Course: Risk Management (Module II: Key Risks and their Measurement) NIBM, Pune

Module II: Key Risks and their Measurement

Section B: Market Risk

Chapter 5: Historical and Monte Carlo Simulation Methods

Prof. Sanjay Basu

Objectives
After studying this chapter, you will understand
 The need for alternatives to Variance-covariance VaR
 The logic of simple Historical Simulation (HS)
 The steps for estimating simple HS VaR
 The strengths and weaknesses of simple HS
 The logic of Monte Carlo Simulation
 The steps for estimating MCS VaR
 The merits and demerits of MCS
Structure
1.0 Introduction
1.1 Simple Historical Simulation
1.2 Steps in Simple Historical Simulation
1.3 Historical simulation: An assessment
2.0 Monte Carlo Simulation
2.1 Fitting distributions
2.2 Generating shocks
3.0 Conclusions

1.0 Introduction
The problems with the Variance-covariance method made practitioners look for alternative
VaR models. Soon they realized that historical or simulated returns could be used in place
of a standard deviation-correlation based framework. Returns capture both the direction

Page 1 of 11
Course: Risk Management (Module II: Key Risks and their Measurement) NIBM, Pune

and severity of future shocks. Moreover, they need not be drawn from the normal
distribution alone. The historical simulation model became very popular, as it was
perceived to be simpler and more tractable than the Variance-covariance approach.
Similarly, the Monte Carlo simulation method is much better at generating future scenarios.
Therefore, with the passage of time, these models gained wider acceptance in financial
markets.
This chapter is organized as follows. Section 1 discusses the rationale for simple HS, the
VaR estimation method and a critical assessment of the concept. In section 2, we follow the
same sequence for MCS. Section 3 concludes.
1.1. Simple Historical Simulation
The Simple Historical Simulation (HS) method predicts changes in current market value of
individual security positions and the portfolio from empirical distributions for historical
security returns. The question we ask is that if future shocks are exactly the same as each of
the observed (logarithmic) returns in the sample dataset, how will the current market
values of individual positions and the portfolio be affected? In other words, we assume that
each historical return captures a possible future scenario. If long (short) positions get
positive (negative) shocks, the security makes profits. Conversely, if long (short) positions
get negative (positive) shocks, the security makes losses. Since returns can be positive or
negative, the bank knows whether a security will gain or lose if such scenarios occur.
Secondly, the chance of gains and losses may not be equal under simple HS, since the
sample dataset can be skewed. Finally, large losses and gains may be more likely than what
the normal distribution suggests. As long as the sample contains such shocks, the impact
will be captured by simple HS.
This is quite unlike shocks based on the standard deviation, for which the direction is
unknown. Therefore, Variance-covariance VaR does not tell the bank whether market value
will rise or fall. Moreover, since standard deviation is symmetric, the chance of gain and
loss is also assumed to be the same. Finally, not only does standard deviation ignore very
large price movements, but the normal distribution also underestimates the likelihood of
extreme shocks. As a result of the double whammy, Variance-covariance VaR fails to
capture stressed losses.
1.2. Simple Historical Simulation – the steps
To assess daily profits and losses, we multiply the current market value(s) by each of the
historical daily returns on individual securities, drawn from the same sample which was
used for the Variance-covariance method. As discussed, each return captures a positive or
negative future shock to individual positions, due to which the bank makes gains or losses.
The portfolio profit or loss is the sum total of the outcomes across all segments. The VaR

Page 2 of 11
Course: Risk Management (Module II: Key Risks and their Measurement) NIBM, Pune

estimate, for a given confidence level, is then derived by sorting the potential gains and
losses and marking off the loss at the appropriate threshold.
Step 1: Estimating Simulated Values of Individual Positions and Portfolio
The simulated value of an individual (e.g. USD) position is:
USD
Vsimt 
 V0USD  1  RtUSD  for t=1,2,…., n (1)

Thus, for n historical USD returns, we obtain n simulated values of the USD position. The
intuition is that the market value of the security will increase/decrease to the extent of the
historical return, if it is used as a potential future scenario. Each shock is equally likely, i.e.
has a probability of (1/n).
The simulated portfolio value is:
P
Vsimt  Vsimt
USD
 Vsimt
equity
 Vsimt
bond1
 Vsimt
bond 2
 ... for t=1,2,…., n (2)

Thus, for n historical returns for each individual security, we obtain n simulated values of
the portfolio.
Step 2: Estimating Simulated Profits / Losses for Individual Positions and Portfolio
The simulated profit / loss for an individual position is:
USD
Gsimt  Vsimt
USD
 V0USD for t=1,2,…., n (3)

The simulated profit / loss for the portfolio is:

P
G simt  Vsimt
P
 V0P for t=1,2,…., n (4)

The inference, from equations (1) and (3), is that the gain/loss is equal to the historical
return multiplied by the current value of a position. Cet. par., a positive (negative) return
increases (reduces) market value and leads to profits (losses) for long positions. If the same
combination of historical shocks is repeated over the horizon, as had occurred in the past,
some segments may gain, while others could lose. Therefore, for a particular set of
scenarios, the portfolio impact is the sum total of the profits and losses across all segments.
This process also captures all cases in which losses in some positions are offset by possible
gains in others. In other words, equation (4) portrays implicit (or built-in) diversification
benefits, without using correlations.
Table 1 presents a snapshot of the simple HS process. It starts with a 10-day sample of
historical returns for each instrument, from which hypothetical profit and loss (P/L)
estimates are assessed. These P/L estimates are then added to arrive at the portfolio P/L
forecast, for each combination of historical shocks. The purpose is to understand how large
losses (on both long and short) may be in the future. In other words, we try to capture the

Page 3 of 11
Course: Risk Management (Module II: Key Risks and their Measurement) NIBM, Pune

extent to which market values could fall or rise in future, if historical shocks repeat
themselves.
Table 1: Simulating Profits and Losses in Simple Historical Simulation
as on 23-09-2016

Historical Returns Simulated Profits and Losses

Date 7952032 USDINR NIFTY 7952032 USDINR NIFTY Portfolio
09-Sep-16 -0.01% 0.15% -0.96% -26.19 1.47 -425.25 -449.96
12-Sep-16 -0.15% 0.53% -1.72% -291.51 5.33 -758.99 -1045.17
14-Sep-16 -0.05% 0.09% 0.13% -98.12 0.93 55.70 -41.49
15-Sep-16 0.05% 0.09% 0.18% 99.86 0.91 80.64 181.40
16-Sep-16 0.01% -0.22% 0.43% 12.78 -2.16 188.00 198.62
19-Sep-16 -0.04% -0.01% 0.32% -73.53 -0.10 143.36 69.73
20-Sep-16 -0.02% 0.22% -0.37% -32.61 2.19 -163.23 -193.65
21-Sep-16 0.00% 0.06% 0.01% -6.34 0.64 6.29 0.60
22-Sep-16 0.27% -0.33% 1.02% 542.87 -3.26 451.98 991.59
23-Sep-16 0.11% -0.21% -0.41% 214.97 -2.07 -179.14 33.76

Step 3: Value at Risk for Individual Positions and Portfolio

After sorting the profits and losses, we pick out the security-wise and portfolio loss
estimates, at a given Confidence Level (e.g. 99%), to determine the instrument-wise and
portfolio VaR. For instance, let us assume that there are 500 P/L estimates, from which we
want to forecast 99% VaR for long positions. By definition, 1% of the worst losses lie above
the VaR threshold. Therefore, we exclude the top five losses (=1% of 500) and choose the
sixth largest loss as the 99% VaR estimate for all instruments, as also the portfolio.
Using the same bank’s data, as in the previous chapter, Tables 2 and 3 illustrate the HS
Security –wise daily VaR at 95% and 99% confidence levels, expressed in absolute amounts
(INR) and as percentage of exposure.

Table 2: Security Wise Daily HSVaR in INR as on 23-09-2016

C.L. 7592029 7962025 8272020 7682023 7492017 7952032 USD NIFTY
95% 159.80 86.28 85.38 60.27 8.78 252.99 4.77 733.36
99% 263.55 156.52 127.79 111.73 15.22 426.18 6.71 995.58

Page 4 of 11
Course: Risk Management (Module II: Key Risks and their Measurement) NIBM, Pune

Table 3: Security Wise Daily HSVaR (% of MV) as on 23-09-2016

C.L. 7592029 7962025 8272020 7682023 7492017 7952032 USD NIFTY
95% 0.13% 0.12% 0.09% 0.11% 0.03% 0.13% 0.48% 1.66%
99% 0.22% 0.22% 0.13% 0.20% 0.05% 0.22% 0.67% 2.25%

Table 3 reiterates the pull-to-par effect – shorter tenor bonds like 7.49% 2017 and 8.27%
2020 lose relatively less, since they mature earlier and are already closer to the face value.
It also shows that NIFTY is the riskiest exposure, while bonds are the safest. Using the same
data, Table 4 illustrates the HS Portfolio daily VaR at 95% and 99% confidence levels, in
absolute amounts (INR) and as percentage of exposure.
Table 4: Portfolio Daily HSVaR in INR and in Percentage as on 23-09-2016
C.L. VaR (INR) VaR (% of MV)
95.00% 1066.52 0.17%
99.00% 1417.92 0.23%

A comparison of Tables 2 and 4 capture the built-in diversification benefits at the 95% and
99% confidence levels – the portfolio VaR estimate is less than the sum of individual VaR
forecasts.

-1,067 1,154
5.0% 90.0% 5.0%
7

5
Values x 10^-4

0
-5000

-4000

-3000

-2000

-1000

1000

2000

3000
0

Figure 1: Portfolio gains and losses for estimating VaR

Page 5 of 11
Course: Risk Management (Module II: Key Risks and their Measurement) NIBM, Pune

Figure 1 shows the sorted portfolio gains and losses, for long positions, when history
repeats itself. We estimate 99% Portfolio VaR from the left tail of the distribution.
1.3: Historical Simulation – An Assessment
It is clear that simple HS is much easier than the Variance-covariance method. It can be
understood and communicated faster to senior and top management. There is no need to
be trained in statistical concepts like standard deviation and correlation. It is also much
simpler to estimate HS VaR than VC VaR, for large and complex portfolios at global banks
and financial institutions, since standard deviations and correlations are not involved at all.
A survey of global banks, which adopted VaR-based capital charges after the Basel-I
amendments in 1996, reports that 73% of the respondents had followed the simple HS
method6.
However, the enthusiasm needs to be mixed with a dose of caution. When we assume that
history repeats itself, we depend too much on historical data. Our sample may not contain
large negative (or positive) shocks, while the future may be very different. As a result,
actual losses may be overestimated (or understated) by the chosen sample. Loss forecasts
from benign phases may underreport actual losses in highly volatile markets, while losses
predicted with data from turbulent periods may overstate losses under normal conditions 4.
Therefore, the sample should be chosen in such a manner that it contains large (both
positive and negative) shocks from the past, so that losses are not at least underestimated.
This exercise requires a sufficiently long dataset on prices and yields, which may not be
available for many instruments.
Even if we are willing to treat past returns as future scenarios, all events may not be
equally likely. Shocks from the recent past may be more probable than those from the
distant past. However, since the basic model is so simple, these adjustments are easy to
make. Declining weights can be assigned to returns further back in the past, to indicate a
lower likelihood of recurrence. VaR estimates, in such cases, will be influenced more by
recent shocks. This method is known as Hybrid or Age-Weighted HS2. If recent shocks are
more adverse (favourable) than past ones, VaR forecasts under Hybrid HS will be higher
(lower) than under simple HS. Similarly, other revisions to simple HS, to include changes in
volatility and correlation over time, can also be made, with little effort 3.
Finally, since HS is not based on parameters like standard deviation, VaR estimates cannot
be linked across confidence levels. In other words, for the Variance-covariance method, it is
easy to convert from one VaR estimate to another, since the multipliers are known for all
confidence levels. However, in case of HS, the VaR forecasts have to be made afresh for each
confidence level. The same problem arises with longer horizons. In the Variance-covariance
approach, T-day VaR can be estimated by multiplying daily VaR by the square root of T. For
example, 10-day VaR99% = 10 × Daily VaR99%. However, in simple HS, 10-day returns may

Page 6 of 11
Course: Risk Management (Module II: Key Risks and their Measurement) NIBM, Pune

have to be computed for the same purpose, leading to loss of intermediate data points and
sharp reduction in sample size.
2. Monte Carlo Simulation
The Variance-covariance method depends on parameters like standard deviations and
correlations and the assumption of normally distributed returns, which tend to
underestimate VaR. Historical simulation presumes that past returns are repeated in the
future. Monte Carlo simulation (MCS) is a possible upgrade on both approaches. It neither
assumes a normal distribution for security and portfolio returns, nor does that history
repeat itself. Instead, it suggests that future scenarios are based on past trends, but are not
identical to past returns.
2.1: Monte Carlo Simulation – fitting distributions
The first step in MCS is to estimate the return distribution for each instrument in the
portfolio. This follows the logic that future shocks should follow the same traits which
historical data exhibits. Therefore, the graph of historical returns is compared with
patterns created by (potentially innumerable) statistical distributions generated by a
software (like best fitTM , from Palisade corporation). The computer-generated pattern that
produces the closest match with the graph of historical returns is the chosen distribution
for the sample data. This is quite like a visual quiz. When we are shown a picture and asked
to identify an object, we eliminate many possible options, based on patterns stored in our
brains. Our response is based on the pattern which resembles the photo the most.
Figure 2 shows the first-choice distribution for 7.59% 2029 and an alternative (rejected)
distribution, for better appreciation.
The historical returns are captured by the solid blue graph. The red curve shows a
loglogistic distribution, while the green curve represents an exponential distribution. It is
clear from the diagram that the Loglogistic distribution fits (or matches) the historical
pattern much better than the exponential distribution. Indeed, many other curves can be
superimposed to show that the loglogistic distribution fits the historical data the best. In
other words, it is the best proxy for historical trends. Hence, the simulated shocks are
based on the loglogistic distribution.

Page 7 of 11
Course: Risk Management (Module II: Key Risks and their Measurement) NIBM, Pune

Figure 2: Comparing fitted distributions for Monte Carlo Simulation

2.2: Monte Carlo Simulation – generating shocks

The second step in MCS is to generate a large number of random shocks from the fitted
return distributions. The term random, in this context, means that all shocks are equally
likely, independent of each other and do not repeat themselves. However, in practice, such
shocks are impossible to obtain. Sooner or later, the series of numbers will be repeated.
Hence, computers produce a sequence of pseudo-random numbers, in which the cycle does
not recur for a very long time. For all practical purposes, it is impossible to distinguish
these shocks from a series of true random numbers. For example, one of the best
generators, the Mersenne twister, creates 219937 – 1 distinct random numbers1.
It is obvious that the scenarios must be correlated, since the historical returns are
interdependent. Therefore, we feed the return correlations, used for Variance-covariance
VaR, into the system and generate 50000 random shocks. The reason for choosing such a
large dataset is that the final results (i.e. VaR estimates) will not vary much from one
sample to another. Table 5 presents a small subset of shocks for all positions.

Page 8 of 11
 Course: Risk Management (Module II: Key Risks and their Measurement) NIBM, Pune

Table 5: Simulated Shocks for Trading Book positions

Iteration 7592029 7962025 8272020 7682023 7492017 7952032 USDINR NIFTY
1 -0.12% 0.14% -0.02% -0.06% 0.01% 0.01% 0.23% 1.82%
2 -0.16% 0.05% -0.01% 0.09% 0.00% 0.13% 0.30% 0.88%
3 -0.04% 0.12% 0.01% 0.10% 0.04% -0.11% 0.23% 0.02%
4 0.00% 0.04% -0.06% -0.01% 0.02% 0.06% 0.19% 0.82%
5 0.01% 0.09% 0.00% 0.02% 0.00% -0.04% -0.84% -0.70%
6 0.44% 0.04% -0.05% 0.12% 0.05% -0.02% 0.34% -0.41%
7 0.01% -0.02% 0.03% 0.04% 0.04% -0.01% 0.06% -0.64%
8 0.05% 0.05% -0.02% 0.04% 0.02% -0.03% -0.08% -0.74%
9 -0.03% 0.06% 0.07% -0.07% -0.02% 0.13% 0.07% 0.05%
10 0.09% 0.00% -0.02% 0.15% 0.02% 0.00% -0.35% 0.20%

The interpretation is simple. For instance, in scenario no. 6, the value of 7.59% 2029 may
rise by 0.44% and USD could appreciate by 0.34%, between the current date and the next
day, though no such shocks may have occurred in the past. As a result of such events, the
impact on Market Value and P/L will be exactly the same as in simple HS. This is shown in
Table 6.
Table 6: Simulated P/L estimates for Trading Book positions
Iteration 7592029 7962025 8272020 7682023 7492017 7952032 USDINR NIFTY
1 -147.08 103.68 -19.87 -31.43 4.61 19.18 2.26 805.00
2 -193.11 33.27 -5.61 52.63 -0.46 264.19 2.96 390.58
3 -47.03 89.63 10.97 59.05 11.60 -214.76 2.32 9.26
4 -4.27 31.29 -59.24 -7.00 5.48 124.11 1.95 362.50
5 15.30 67.22 -3.38 13.59 -0.50 -84.93 -8.42 -307.06
6 528.35 26.90 -46.29 67.40 14.50 -30.55 3.40 -182.35
7 10.30 -15.25 33.49 19.93 13.00 -18.56 0.61 -282.77
8 58.47 37.85 -17.08 20.46 4.79 -49.75 -0.85 -325.68
9 -33.11 45.41 70.16 -39.07 -7.27 260.30 0.65 19.90
10 110.84 2.61 -14.94 87.28 6.89 3.48 -3.49 89.83

Page 9 of 11
Course: Risk Management (Module II: Key Risks and their Measurement) NIBM, Pune

As in simple HS, we sort the P/L estimates and compute VaR at different confidence levels.
The security-specific results are presented in Tables 7 and 8.
Table 7: Security Wise Daily MCSVaR in INR as on 23-09-2016
c.l. 7592029 7962025 8272020 7682023 7492017 7952032 USDINR NIFTY
95% 162.94 87.90 83.89 65.78 9.87 275.91 5.07 670.65
99% 248.69 136.06 126.47 101.92 16.91 415.22 7.98 1057.75

Table 8: Security Wise Daily MCSVaR (% of MV) as on 23-09-2016

c.l. 7592029 7962025 8272020 7682023 7492017 7952032 USDINR NIFTY
95% 0.14% 0.12% 0.09% 0.12% 0.03% 0.14% 0.51% 1.52%
99% 0.21% 0.19% 0.13% 0.18% 0.05% 0.21% 0.80% 2.40%

The pull-to-par effect is evident once again – shorter-tenor bonds lose relatively less (as
share of MV), since they are closer to maturity. We also report portfolio MCS VaR (in INR
and as share of MV) in Table 9.
Table 9: Portfolio Daily MCSVaR in INR and in Percentage as on 23-09-2016
c.l. VaR (INR) VaR (% of MV)
95% 713.39 0.11%
99% 1123.65 0.18%

The MCS process is superior to both Variance-covariance and HS techniques for a number
of reasons. First, it does not restrict the analyst to the assumptions of the Variance-
covariance model – future shocks can be much larger (and more likely) than what the
normal distribution suggests. Secondly, future scenarios may be very different and much
more than what simple HS captures. For instance, assuming that there are 250 trading days
per year, an MCS exercise with 50000 shocks generates two hundred years’ (=50000/250)
worth of relevant scenarios, from limited historical data. While a very long historical series
may be unavailable or outdated, the set of simulated events is still pertinent to the
problem. Hence, MCS is useful for structured scenario analysis, especially when historical
data is scarce. In particular, it is very popular for complex derivatives or illiquid positions,
for which market prices are not readily available.
However, MCS needs advanced computing power for real-life portfolios. Otherwise, it is
tough to generate so many scenarios, on a regular basis, for a large number of instruments

Page 10 of 11
Course: Risk Management (Module II: Key Risks and their Measurement) NIBM, Pune

and revise them from time to time as the composition of the trading book changes. If it is
difficult to estimate portfolio VaR with the Variance-covariance method, MCS poses a far
greater challenge. Furthermore, it requires sophisticated statistical skills. While the
Variance-covariance model presumes prior knowledge of only the normal distribution,
which is quite common anyway, the MCS process is based on an in-depth understanding of
many probability distributions, which may represent historical trends. In particular, choice
of the wrong distribution may distort the nature of random shocks and VaR estimates. In
other words, poor knowledge of statistics and/or a tendency to use convenient short-cuts
may lead to enormous model risk in MCS5. Therefore, MCS is often viewed as a black box –
hard to grasp and implement in practice.
3. Conclusions
As discussed in the previous chapter, the Variance-covariance approach has a number of
flaws. Hence, simulation methods may be used as alternative tools to forecast losses. While
simple HS is intuitive and implementable, the results may crucially depend on the nature
and size of the sample. On the other hand, while MCS is theoretically superior to other
measures and very comprehensive, it is hard to execute for large and dynamic real-life
portfolios. Therefore, as a compromise, most banks use simple HS to estimate VaR-based
capital charges. Its weaknesses were exposed during the global crisis, leading to a new set
of Market Risk guidelines in 2016.
References
1. Alexander, C. (2008): Market Risk Analysis Vol. IV: Value-at-Risk Models, John Wiley and
Sons, Chichester, England.
2. Boudoukh, J, M. Richardson, and R. Whitelaw (1998): The Best of Both Worlds, RISK,
May, 64-67.
3. Dowd, K. (2005): Measuring Market Risk, 2nd Edition, John Wiley and Sons, Chichester,
England.
4. Finger, C. (2006): How Historical Simulation has made me lazy, Riskmetrics Research
Monthly, April.
5. Jorion, P. (2015): Financial Risk Manager Handbook, 6th Edition, John Wiley and Sons,
New Delhi.
6. Pérignon, C. and D. Smith (2010): The Level and Quality of Value-at-Risk Disclosure by
Commercial Banks, Journal of Banking and Finance, 34(2), 362-377.

Page 11 of 11