 981 79; 616 62; 469 55; 113 45; 541 38; 132
Joint 22,079 17,893 85,264 119; 565 90; 148 73; 522 65; 903 55; 120 45; 549
64 F. Jamshidian, Yu Zhu
Fig. 2. 30-day market exposure distribution of swap portfolio
Fig. 3. Joint exposure distribution of both market risk and credit risk (Sample portfolio of
swaps)
Scenario Simulation 65
Fig. 4. Histogram of default losses for a sample portfolio (Sample portfolio of swaps)
7. Concluding remarks
The Scenario Simulation model provides a rigorous yet practical framework
for quantitative analysis of multi-currency portfolio risk. The model is highly
ecient computationally, and can be applied to a wide range of risk manage-
ment tasks such as estimation of portfolio market risk and credit exposure, and
calculation of value-at-risk for a nancial institution.
The basic idea is to discretize the multivariate distribution of market vari-
ables into a manageable number of scenarios, e.g., 105 yield curve scenarios for
each currency. When there are several currencies, we can take a large Monte
Carlo sample, e.g., of size one million, to achieve numerical stability; yet, we
only have to evaluate each transaction in the portfolio 105 times.
1
Value-at-risk was rst introduced by G30 in 1993. See Global Derivative Study Group
(1993). BIS formally adopted the concept in 1995 as the determinant of market risk capital
requirement. For various methods to estimate VaR, see Value-At-Risk, Risk Magazine
Special Supplement, June 1996.
2
We emphasize that in this model, all distributions, Brownian motions, etc., are with respect
to the actual measure, rather than the risk-neutral measure. As such, the model is not intended
for valuation of derivatives, rather, it generates yield curve scenarios for future dates to be
used as input to appropriate valuation models. The model is nevertheless arbitrage free, as
the uncertainty of each key rate carries its own market price of risk. The latter are implicitly
determined by either of the two expectation hypotheses (4a) or (4b). In this connection, we
view the later reduction to three factors as an approximation of the covariance structure,
rather than exact modelling with three factors.
3
See, e.g., Murphy et al. (1995) and Kahn (1989).
4
The historical covariance data are from J.P. Morgan RiskMetrics (5/30/1995 monthly data)
which is a trademark of J.P. Morgan. For its estimation method, see J.P. Morgans technical
66 F. Jamshidian, Yu Zhu
document (1995). It should be noted that the Scenario Simulation model does not depend
on a particular data source.
5
For these derivations, it is useful to note that if R denotes the matrix (R
ij
); then, RR
t
=
R; R
t
R = diag([); and R
1
= diag(1=[)R
t
.
6
For example, Cox and Rubinstein (1985).
7
To determine the appropriate number of scenarios is of course an empirical issue.
8
In Table 2, we use 7 5 3 to represent 7, 5, and 3 scenarios for the rst, second and
third factor respectively. The total number of scenarios is equal to 7 5 3 = 105. Though
not reported here, the comparison of percentiles of the two methods gives similar results.
9
See, e.g., Karlin and Taylor (1981).
10
See, e.g., Billingsley (1986). Note here the symbol means asymptotic to.
11
The correlation between two principal factors of two dierent countries is obtained from
Eq. (9a) applied to both countries.
12
These implied volatilities are distinct from the volatilities in Eq. (2) that drive the key
rates.
13
This is because if Y is a k-dimensional variate with an identity covariance matrix, then
X = AY will have correlation matrix Q.
14
A good reference for the numerical methods mentioned in this paragraph is Press et al.
(1992).
15
Moodys estimated that the average recovery rate (i.e., 1 L) for senior secured bonds is
as high as 53.05%, and for senior unsecured bonds is 45.88%. For senior surbordinated and
subordinated debt the recovery rates become 37.02% and 29.57%, respectively. For junior
subordinated debt, it is merely 16.40% (see Carty and Lieberman, 1996).
16
The expected credit loss for time period t, E[D
t
], can be calculated as
s
max [V
t
(s); 0]
p
t
(s) q
t
, where V
t
(s) is a counterpartys termination value at time t in scenario s, p
t
(s) is
the probability of scenario s, and q
t
is the marginal default probability of the counterparty.
If P(t) is the discount bond price maturing in t, then the total expected credit loss can be
expressed as E[D] =
t
P(t)E[D
t
], which can be interpreted as the fair insurance cost of
the counterpart default.
17
For detailed discussion, see Gluck and Clarkson (1993).
18
These are in fact Standard & Poors 5-year cumulative stressed default probabilities but
used for 1-month period. By taking the 12-th root of S&Ps yearly probability transition
matrix we can obtain one-month default probabilities, which are respectively 0.0202, 0.0243,
0.0354, and 0.0978% for the above rating categories. In other words, the default probabilities
used in Table 7 are more than 50 times larger than S&Ps stressed one-month default prob-
abilities. S&P emphasizes that stressed probabilities are advisable particularly for estimation
of potential credit losses over a short period of time (see Bahar and Gold 1995).
References
Bahar, R., Gold, M.: Structuring derivative product companies: Risk and safeguards. In:
Derivative credit risk: Advances in measurement and management. Risk Publications,
1995
Basle Committee on Banking Supervision: An internal model-based approach to market risk
capital requirements, April 1995
Billingsley, P.: Probability and measure, 2nd edn. New York: John Wiley 1986
Cox, J., Rubinstein, M.: Options markets. Englewood Clis: Prentice-Hall 1985
Global Derivative Study Group (G30): Derivatives: Practices and principles, 1993
Gluck, J., Clarkson, B.: Moodys approach to evaluating derivative products subsidiaries.
Moodys Investors Services, August 1993
Kahn, R.: Risk and return in the U.S. bond market: A multifactor approach. In: Fabozzi,
F. (ed.) Advances and innovations in the bond and mortgage markets. Burr Ridge, IL:
Probus 1989
Scenario Simulation 67
Karlin, S., Taylor, H.: A second course in stochastic processes. New York: Academic Press
1981
RiskMetrics-Technical Document, 3rd edn. Hove, UK: J.P. Morgan 1995
Carty, L., Lieberman, D.: Corporate bond defaults and default rates 19381995. Moodys
Inverstors Service, Jan. 1996
Murphy, B.P., Won, D., Gulrajani, D.: Valuation and risk analysis of international bonds.
In: Fabozzi, F., Fabozzi, T. (eds.) The handbook of xed income securities, 4th edn.,
Homewood: Irwin 1995
Press, W., Teukolsky, S., Vetterling, W., Flannery, B.: Numerical recipes in C: The art of
scientic computing. Cambridge: Cambridge University Press 1992
Value-at-Risk, Special Supplement. Risk, June 1996