Professional Documents
Culture Documents
RiskMetrics Monitor
First quarter 1996
BIS revises market risk supplement to 1988 Basle Capital Accord
Expanding the parametric VaR approach for the treatment of options
Introducing a new tool for loading RiskMetrics data into an Excel spreadsheet
Lessons learned from implementing the RiskMetrics data synchronization algorithm
This quarter, along with the section titled RiskMetrics News, we address the following
subjects:
On December 12, 1995, the Basle Committee on Banking supervision released a
communiqu announcing the publication in January 1996 of a revised supplement to
the 1988 Basle Capital Accord which deals with market risk.
Responding to the reactions by the banking community to the initial draft of the proposal to
allow financial institutions to use internal models to estimate market risk as a basis for setting
capital requirements, the Basle Committee has modified one of the quantitative provisions of
the model-based approach: Subject to approval by domestic regulators, banks will be allowed
to account for correlation effects across asset classes.
While simulation is often portrayed as the preferable solution in estimating market
risks for portfolios which contain sizeable positions in nonlinear instruments,
parametric approaches can be refined and implemented to effectively deal with nonnormal P/L distributions.
This article looks at two methodologies which use a basic delta-gamma parametric VaR
precept but achieve similar results to the simulation approach at a much lower cost in
terms of computational intensity.
To enable users to develop customized Excel spreadsheet based applications, we
have developed an Add-In tool that can both access volatility and correlation data from
the datasets and perform basic VaR calculations.
The RiskMetrics Add-In tool will be available by February 1, 1996 from our Internet
site. It will be accompanied by a set of RiskMetrics volatility and correlation datasets
specifically formatted for use with the Add-In. These files will be updated on a daily basis
along with the standard sets.
New York
January 23, 1996
RiskMetrics News
Upcoming changes to the RiskMetrics datasets
At the end of February 1996, the following changes will be made to the RiskMetrics
volatility and correlation datasets which are posted daily on the Internet:
1. The revised methodology outlined in the last edition of the RiskMetrics Monitor (Fourth
quarter 1995) to estimate volatility over a 25-day (1-month) horizon will be implemented.
2. The data synchronization algorithms detailed in the RiskMetrics Monitor (Third quarter
1995) and revised in the fourth article of this document will be implemented.
3. The 1-day and 1-week RiskMetrics vertices for money market interest rates will be dropped
from the daily volatility and correlation files. In testing the data synchronization algorithms,
we discovered that the infrequent daily changes in these rates interfered with the process used
to adjust data for time zone differences. Also, given the low level of price volatility at the very
short end of the curve, we felt that users could safely anchor cash flows at the 1-month vertex.
The global benefit achieved by adjusting the data for timing differences outweighed the effect
of dropping the two low-end vertices.
Summit System, Inc. joins ranks of RiskMetrics developers
Summit System, Inc.
20 Exchange Place, New York, NY 10005
Harvey Rand (1-212) 269-6990, FAX (1-212) 269-6941, sumsales@pipeline.com
Summit-VaR is an extensive family of tools that has been enriched with a comprehensive
collection of VaR analyses which includes among other analytic capabilities, the
J.P.Morgan RiskMetrics methodology.
Summits RiskMetrics methodology is geared toward estimating the global risk of linear
positions. A complementary scenario analysis, based on stress testing, has been implemented
for options products to capture gamma risk. Summit VaR series also include specific
analyses for bond-related products in the emerging markets, where the lack of liquidity of
some portion of the yield curve may not allow traditional interest rate based analysis to
capture the risk properly.
As with its coverage of interest rate products, Summit also provides a Forex VaR implementation utilizing the RiskMetrics approach. Finally, with an optional parametric approach,
Summit generates historical VaR figures based on the past behavior of the markets.
New York
January 23, 1996
RiskMetrics Monitor
page 3
On December 12, 1995, the Basle Committee on Banking Supervision announced that it would
publish a revised version of the Supplement to the 1988 Basle Capital Accord during the course
of the month of January. The original version of this supplement which incorporated proposals
defining the scope and usage of internal bank models to estimate market risks was made public
in April 1995 and opened for comment by financial institutions.
The revised proposals, set to be implemented by the end of 1997, essentially address two of
the concerns banks expressed in their official responses to the original draft:
Internal bank models will be allowed to use correlations both within and across asset classes
to estimate a fully diversified Value at Risk.
Banks will be allowed to scale up their (mostly) daily internal Value at Risk estimates to the
10-day horizon required by the regulatory authorities. Banks had argued in their response to
the original proposal that it would be burdensome for them to maintain two frameworks for
estimating market risk using different horizons and that it would be prohibitively expensive to
back-test models for the 10-day horizon as firms would be required to keep position and price
data for 10 overlapping portfolios.
None of the other quantitative standards (10-day risk horizon, 99% confidence, at least 1year historical volatility data, capital charge based on the higher of the VaR estimate or
three times the average VaR over the preceding 60 business days) have been modified.
The changes do not affect the RiskMetrics Regulatory
dataset as it is currently produced. The estimates of volatility which are published daily are based on a 1-day, 95%
confidence basis and can easily be rescaled. Correlations
both within and across asset classes are included in the
dataset files. Further the RiskMetrics regulatory datasets
have recently been expanded to incorporate all information
which was previously only included in the standard sets
such as yield data.
While addressing some of the concerns of financial institutions, the revised proposals are
still unclear on a number of issues:
While the document mentions that banks will be allowed to use correlations across asset
classes, the document makes their use subject to approval by individual country supervisors, which will judge the integrity of the methodology used. Supervisors will therefore
be burdened with the added responsibility of verifying whether correlations are updated
regularly and tested for stability (which can have a different meaning whether one is speaking
to an economist or statistician). How this will be implemented in practice remains to be seen.
Banks will be allowed to scale up their daily VaR estimates to arrive at the 10-day holding
period. It is not clear which methodologies will be permitted (square root of time?) and what
impact this will have on back testing the results. It has been our experience that back-testing a
daily VaR estimate would tell little about the accuracy of a 10-day holding period model.
These points may be clarified when the complete document is released some time this
month. We suspect, however, that the proposals may be interpreted differently by country
supervisors and that the exact modus operandi of the proposals will be made clearer through
implementation experience.
New York
January 23, 1996
RiskMetrics Monitor
page 4
where -/+ 1.65 are the 5th/95th percentiles of the standardized normal distribution. Over short
horizons, the estimate of E(Rp) is often set to zero to reduce the noise in estimating the sample
mean.2
In general, when a portfolios payoff is a nonlinear function of some underlying returns,
even if these returns are distributed normally, the confidence interval estimate for the
expected value of the portfolio using CV is inappropriate. For example, portfolios with
nonlinear payoffs may have skewed return distributions. Skewness invalidates the application of symmetry imposed by the scale factors +/- 1.65 (the quantiles of the standard normal
distribution). In addition, nonlinearities transform the moments (e.g., mean, variance,
skewness, etc.) of the underlying return distribution. Therefore, assumptions placed on the
expected values of underlying returns do not necessarily carry over to a portfolios expected
values.
In order to properly evaluate the risk of a portfolio that contains nonlinear instruments,
researchers often propose full simulation routines. According to this methodology, a path of
future underlying prices are generated and the portfolios value, which consists of options,
is revalued at various prices along the path. A specific type of full simulation, known as
Structured Monte Carlo, is outlined in the RiskMetrics Technical Document. A major
disadvantage of the full simulation approach is that it is computationally and time intensive.
This study presents two methods to compute the Value at Risk estimates of portfolios with
nonlinear payoffs that do not require full simulation. Its goal is to present a methodology
that is relatively simple to implement and does not require a lot of computer time. The methodology is developed from first principles and is used to compute the VaR over a five-week
horizon of the following position.3
On April 18, 1995, a U.S. dollar based investor buys a USD1,000,000 nominal value 2-year
French franc government bond (OAT Strip) at a yield of 7.147%. In order to hedge FX exposure, the investor buys a 5-week (which corresponds to the investment horizon) FRF/USD put
1
New York
January 23, 1996
RiskMetrics Monitor
page 5
option on a notional USD 870,994 at the money forward FX rate of 4.864. The current value of
the option is FRF/USD 0.04616. Therefore, it costs USD 8,289 to hedge USD 870,994.
Market risk is often analyzed in terms of its delta (1st order) and gamma (2nd order) risk.
Below, in the case where VaR incorporates gamma risk, we compare results obtained using
the methods described herein to those given by full simulation which serves as our benchmark. In so doing we highlight important differences and similarities among the different
methodologies. To facilitate the discussion we will use the following definitions and
parameter settings.
PV = Amount of current position in USD (present value)
= USD870,994
RB = 5-week return on 2-year OAT
B = 5-week forecast standard deviation of bond price
= 0.7757% (yield standard deviation=5.83%)
RX = 5-week return on FRF/USD exchange rate
X = 5-week forecast standard deviation of FX rate
= 3.117%
= 0.532
RO = return on FX option
We analyze the risk of the synthetic portfolio as follows. First, the option is ignored and we
compute the Value at Risk of the position that consists only of the purchased OAT. We then
introduce the option to hedge foreign exchange risk on the OAT. We focus on two specific types
of market risk associated with holding the option delta risk and gamma risk. Initially we focus
exclusively on delta risk since the standard VaR methodology is still applicable. Next, gamma
risk is added. Now, the nonlinearity of the options payoff dominates the portfolios return
distribution. Since standard VaR methodology is no longer appropriate, we use a normal
analytical approximation known as the Cornish-Fisher expansion to find the percentiles of this
portfolios distribution. These percentiles are then used to estimate VaR. Finally, we conduct an
experiment to determine how the normal approximation performs. The method used in this
experiment, which we refer to as partial simulation, actually turns out to be an alternative
technique that may be used to compute VaR in the presence of gamma risk.
The unhedged position
Suppose the investor buys the OAT but does not hedge its foreign exchange exposure. In this
case the return on the portfolio, RP, is simply the sum of the returns on the FRF/USD and OAT
which is written as RP=RX + RB. Its standard deviation is:
[1]
1 = 2B + 2X + 2 * B, X B X
Assuming that RX and R B are normal, as in standard VaR, we know that RP is also normal so the
Value at Risk of holding the foreign bond is
[2]
Hedging FX exposure
In an effort to reduce foreign exchange market risk the investor buys a put option on the FRF/
USD exchange rate. Having purchased the put option, the return on the portfolio that also
New York
January 23, 1996
RiskMetrics Monitor
page 6
consists of the OAT is now RP=RB+RX+RO . In order to compute the return on the portfolio, RP,
we need an expression for the return on the option, RO. We value the FX option using the
Garman-Kohlhagen formula. Specifically, for a given set of parameters denote the options
value by V(PX(t),K,,,X) where:
PX(t) = FRF/USD spot rate at time t
K
A first step toward obtaining an expression for RO is to approximate the future value of the
option V( PX(t +1) ,K,, r, X) with a 2nd-order Taylor series expansion around the current values
(spot rates), PX(t) , K, , r, and X. This yields4
1
2
[3] V ( PX (t +1) , K , r, , X ) Vo ( PX (t ) , K, r, , X ) + ( PX (t +1) PX (t ) ) + 2 ( PX (t +1) PX (t ) )
1
(dPX ) 2 where the options delta
2
[4]
or
dV = Px(t) * * (R x ) +
1
2
2
* Px(t)
* * (R x )
2
Here, dV relates an absolute change in the value of the option to a relative change in the foreign
exchange rate. However, since dV is still in units of PX we need to standardize it to make it
unitless. This allows us to obtain the relative return on the option
[5]
R o = * (R x ) +
1
* Px(t) * * (R x )2
2
R P = RB + RX + R X +
1
2
PX(t) (R X )
2
In the following expression, delta () is the first derivative of V with respect to PX and gamma () is the second
derivative of V with respect to PX.
New York
January 23, 1996
RiskMetrics Monitor
page 7
Since the investor is purchasing a put to hedge foreign exchange risk, < 0. For ease of exposition we will use = in the following analysis. Notice from [6] that we can compute two
types of Value at Risk estimates. The first incorporates the linear risk of the option. In this case,
we use only the first three terms of [6]. Alternatively, we can capture the nonlinear features of
the options return distribution by also including the gamma effect (the last term).
If the investor only wants to account for the delta component of the option, the return on the
portfolio is:
[7]
R P = R B + (1 * )R X
2 = 2B + (1 * ) 2 2X + 2 * (1 * ) 2B, X
Assuming that both RB and Rx are normal implies that RP is also normal. The Value at Risk of
this hedged position accounting only for the delta risk of the option is
[9]
Intuitively, it can be seen from [8] that when *=1, the FX risk is completely hedged and all that
is left is interest rate risk. On the other hand, *=0 implies there is no hedge and the position is
as if the investor held a foreign bond (see [1]). For values of * between 0 and 1, the VaR of
holding a foreign bond and FX option will be lower than if no option was held.
As previously shown, the portfolios return that accounts for both the delta and gamma effect of
the option is
[10]
R GP = RB + (1 * ) RX +
1
2
PX(t) (R X )
2
A consequence of including the term is that R GP s distribution becomes right skewed. To see
how the options delta and gamma components effect the portfolios return distribution, chart 1
presents probability density functions (pdf) for two portfolio return series. One pdf is based only
on the delta component (see [7]), the other is based on both the delta and gamma components
(see [10]).
Chart 1
delta
50
40
30
20
delta+gamma
10
0
-0.04
-0.02
0.02
0.04
0.06
New York
January 23, 1996
RiskMetrics Monitor
page 8
A striking feature from chart 1 is the skewness embedded in the return distribution that includes
the gamma effect. In fact, the distribution (grey line) that only accounts for delta is simply a
scaled version of the normal distribution. Finally, for future use we need to derive the standard
deviation of R GP which is presented below
2
3 = B2 + (1 * )2 * X2 + 2 * (1 * ) * B,X
+
[11]
1
* PX2 (t ) * * X4
2
This expression5 follows from the assumption that RB and RX are normal.
Accounting for the distributional features of R GP
In the presence of gamma risk, normal VaR methodology which relies on the critical values +/1.65 will give misleading risk estimates. The reason for this is simple: +/- 1.65 come from the
normal distribution, however, as seen from chart 1, the gamma component of the option causes
the portfolio return distribution to be highly skewed. The inaccuracy of normal VaR is shown in
chart 2 below. We plot R GP s distribution (grey line) and that given by the normal distribution
with a zero mean and a variance equal to 32 (black line).
Chart 2
Normal VaR
True VaR
30
20
10
0
-0.08
-0.04
0.04
0.08
Essentially, by incorporating gamma risk the investor reduces his risk by the difference between
true VaR and normal VaR. The discrepancy between true and normal VaR leads us to search for
methods that augment the standard VaR methodology to account for the skewed return distribution. In particular, we seek counterparts to the quantiles +/- 1.65 that capture the skewness of
R GP s distribution.
Focusing on analytical solutions, there are basically three approaches we could take to find the
percentiles of R GP s distribution. First, we could match the moments of R GP to a general family
of distributions (Pearson family), second, we could construct the distribution of R GP as a
deformation of a standard normal variables, and, third, we could use the moments of R GP and a
normal analytical approximation to estimate the percentiles of R GP . In this article we describe
how to apply this last method. We find the critical points of R GP s distribution (i.e., the counterparts to -/+ 1.65) by applying a formula known as the Cornish-Fisher expansion.
Applications of normal analytical approximations are motivated by the understanding that any
New York
January 23, 1996
RiskMetrics Monitor
page 9
any distribution can be viewed as a function of any other one. For example, the 5th and 95th
percentiles of R GP s distribution denoted cv.05 and cv.95 can be calculated as a function of the
standard normal percentiles z.05 =-1.65 and z.95 =1.65, and R GP s estimated moments. To be
more specific, consider again the normal 90% confidence interval around the mean portfolio
return E[RP]
[12]
Under the maintained assumptions, when Rp is no longer normal, that is, when Rp becomes R GP
we can write the approximate confidence interval for E[ R GP ] as
[13]
{[ ]
= {E[ R ] + (cv
[ ]
[ ] + (cv
.05 ) * , E
RPG
.95 ) *
The main purpose of the correction term s is to adjust for skewness. To a lesser extent it corrects
for higher-order departures from normality. In the case of the normal approximation interval,
s.05=s.95=0. In practice, the Cornish-Fisher expansion allows us to compute the adjusted critical
values cv.05 and cv.95 as a function of the normal critical values z.05 and z.95 directly.6
[14]
cv = z +
1 2
1 3
1
( za 1) * 3 +
( z 3z ) * 4
(2 z3 5z ) * 32
24
36
6
where
3=E[( R GP -E[ R GP ])3/3 measures R GP s skewness
4=E[( R GP -E[ R GP ])4/4-3 measures R GP s kurtosis
For example, if we wanted to compute the adjusted percentile cv.05 associated with -1.65, we
would use:
[15]
cv.05 = 1.65 +
1
1
1
((1.65)2 1) * 3 +
((1.65)3 3(1.65)) * 4
(2(1.65)3 5(1.65))32
6
24
36
Under the assumption that returns are normally distributed, 3 and 4 can be written
directly as a function of the variances and covariances of RB and R X. This result is very
useful since multivariate extensions are straightforward and standard VaR calculations
already require a covariance matrix. The measures 3 and 4 depend on the cumulants of R GP
where the first four cumulants of R GP , denoted {1, 2, 3, 4 } are defined as
1 = E[ R GP ]
[16]
In this article we present only the first 4 terms of the Cornish-Fisher expression. For a sample size n, this
approximation has an error of order O(n-3/2). For a more complete version, see Johnson and Kotz (1970).
New York
January 23, 1996
RiskMetrics Monitor
page 10
5th
Normal
Cornish-Fisher approximation
Relative difference
-1.650
-1.176
+28.7%
95th
1.650
2.029
+22.9%
Having calculated the adjusted percentiles, the 90% confidence interval for the expected return
on the portfolio that consists of an OAT and a put option is:
C G = {1.176 P + E(R GP ), E(R GP ) + 2.029 P}
where p = 1.78% and E(R GP ) = 0.745% . Using these results, the VaR of this portfolio is
[17]
When applying [15] it is important to remember that this expression is exact only when the true
values of the standardized cumulants are used. In practice, however, we evaluate [14] using
sample estimates of the standardized cumulants. When sample estimates are used to evaluate a
mathematical expression we face what is known as a certainty equivalence problem. Essentially what happens is that the estimation error embedded in the sample estimates is carried over
to the numerical value produced by [14]. Consequently, if there is a lot of estimation error, the
Cornish-Fisher critical values (cvs) will be inaccurate. To determine how estimation error
affects the Cornish-Fisher approximation, we find the critical values of R GP by simulating its
distribution and then finding the 5th and 95th percentiles. Note that this is not the same as full
simulation mentioned earlier because here nothing is revalued. All that is required is that we
generate a matrix of normal random numbers denoted Y and then apply [10].
This partial simulation approach works as follows. Let N denote the number of simulated
random variables and define an Nx2 matrix of independent normal random variablesY=[Y1 Y2]
where Y1 and Y2 are both N x 1 random. Using Y and the covariance matrix of RB and R X
denoted , simulate X=[RB RX] , an Nx2 matrix of correlated normal random variables vectors.8
Defining =[1, 1-] (2x1), =[0,] (2x1) and PX as the spot FRF/USD exchange rate, the
distribution of R GP is generated using the expression:
[18]
R GP = X * +
1
Px * X 2 *
2
Exact formulae for the cumulants are provided in a Technical Appendix that is available from the author upon
request
8 See RiskMetrics Technical document (3rd edition) for details
New York
January 23, 1996
RiskMetrics Monitor
page 11
Denoting the 5th percentile of the standardized distribution of R GP by m.05, VaR under partial
simulation is:
[19]
Table 2 summarizes the results of this section and presents VaR estimates produced by full
simulation.
Table 2
Normal
Lower 5% (VaR)
Upper 5%
-20,698
20,698
Cornish
Fisher
-18,285
31,543
Partial
simulation
Full
simulation
-19,243
29,789
-19,596
33,538
The Cornish-Fisher expansion and the partial simulation approach give similar results and both
are an improvement over the normal model (in comparison to full simulation). In full simulation,
the 5th percentile of the profit/loss distribution is USD -4,008. Since the mean of the distribution
is USD 15,588, the adverse price move from the mean (VaR) is USD 19,596.
The expected value of a portfolio with gamma risk
In the previous section we established how accounting for the gamma effect
1 2 * PX(t) * * (R X )2 skews the portfolios return distribution. Another feature of including this
term is that even if it is assumed that the underlying returns have a zero mean
(i.e., E[R B]=E[ RX ]=0), the portfolios expected value, E[ RPG ] , is not necessarily zero since
1
* * PX(t) * X2
2
= 0.745%
[20]
E[R GP ] =
or
Using this as the appropriate mean portfolio return, we can compare the lower and upper bounds
of 90% confidence intervals generated by normal, Cornish-Fisher, partial and full simulation
methods. These bounds are presented in table 3.
Table 3
Normal
-14,210
27,186
Cornish
Fisher
Partial
simulation
-11,797
38,031
-12,755
36,277
Full
simulation
-4,008
49,126
New York
January 23, 1996
RiskMetrics Monitor
page 12
Finally, recall that the coverage cost of USD 870,994 is USD 8,289. Since this is a sunk cost,
when computing the confidence bands it should be subtracted from the portfolios expected
return. Table 4 presents confidence bands which are adjusted for the both the portfolios
expected return and the options cost.
Table 4
Normal
Cornish
Fisher
Partial
simulation
Lower 5%
Upper 5%
-22,499
35,475
-20,036
46,320
-21,044
44,566
Full
simulation
-12,297
40,837
Conclusion
This article describes two alternative methodologies to estimate VaR on portfolios that include
options when accounting for gamma risk. The Cornish-Fisher expansion provides an analytical
approximation to the percentiles of a portfolios true return distribution. In VaR estimation,
these percentiles are used in place of their normal counterparts +/- 1.65. Based on a synthetic
portfolio that consists of a French government bond (OAT) and a foreign exchange option, this
approximation offers an improvement over normal VaR estimates. However, in practice, the
Cornish-Fisher expression may yield inaccurate results because it is evaluated at sample
estimates of skewness and kurtosis rather than at their true values. To address this issue, a
partial simulation approach is suggested that requires the simulation of correlated multivariate
normal random variates. These variates are then transformed into the portfolios return distribution from which the appropriate percentiles are found. While partial simulation does not require
the calculation of sample estimates, its main drawback is that the percentiles it produces are
subject to simulation error. Ultimately, evaluating the performance of these adjustments is an
empirical issue. This requires estimating VaR on portfolios of different size and composition.
Those interested in such studies as well as the technical details of this paper including extensions to larger portfolios should contact the author using the number (e-mail address) listed
above.
New York
January 23, 1996
RiskMetrics Monitor
page 13
macro name
Turn error logging on
clears matrix from memory,e.g.,rm2dly
loads specified matrix
Note that the volatility and correlation file name prefix is included in the second and third
command.
The last parameter passed in the VAR_TS_LOAD_MATRIX command, i.e., list, refers to the
list of the instrument names you wish to load. This can include from one to all of the
RiskMetrics instruments. This list can be a reference to a range in a spreadsheet or macro file
or a separate text file.
Shown below is a sample list of instruments. Instrument names used by the Add-In are case
New York
January 23, 1996
RiskMetrics Monitor
page 14
1
2
3
4
5
6
7
8
9
AUD.XS
AUD.SO2
AUD.Z02
BEF.S02
BEF.Z02
BEF.XS
CAD.XS
CHF.XS
Example
For this example we suppose you are USD based and have the following positions (expressed in
USD equivalents):
FRF
GBP
GBP
Assume the daily dataset for December 15, 1995 (rm2dly.sit) has been downloaded, decompressed, and you have renamed the two files it contains to have the same prefix.
Volatility file, dv121595.vol has been renamed to rm2dly.vol
Correlation file, dc121595.cor has been renamed to rm2dly.cor
Open or load the Add-In file. (Refer to appendix for platform specifics).
Next, open a spreadsheet that could look like the one pictured below. The order of the positions
is not important, e.g., GBP.S05 could have been listed first or last. Note that because you are
USD based you have foreign exchange risk as well as interest rate risk. The position and
instrument code columns can be reversed provided the parameters in the function call are
referenced to the appropriate cell.
A
1
2
3
4
5
6
7
8
9
1 0
Market
value
Position
-80
55
25
55
-55
Instrument
GBP.S05
FRF.S03
GBP.Z07
FRF.XS
GBP.XS
Diversification effect
Total VaR
DEaR
Volatility
30.1594
0.37699
20.9803
0.38146
15.2091
0.60837
44.1451
0.80264
15.2519
0.61008
70.0060
55.7398
The formula that would be entered in C4 for DEaR of the GBP 5-year swap position:
New York
January 23, 1996
RiskMetrics Monitor
page 15
J
1
2
3
4
5
6
7
8
9
Correlations
Instrument 1
DEM.S05
DEM.S05
DEM.S05
DEM.S05
FRF.S03
FRF.S03
FRF.S03
Instrument 2
FRF.S03
GBP.Z07
FRF.XS
GBP.XS
GBP.Z07
FRF.XS
GBP.XS
Value
0.02784
0.66728
0.34253
0.02825
0.40524
-0.25479
-0.15662
New York
January 23, 1996
RiskMetrics Monitor
page 16
Appendix
Loading the Add-In
PC (Windows)
Open the JPMVAR.XLL file from the File menu or
Add JPMVAR.XLL to the Add-In Manager list of Add-Ins
MacIntosh
Double-click on the JPMVAR icon to start Excel
Open the JPMVAR file from the File menu
Add JPMVAR to the Add-In Manager list of Add-Ins.
Add-In Verification
You can verify that the Add-In has been properly loaded by selecting a cell on a spreadsheet
and typing:
=JPMVAR(VERSION)
This will return JPMVAR Add-In Version 6.00 if the Add-In is loaded.
Unloading the Add-in
PC (Windows) - UNREGISTER(JPMVAR.XLL)
MacIntosh - UNREGISTER(JPMVAR)
If the Add-In was loaded by adding it to the Add-In Manager list, it can be unloaded by
removing it from the list.
Add-In function call general syntax
= JPMVAR(Add-in Function Name, Pararmeter 1, , Parameter N)
Sample Formulas
=JPMVAR(VAR_TS_VOL,rm2dly,AUD.XS) for volatility or
=JPMVAR(VAR_TS_CORR,rm2dly,AUD.XS,DEM.XS) for correlation
Errors
Errors conditions are detected and recorded in a file called error.log. When unexplained
errors occur, users should ensure that error logging is ON and examine the error.log file. Its
location is:
PC(Windows) - C:\ root directory
MacIntosh - desktop
Errors can also be returned by using the Error_get_log function (see below).
New York
January 23, 1996
RiskMetrics Monitor
page 17
Analytics
The VaR calculation works in two steps. First it takes the position array and converts it into an
intermediate position risk array using the position vector and the respective price volatility for
each instrument:
PositionRiskArray=Position Array * Volatility Vector
Next the routine computes Value at Risk.
Valid Input
Valid prefix
Valid Input
(A)rray or (F)ile
N/A
New York
January 23, 1996
RiskMetrics Monitor
page 18
VAR_TS_CALC
Returns the RiskMetrics DEaR estimate for the single or set of positions based upon the
instrument defined amount..
=JPMVAR(VAR_TS_CALC,rm2dly,Range 1,Range 2)
Input Parameters
Valid Input
N/A
N/A
Note: When entering in the formula as an array you must press the appropriate keys to have the
input recognized as an array; see below. Also range 1 and range 2 must have the same number
of rows or columns. Output is one VaR per row/column pair.
PC (Windows) - Control+Shift+Enter
MacIntosh - Command+Enter
(Use the Enter key and not the Return key)
VAR_TS_VOL
This routine returns the volatility of a given instrument, e.g. Australian dollar against the
U.S. Dollar (AUD.XS) or the five year U.S. government zero rate (USD.Z05). Each volatility
represents 1.65 standard deviations. The matrix must already be loaded in memory with the
VR_TS_LOAD_MATRIX function. The instrument name argument is matched to an instrument name associated with the volatility vector in memory.
=JPMVAR(VAR_TS_VOL,rm2dly,A2)
Input Parameters
Valid Input
VAR_TS_CORR
This routine returns the correlation between two given time series. The matrix must already
be loaded in memory with the VAR_TS_LOAD_MATRIX function. The time series names
arguments are matched to time series names associated with the correlation matrix in memory.
New York
January 23, 1996
RiskMetrics Monitor
page 19
=JPMVAR(VAR_TS_CORR,rm2dly,Reference 1, Reference 2)
Input Parameters
Valid Input
ERR_LOG
This routine turns error logging on and off (the default is off).
=JPMVAR(ERR_LOG,1)
Input Parameters
Valid Input
1 is On
0 is Off
ERR_GET_LOG
This routine returns the last messages in the error.log file, up to 20 messages. The ERR_LOG
must be turned on before this routine is used. Output is refreshed when the function is
recalculated. You MUST force recalculation by editing the formula or via macro logic.
=JPMVAR(ERR_GET_LOG)
There are no input parameters.
New York
January 23, 1996
RiskMetrics Monitor
page 20
for k j
for k = j
3) Calculate the eigenvalues of the matrix KR-1. Denote these eigenvalues by di for i=1,2,...,N.
4) Select the largest allowable correction factor where
1
0=
, 1
min di
5) Form the corrected covariance matrix M such that M=R+*K and its associated
correlation matrix Mcorr.
6) Check to make sure the following two conditions are satisfied:
i. Mcorr is positive definite.
ii. Mcorr does not contain any elements are larger than one in absolute value.
7) If (i) or (ii) is not satisfied, reduce and repeat steps 5 and 6. Otherwise keep Mcorr as
the new correlation matrix.
The time at which a price or yield is recorded determines whether or not a time series will be
included in the adjustment process. If a pair of series are recorded at times that differ by
eight hours or more, then the resulting correlation estimate was adjusted. In summary, the
following instruments have been included in the process.
New York
January 23, 1996
RiskMetrics Monitor
page 21
Asset Class
Market
Maturities
Money market
Government bonds
Interest rate swap
Equity
All
All
All
All
Refer to the RiskMetrics Technical Document (Section D) for the schedules of times.
Recall that the algorithm requires the covariance matrix to be an n-dimensional square matrix
and positive definite. We synchronized matrices of varying size, ranging from 2x2 to 100x100,
and checked the reasonableness of the results. Reasonable was defined with respect to the
original correlations an unreasonable result would be one whose synchronized correlation
series was markedly more noisy than the original. Incidentally, this noise may be due to the
computers imprecision and not from estimation. In the end, we determined that a square matrix
of order 2 produced the most satisfactory adjusted correlation estimates.
Further, note that the algorithm for the corrected covariance matrix, M, was given as
M=R+*K. Also note that the unadjusted (or, current, RiskMetrics) covariance matrix, R, is
estimated as
2 current, t = 1 t21 + (1 1 ) Xt Yt
where 1 = 0.94
where 2 = 0.98
2
2
The new elements of the final matrix are then 2 new,t = current,t
+ adjusted,t
.
By decrementing by f in - 0.0001 steps we arrived at the most reasonable results and limited the
computational costs.
We based our experiments and observations presented in this section on the RiskMetrics daily
horizon estimates. In the fourth quarter 1995 edition of the RiskMetrics Monitor, we outlined
an alternative volatility forecasting method for the RiskMetrics monthly horizon estimates.
Denoted as alternative RiskMetrics (or Alt R in the charts and tables), this new method allow
us to apply the same techniques to adjust the monthly correlation estimates as well.
New York
January 23, 1996
RiskMetrics Monitor
page 22
New York
January 23, 1996
RiskMetrics Monitor
page 23
New York
January 23, 1996
RiskMetrics Monitor
Morgan Guaranty Trust Company
page 24
Risk Management Services
Jacques Longerstaey (1-212) 648-4936
riskmetrics@jpmorgan.com
RiskMetrics products
Chicago
Mexico
San Francisco
Toronto
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 judgment 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.