RiskMetrics (Monitor) 3
RiskMetrics (Monitor) 3
Morgan/Reuters
RiskMetrics Monitor
RiskMetrics News
Version 1.2 of FourFifteen J.P. Morgans system for evaluating, exploring and reporting
market risk has been released.
Please note that on or about July 17, 1997 it is our intention to cease updating the datasets at
the J.P. Morgan web site. They are now and will continue to be available but only at the Reuters web site.
Reuters Ltd
International Marketing
Martin Spencer
(44-171) 542-3260
[email protected]
In the article On measuring credit exposure which appeared in the last issue of the RiskMetrics Monitor we presented three simple models to calculate the credit exposure of market
driven transactions. Since the publication we have received several inquiries concerning the
details of the credit exposure calculation, in particular, as they pertain to the methodology
that relies on monte carlo simulation. In this article, we present a step-by-step analysis of the
credit exposure calculation of a foreign exchange forward contract that employs monte carlo
simulation. The purpose of this article is to offer the reader a more complete understanding of
the credit exposure calculation and to make the reader aware of potential shortcomings of
such a simple credit exposure calculation.
A general approach to calculating VaR without volatilities and correlations
19
25
RiskMetrics Monitor
Second Quarter 1997
page 2
RiskMetrics News
Scott Howard
Morgan Guaranty Trust Company
Risk Management Advisory
(1-212) 648-4317
[email protected]
basket currencies
RiskMetrics Datasets
Reuters has been producing the datasets in parallel for the past 2 months. Therefore J.P. Morgan intends
to cease updating its web site on or about July 17, 1997. We expect that a month will allow those who
currently access J.P. Morgan to make the transition to Reuters. Their web addresses are
http://www.riskmetrics.reuters.com/WDown4.htm and ftp://ftp.riskmetrics.reuters.com/datasets/.
RiskMetrics and FourFifteen@ are registered trademark of J.P. Morgan in the United States and in other countries. They are written with the symbol
on its first occurrence and RiskMetrics and FourFifteen thereafter.
RiskMetrics Monitor
Second Quarter 1997
page 3
In the article On measuring credit exposure which appeared in the last issue of the RiskMetrics Monitor we presented three simple models to calculate the credit exposure of market driven transactions.
Since the publication we have received several inquiries concerning the details of the credit exposure
calculation, in particular, as they pertain to the methodology that relies on monte carlo simulation. In
this article, we present a step-by-step analysis of the credit exposure calculation of a foreign exchange
forward contract that employs monte carlo simulation. In so doing, we highlight some of the important
assumptions underlying the calculation. The purpose of this article is to offer the reader a more complete understanding of the credit exposure calculation and to make the reader aware of potential shortcomings of such a simple credit exposure calculation.
The rest of the article is organized as follows:
In section 1 we present the specifications of a foreign exchange forward contract where a US
importer agrees to purchase German marks (DEM) with US dollars (USD) from a Bank at
some future point in time. We identify the forward contracts cashflows, find its mark-to-market value and compute the forwards current credit exposure.
Section 2 provides the details behind the forwards potential credit exposure calculation.
Potential exposure calculations such as expected, maximum and average exposure, are computed via monte carlo simulation. More specifically, we
- describe the sampling times when we compute the forwards expected and maximum
exposures
- explain how to construct forward foreign exchange and interest rate curves, and
- explain how to produce cashflow maps at each sampling time
Before simulating foreign exchange and interest rates we:
- review the basic assumptions of the model used in the simulation
- explain how forward rates are used to determine the drift component in foreign
exchange and interest rates, and
- describe the technique used to generate returns at each sampling time
Having simulated the rates, we explain how to revalue the forward contract at each sampling
time and how we compute expected and maximum exposure. Finally, we compare the forwards exposure profile to the exposure profile of an interest rate swap.
Section 3 summarizes the main results of the article and provides direction for future research.
RiskMetrics Monitor
Second Quarter 1997
page 4
transaction on April 24, 1997 which is referred to as the analysis date. Table 1 provides the details of
the contract.
Table 1
Foreign exchange forward description
Analysis date: April 24, 1997
Trade date:
Maturity date:
Notional amount
Spot exchange rate
Strike price
US importer
Bank
The forward contracts cashflows, from the importers perspective, are presented in chart 1.
Chart 1
Forward contracts cashflows
US importer perspective
Receive
1 million DEM
23 Jan 97
Trade date
24 Apr 97
Analysis date
1 Nov 97
2 May 01
Maturity date
Pay
693,700 USD
To calculate the US importers current credit exposure we must find the forwards mark-to-market
(MTM) value. In general, the forwards MTM is calculated as follows:
[1]
On April 24, 1997, the current forward USD/DEM exchange rate (from the analysis date to maturity)
is 0.6379 USD/DEM. The US interest rate for the same period (4.022 years) is 6.783%. Using a con1
tinuous discounting procedure we find the MTM to be -42,419 USD, that is,
[2]
1, 000, 000 ( 0.6379 0.6937 ) exp ( 0.06783 4.022 ) = 42, 419 USD
Note that the forwards MTM is negative. This implies that at the analysis date the forward is out-ofthe-money for the importer. Consequently, the US importers current credit exposure is zero. Chart 2
shows the foreign exchange forwards profit and loss (P&L) profile as a function of the current forward
rate.
RiskMetrics Monitor
Second Quarter 1997
page 5
Chart 2
Forwards P&L profile
P&L is reported in USD
Forward's P&L
out-of-the-money
in-the-money
-42,419
0.6379
0.6937(strike)
Intuitively, the foreign exchange forwards value to the US importer is negative because on the trade
date, January 23, 1996, the importer contracted to pay 0.6937 USD/DEM and the forward price has
since fallen. Alternatively expressed, the importer agreed to pay 693,700 USD to buy 1,000,000 DEM
on May 2, 2001. At the analysis date, however, the current price for 1,000,000 DEM to be received on
May 2, 2001 is 637,900 USD. Hence, the USD price of DEM has fallen relative to the USD price of
DEM at the trade date. Consequently, on the analysis date it would cost the US importer 42,419 USD
less to purchase DEM at the current forward price rather than at 0.6937 USD/DEM.
Now, suppose instead of decreasing to 0.6937USD/DEM, the current forward price of DEM increased
relative to the strike price. In this case we would find that the US importers forward contract would
have a positive mark-to-market value (refer to chart 2) it would be in-the-moneysince the importer
could buy DEM (with USD) at a discount relative to the market prices at the analysis date.
2.1 Background
The first step towards measuring potential credit exposures is to define sampling times. Sampling times
are future dates when exposures are calculated. Table 2 presents the sampling times for the US importers forward contract.
RiskMetrics Monitor
Second Quarter 1997
page 6
Table 2
Sampling times
In days and years; calculations assume 1 year = 365 days
Days from analysis date
Years from analysis date
30
0.083
192
0.526
373
1.021
557
1.526
738
2.022
922
2.526
1102
3.022
1287
3.526
Having determined the sampling times, next we need to identify the RiskMetrics interest rate and foreign exchange rate nodes that will be used to map the forwards cashflows. Remember that cashflow
mapping is the process of redistributing the observed cashflows onto so-called RiskMetrics nodes, to
2
produce RiskMetrics cashflows . RiskMetrics provides volatilities and correlations for the following
US and German interest rate nodes:
Table 3
RiskMetrics interest rate nodes
For US and German interest rates
Years
0.083
0.25
0.50
1.0
2.0
3.0
4.0
5.0
7.0
9.0
10.0
15.0
20.0
In addition, RiskMetrics provides volatilities and correlations associated with the logarithmic change
in the USD/DEM exchange rate.
In order to compute potential exposures we will need to simulate a distribution of the forwards value
at each sampling time. It follows from interest rate parity (IRP), that the simulation of the forwards
3
future value can be done in two (equivalent) ways :
1. At each sampling time, simulate forward (USD/DEM) exchange rates and US interest rates and
compute a distribution of future values of the forward contract, or
2. At each sampling time, simulate German and US interest rates as well as spot foreign exchange
rates and compute a distribution of future values of the forward contract.
2.2 Constructing foreign exchange and interest rate curves and cashflow mapping
As a first step, to determine the future values of the forward contract at each sampling time we must
build foreign exchange forward and interest rate curves. Chart 3 shows the US and German term structure of interest rates on April 24, 1997.
2
3
RiskMetrics Monitor
Second Quarter 1997
page 7
Chart 3
USD and DEM zero rates
April 24, 1997
4.50
6.75
USD
6.50
4.00
6.25
DEM
3.50
6.00
3.00
5.75
5.50
2.50
0.1
0.5
1.5
2.5
3.5
Based on these zero rates and the spot USD/DEM exchange rate we can construct USD/DEM forward
rates at each sampling time. We now explain how to construct the forward USD/DEM curve. In general,
us
de
if we let S t denote the spot USD/DEM exchange rate at time t and let r t, T and r t, T denote all US and
German zero rates at time t, for the period between time t and T, then it follows from interest rate parity
that the forward foreign exchange rate F t, T is given by the expression
us
exp r t, T years
F t, T = S t --------------------------------------------[1]
de
exp r t, T years
where years represents the times presented in table 2. Using the appropriate spot exchange rate and zero
interest rates, the foreign exchange forward curve is presented in chart 4.
Chart 4
USD/DEM forward exchange rate curve on April 24, 1997
Spot USD/DEM foreign exchange rate is 0.5824.
USD/DEM forward exchange rate
0.650
0.625
0.600
0.575
0.550
0.525
0.1
0.5
1.5
2.5
3.5
Years
Table 4 provides the US and German interest rates and the corresponding foreign exchange forward
curve at each of the 8 sampling times
RiskMetrics Monitor
Second Quarter 1997
page 8
Table 4
US and German interest rates and foreign exchange forward curve
Spot USD/DEM foreign exchange rate is 0.5824
Years from
analysis date
0.083
0.526
1.021
1.526
2.022
2.526
3.022
3.526
us
de
r t, T (%)
r t, T (%)
Ft
5.687
6.062
6.427
6.450
6.475
6.576
6.674
6.729
3.221
3.266
3.384
3.494
3.607
3.838
4.066
4.294
0.584
0.591
0.601
0.609
0.617
0.624
0.630
0.634
In addition to the foreign exchange forward rates we also need to construct US and German forward
interest rate curves. We must construct such curves at each sampling time. Further, since we will map
the forwards cashflows to the RiskMetrics nodes, the forward rates must coincide with the RiskMetrics
nodes. We will now demonstrate the construction of forward interest rates with an example.
Assuming that bond yields are continuously compounded, we can relate the so-called short rate, r 0 ,
and forward rate, r f to the long rate, r 1 as follows:
e
[2]
r0 T 0 r f T f
= e
r1 T 1
where T 0 , T f and T 1 are the time periods for the short, forward and long rates, respectively.
Chart 5 illustrates the relationship between the time periods when the short, forward and long rates prevail.
r0(T0)
rf(Tf)
Maturity date
Analysis date
Chart 5
Time profile of short, forward and long rates
r1(T1)
r1 T 1 r0 T 0
r f = --------------------------Tf
Given the short and long rates, we can use [3] to solve for the relevant forward rates. For example, suppose the sampling time, T 0 , is 0.526 (about one-half of a year from the analysis date). The (short) US
us
interest rate corresponding to this time period, r 0 , is 6.062%. Our goal is to construct forward interest
us
rates, r f , at the following RiskMetrics nodes; 0.083, 0.25, 0.50, 1.0, 2.0, 3.0, and 4.0 years. We can
now use the short and long rates to solve for the forward rates corresponding to the RiskMetrics nodes.
RiskMetrics Monitor
Second Quarter 1997
page 9
us
These forward rates are presented in table 5 along with the corresponding long rates, r 1 , and the time
at which they occur. Note that the T 1 = T f + T 0 .
Table 5
Constructing US interest forward rates at second sampling time
us
All times are presented in years, T 0 = 0.526; r 0 = 6.062%
Tf
T1
0.083
0.250
0.50
1.00
2.00
3.00
4.00
0.609
0.776
1.026
1.526
2.526
3.526
4.526
us
us
6.126
6.254
6.427
6.450
6.576
6.729
6.819
6.531
6.659
6.813
6.654
6.712
6.846
6.918
r 1 (%) r f (%)
These forward interest rates correspond to the relevant RiskMetrics nodes that are used for mapping at
the second sampling time. Having created the forward rates, the next step is to map the forwards cashflows to the RiskMetrics nodes. To demonstrate the mapping procedure we continue to focus on the
cashflow map that occurs at the second sampling time (i.e., at 0.526 years from the analysis date).
Since the time from the analysis date until maturity is 4.022 years, the forwards cashflows (1,000,000
DEM and -693,700 USD) occur approximately 3.50 years from the second sampling time. Chart 6
shows the relationship between the forwards cashflows and the sampling time.
Chart 6
Forwards cashflows at the second sampling time
4.022 years.
3 months
23 Jan 97
Trade date
24 Apr 97
Analysis date
3.50 years
1 Nov 97
Receive
1 million DEM
2 May 01
Maturity date
Pay
693,700 USD
Since these cashflows occur between years 3 and 4, we map these cashflows to the 3 and 4 year RiskMetrics nodes. Table 6 presents the cashflow maps for the DEM and USD positions at the second sampling time
RiskMetrics Monitor
Second Quarter 1997
page 10
Table 6
Cashflow maps and RiskMetrics nodes
Cashflow occurs in 3.50 years from 0.526 year sampling time
RiskMetrics nodes
DEM position
USD position
3 year
501,369.86
347,800.27
4 year
498,630.14
345,899.73
For a more complete picture of the cashflow maps table 7 presents the cashflow maps and the corresponding RiskMetrics nodes for each of the 8 sampling times. The column denoted Cashflow date
represents the time at which the forwards cashflows occur, adjusting for the sampling time.
Table 7
Cashflow maps for all sampling times
Sampling
time (yrs)
0.083
Cashflow
date (yrs)
3.940
0.526
1.021
1.526
2.022
2.526
3.022
3.526
3.500
3.003
2.500
2.003
1.500
1.003
0.498
345,899 (4)
1,900 (4)
345,899 (3)
1,900 (3)
345,899 (2)
1,900 (2)
698,898 (0.50)
498,630 (4)
2,739 (4)
498,630 (3)
2,739 (3)
498,630 (2)
2,739 (2)
994,520 (0.50)
S t + 1 = S t exp ( t )
ln ( S t + 1 ) = ln ( S t ) + t
where
2
RiskMetrics Monitor
Second Quarter 1997
page 11
S t is a generic variable representing the USD/DEM foreign exchange rate and the relevant US and German interest rates.
Now, suppose we wanted to simulate the USD/DEM foreign exchange rate over one day. To do so, we
2
would need estimates of the one-day mean, fx, t , and variance, fx, t . We would then set S fx, t to the
current spot foreign exchange rate (0.5824 USD/DEM, as of the analysis date) and simulate a random
2
set of numbers, t , from a normal distribution with a mean fx, t and variance fx, t . We would then
substitute these normally distributed numbers as well as the current spot rate into [4] to produce a distribution of foreign exchange rates for the next day.
Extending this approach, if one were to simulate prices over, say, the next T days we would modify [4]
to become
[6]
St + T
= S t exp
t=1
= S t exp ( z t )
2
Note that z t is normally distributed with a mean T t and variance T t . Therefore, if we were to simulate foreign exchange rates T days from the analysis date we would simulate normal random variates
2
with mean T fx, t and variance T fx, t , and then using the current spot rate( S fx, t ) would apply [6].
In practice, while the variance is often estimated from historical data, the mean, or drift parameter is
derived from the current forward rate. We derive an expression for the mean t as follows:
1. Find the expected value of the price T days forward, S t + T . The mathematical expectation of [6] is
given by the following expression
[7]
T
E ( S t + T ) = S t exp T t + ---------t-
2. Set the exchange rates forward value, F t, T equal to the spot rates expected value, i.e,
F t, T = E ( S t + T ) and solve for the mean T t . This yields the expression
2
[8]
F t, T T
T t = ln ---------- --------t St 2
Therefore, for a given forward rate, which was derived in section 2.2, and an estimate of the daily variance of foreign exchange returns, we can back-out the foreign exchange rates drift parameter. At the
second sampling time, the foreign exchange forward rate, ( F t, T ) , is 0.591 and at the analysis date the
spot rate, ( S fx, t ) , is 0.5824. The one-day variance, 2 fx, t , for the returns on the USD/DEM exchange rate is 0.0000211. Substituting this information into [8] yields T fx, t = 1.4%.
Note that we can solve for the interest rate drift parameters in exactly the same way.
Using the drift parameters and the estimated daily variances (and covariances) we can simulate foreign
exchange and interest rates, taking into account the correlation among the rates, by applying any of the
techniques presented in Appendix E of the RiskMetrics Technical Document, 4th edition. Specifically,
RiskMetrics Monitor
Second Quarter 1997
page 12
we will demonstrate how to perform monte carlo simulation by applying the singular value decomposition of the estimated covariance matrix.
Suppose we wish to simulate 1000 foreign exchange and interest scenarios (simulations) at the second
sampling time. We will generate a 1000 x 5 matrix of future prices, Y. At the second sampling time, the
columns of Y represent the USD/DEM foreign exchange rate, the US 3 and 4 year rates and the German
3 and 4 year rates, respectively, 0.526 years from the analysis date. We now demonstrate how we construct Y in three steps.
First, we assume that each row of Y, y i , (i=1,..,1000) is distributed according to the multivariate (5-variate) normal distribution with a mean m i and covariance matrix i . i has dimenus
us
de
de
sion 5 x 5.The mean m i is defined as m i = f x, t 3 y, t 4 y, t 3 y, t 4 y, t
T
RiskMetrics Monitor
Second Quarter 1997
page 13
Chart 7
Histogram of simulated USD/DEM exchange rates
At 0.526 years from analysis date; forward rate is USD/DEM 0.591
Frequency
25
20
15
10
0
0.489
0.516
0.542
0.569
0.595
0.622
0.648
0.675
0.701
Charts 8 shows the histograms of the simulated 3 and 4 year forward interest rates, 0.526 years from
the analysis date.
Chart 8
Histogram of 3 and 4 year US interest rate
Frequency
Frequency
25
25
20
20
15
15
10
10
0
4.66
5.27
5.87
6.48
7.08
7.69
8.29
8.90
9.50
0
4.73
5.34
5.95
6.57
7.18
7.79
8.41
9.02
9.63
RiskMetrics Monitor
Second Quarter 1997
page 14
For example, consider computing the forward foreign exchange rate at the second sampling time. In
this case, the forwards cashflow occurs in 3.5 years from the sampling time and it is mapped to the 3
and 4 year RiskMetrics nodes. Therefore, we will calculate two forward foreign exchange rates corresponding to each node in the cashflow map. If we let SF X 0.526 represent the simulated (spot) foreign
us
us
de
exchange rates at the second sampling time and r 0.526, 3 yr ( r de
) and r 0.526, 4 yr ( r 0.526, 4 yr ) rep0.526, 3 yr
resent the simulated US (German) interest rates for each cashflow map at the 3 and 4 year nodes, then
the two forward foreign exchange rates are given by the following expressions.
[9]
us
exp r 0.526, 3 yr 3
---------------------------------------------FW D 0.526, 3 yr = S F X 0.526
de
(corresponding to the 3 year node)
exp r 0.526, 3 yr 3
and
us
exp r 0.526, 4 yr 4
FW D 0.526, 4 yr = S F X 0.526 --------------------------------------------- de
(corresponding to the 4 year node)
exp r 0.526, 4 yr 4
The distribution of profit and loss on the forward contract 0.526 years from the analysis date is (see
table 7):
[10]
[11]
us
us
Alternatively, it follows from interest rate parity that we can compute the profit and loss on the forward
contract as:
de
de
us
RiskMetrics Monitor
Second Quarter 1997
page 15
$24.2 $46.9
P&L($000)
The P&L distribution has a mean value of -44,439 USD and a standard deviation of 32,941 USD. The
negative mean is the forwards mark-to-market at the second sampling time. This implies that the current exposure at the second sampling time is zero since the forwards forward value is out-of-the-money.
Recall that a party has positive credit exposure only if the contract it currently holds is in-the-money.
Consequently, the distribution of the forwards credit exposure corresponds to the non-negative values
of the P&L distribution. The distribution of the forwards credit exposure at the second sampling time
is presented in chart 10.
Chart 10
Histogram of forward contracts credit exposure (in USD)
At 0.526 years from analysis date
Frequency
65
60
55
10
0
$0.0
$6.9
Chart 10 is the same as chart 9 but with all negative values converted to zero. The mean of the credit
exposure distribution which represents the expected exposure 0.526 years after the analysis date is
RiskMetrics Monitor
Second Quarter 1997
page 16
1,341USD. The maximum exposure of the forward contract at the second sampling time, which is the
95th percentile of the credit exposure distribution, is 9,947USD. The average exposure of the forward
contract, which is a measure of how much, on average, the US importer expects to lose over the life of
the forward contract, is 9,687USD. Chart 11 shows the expected and maximum exposure profiles over
the life of the forward contract for the US importer.
Chart 11
US importers foreign exchange forward exposure profile
Exposure
$140,000
$120,000
Maximum exposure
$100,000
$80,000
$60,000
$40,000
Expected exposure
$20,000
$0
0.08
0.53
1.02
1.53
2.02
2.53
3.02
3.53
Years
Note that the exposure profile of the forward increases over the life of the contract. This reflects the
forwards single payment (at maturity) as well as the increase in the forwards volatility with time.
Compare the forwards exposure profile to that of an interest rate swaps (Chart 12).
Chart 12
IR swaps exposure profile
Exposure (percent of notional)
4.5%
4.0%
Maximum
3.5%
3.0%
2.5%
2.0%
1.5%
Expected
1.0%
0.5%
0.0%
0
Sampling times
Note that whereas the interest rate swaps profile has a hump shape, the forwards exposure profile increases towards the forwards maturity. The difference in the appearance of the exposure profiles is due
to the amortization of the swaps payments. In other words, unlike the forward whose cashflows remain
RiskMetrics Monitor
Second Quarter 1997
page 17
constant over the life of the contract, the swaps cashflows decrease (as more payments are made/received) over its life. Therefore, for the swap the volatility effect dominates at the beginning of the
swaps life and then is overtaken by the so-called amortization effect. However, in the case of the
forward, the volatility effect dominates the forwards exposure profile since there is no amortization.
RiskMetrics Monitor
Second Quarter 1997
page 18
RiskMetrics Monitor
Second Quarter 1997
page 19
RiskMetrics Monitor
Secpmd Quarter 1997
page 20
If there are a large number of benchmarks relative to the number of returns per benchmark, the
dataset of benchmark returns is smaller than the correlation matrix
Better numerical precision
An intuitive interpretation of how Monte Carlo returns are generated
Fast marginal risk analysis.
We now explain how users can work directly with the return time series (RTS). Define the RTS dataset
as follows. For benchmark i, let r i be the T 1 vector of returns (with mean 0), where T is the number of returns (historical observations) for each benchmark. Define r i, t to be the return in period t of
benchmark i . Let R = r 1, r 2, ...., r n , where n is the number of benchmarks. We can write the T x n
matrix of returns R as follows.
r 11 r 1n
[1]
R =
r JJ
r T 1 r Tn
Since R is a T n matrix, the RTS dataset has Tn values. Compare this to the VCV dataset which
has n standard deviations, and n ( n 1 ) 2 correlations, for a total of n ( n + 1 ) 2 values. For situations
where the number of benchmarks is more than twice the number of observations, RTS requires less
storage, i.e., R requires less storage than its corresponding covariance matrix.
[2]
C = T R R.
T
where R is the transpose of R. Hence, R provides a simple factoring of the covariance matrix. This
will be useful later.
VaR = 1.65 w Cw
assuming that underlying returns are distributed according to the conditional multivariate normal distribution. Now, it follows from [1] that
[4]
T 1 T
Rw .
As we can see from the right-hand side of equation [3], the VaR calculation depends only on the benchmark weights, w , the underlying return matrix, R, and the number of historical observations T. Because
the computational effort varies linearly with the number of benchmarks, using the RTS matrix is faster
RiskMetrics Monitor
Second Quarter 1997
page 21
than using the covariance matrix C (provided the number of benchmarks is more than twice the number
of observations in each benchmark).
The preceding analysis demonstrates how to compute VaR by the VCV method when the data are equally weighted. However, we note that this methodology is general and applies equally well when the data
are weighted exponentially.
[5]
r 11
r 1n
r 21
r 21
R =
T1
J1
rT 1
r JJ
T1
r Tn
Now, we can define the covariance matrix based on exponential weighting simply as
[6]
C =
i 1
i=1
1
T
R R
T
R R
where
T
[7]
i1
i=1
It follows immediately from the results presented in [4] that VaR in this case is
[8]
VaR = 1.65
R w
Once a decay factor, , is selected in the VCV method, the correlation matrix is computed, and it is no
longer possible to change the weight without recomputing the entire covariance matrix. However, since
there are no assumed weights in the RTS dataset, we are free to choose different values without any
additional computational burden.
RiskMetrics Monitor
Secpmd Quarter 1997
page 22
The decomposed matrix does not easily provide an intuitive understanding of how the deviates
are generated
Changing a single return value of one benchmark requires a new decomposition
Cholesky decomposition requires that the correlation matrix be PD (positive definite), and
SVD requires PSD (positive semi-definite).3
Using the RTS dataset, we can take a more direct approach that suggests an intuitive interpretation.
Consider a row of the R matrix. It represents one observation interval--a snapshot of all benchmark returns. Suppose we multiply each row by a normally distributed random variate, and add the rows. The
result is a vector of Monte Carlo returns that have the correct variance and correlations. In other words,
Monte Carlo returns are simply the sum of random multiples of the benchmark snapshots.
We now show how to perform Monte Carlo with the RTS that avoids any volatility and correlation calculations.
Let = { 1, 2, ...., T } be a T 1 vector of independent, N(0,1) random variates. Then we are
T
interested in simulating a 1 n vector of correlated returns, R . Consider the ith component,
T
T
ri = r i . Note that R is just a sum of independent, normally distributed random variables (all with
mean zero, and standard deviation vector r i ). So the mean of the sum is the sum of the means which
are zero. And the variance is the sum of the variances, which is the variance of r i . So the mean and
variance of ri is the same as that of r i . Unlike r i , ri is truly normally distributed.
What about the correlation of ri with other simulated returns? Consider another simulated return, rj .
The correlation is = cov ri, rj i j . Since ri and rj have mean zero,
ij
cov ( r i, r j ) = E [ r i r j ]
T
= E t = 1 t r i, t t = 1 t r j, t
[9]
= E t = 1 t r i, t r j, t + 2 t = 1 s = t + 1 ( t s r i, t r j, t )
Because the i are independent, the cross terms have zero expected value. Since the expectation of the
sum equals the sum of the expectation, the right-hand side becomes
[10]
E t = 1 t r i, t r j, t
= t = 1 r i, t r j, t E t = t = 1 r i, t r j, t = cov ( r i, r j )
Hence, covariance and correlations are also preserved. In general, we can show that the covariance maT
T
trix of the 1 x n vector of random variables y= R is C= R R . This follows immediately from the definition of the variance of y where the mean of y is zero. Letting E(x) denote the mathematical
expectation of x, we can write the variance of y as follows:
The correlation matrix cannot be PD if there are more benchmarks (columns of R) than there are observations in each
benchmark (rows of R). In fact, due to roundoff errors, even PSD is rarely satisfied. These problems can be dealt with by
tweaking the correlation matrix to become PSD (which introduces other errors), or by increasing the number of observations in each benchmark (which bloats the dataset without adding significant information, and may require data that is
unavailable, or outside the sample of interest). The RTS dataset allows analysis with fewer observations. Of course, one
must still be careful that the chosen dataset is sufficiently representative of the relationships between benchmarks.
RiskMetrics Monitor
Second Quarter 1997
page 23
[11]
T
T T
Variance ( y ) = E y y = E R R
T
T
= R E R
= C
Now, certainly Monte Carlo with a covariance matrix represents a performance improvement in terms
of pre-processing (i.e. in decomposing a correlation matrix), since we do not need any pre-processing.
However, is it a fast way to generate deviates? Once the T independent normal deviates are generated,
simulating each benchmark uses T multiples. This compares with k (where k is the rank of R) multiplications when using a Cholesky or SVD decomposition.
However, we can make return generation still faster. By randomly sampling the observation vectors
(i.e. using only a subset), we can still preserve correlations and variances. In fact, we can generate our
Monte Carlo returns with just one random normal deviate per trial, and one multiply per benchmark.
Details are left for future discussion.
3. Other applications
The RTS dataset also allows more powerful tools for marginal analysis, and ad hoc manipulation of the
benchmark returns. The Monte Carlo sampling technique discussed above, as well as marginal analysis
techniques based on RTS, are employed by CreditManager (J.P. Morgans credit risk calculator) for
Monte Carlo simulation of credit portfolios. Details of these methods may be covered in a subsequent
note.
4. Conclusions
As the number of benchmarks grows relative to the effective number of observations per benchmark,
it becomes more efficient to use the RTS dataset. Moreover, using the RTS dataset provides a number
of benefits over VCV-based approaches in terms of flexibility. In particular, weve shown that it provides an intuitively appealing technique for generating Monte Carlo samples.
RiskMetrics Monitor
Secpmd Quarter 1997
page 24
RiskMetrics Monitor
Second Quarter 1997
page 25
A look at two methodologies that use a basic delta-gamma parametric VaR precept but achieve
results similar to simulation.
RiskMetrics Monitor
Fourth quarter 1996
page 26
RiskMetrics Monitor
Second Quarter 1997
page 27
RiskMetrics Monitor
Second quarter 1997
page 28
RiskMetrics products
New York
Chicago
Mexico
San Francisco
Toronto
North America
Europe
London
Brussels
Paris
Frankfurt
Milan
Madrid
Zurich
Asia
Singapore
Tokyo
Hong Kong
Australia
RiskMetrics is based on, but differs significantly from, the market risk management systems developed by J.P. Morgan for its own use. J.P. Morgan does not warrant any results
obtained from use of the RiskMetrics data, methodology, documentation or any information derived from the data (collectively the Data) and does not guarantee its sequence,
timeliness, accuracy, completeness or continued availability. The Data is calculated on the basis of historical observations and should not be relied upon to predict future market
movements. The Data is meant to be used with systems developed by third parties. J.P. Morgan does not guarantee the accuracy or quality of such systems.
Additional information is available upon request. Information herein is believed to be reliable, but J.P. Morgan does not warrant its completeness or accuracy. Opinions and estimates constitute our judgement and are
subject to change without notice. Past performance is not indicative of future results. This material is not intended as an offer or solicitation for the purchase or sale of any financial instrument. J.P. Morgan may hold a
position or act as market maker in the financial instruments of any issuer discussed herein or act as advisor or lender to such issuer. Morgan Guaranty Trust Company is a member of FDIC and SFA. Copyright 1996 J.P.
Morgan & Co. Incorporated. Clients should contact analysts at and execute transactions through a J.P. Morgan entity in their home jurisdiction unless governing law permits otherwise.