You are on page 1of 24

BEAM046

FINANCIAL MODELLING

Dr J. Shen

Lecture 6
Risk Measurement
Intended learning outcomes

By the end of this lecture students should

• Be familiar with the different types of risk in financial markets


• Understand the concept of value at risk
• Be able to calculate value at risk using the variance-covariance approach
• Be able to calculate value at risk using the historical simulation approach
• Be able to undertake back testing of value at risk models

Reading

• S. Benninga, Financial Modeling, MIT Press, 2008, Chapter 15


• Kevin Dowd, Beyond Value at Risk, Wiley, 1998. Chapters 2, 3, 4 and 5.
• Philippe Jorion, 2006, Value at Risk: The New Benchmark for Managing Financial Risk, McGraw Hill,
Chapters 5 and 6.
• http://www.gloriamundi.org

2
Introduction

• Risk is inherent in any business activity that involves uncertain cash flows

• There are six main types of risk that firms may face

Business risk: uncertainty over the business environment that is specific to the firm
Market risk: uncertainty over the price of financial assets
Credit risk: uncertainty over the repayment of debt
Operational risk: uncertainty over the internal systems of a company and the people who operate them
Liquidity risk: uncertainty over the ability to find an immediate market for financial assets
Legal risk: uncertainty over the enforceability of legal contracts

• Ignoring risk can have disastrous consequences (e.g. Barings Bank, Orange County, Metallgesellschaft,
Sumitomo Corp, Daiwa Bank, Long Term Capital Management)

• Most businesses engage in risk management

• However, before risks can be managed, they must be measured

3
Value at Risk

• The single most important risk measurement tool is Value at Risk, or VaR, which embodies the market
risk of a portfolio, or indeed of a whole firm, in a single number

• VaR is defined as the maximum loss on a portfolio that can be expected over a certain time interval with
a certain degree of confidence

Example
If a portfolio whose current market value is £10 million has a 10-day 95% VaR of £200,000, then over
the next ten days, with 95% certainty, the largest loss that should be expected is £200,000.

• VaR is therefore (the negative of) the appropriate quantile (or percentile) of the distribution of profit and
loss (or return) of the portfolio

• VaR is a measure of the maximum loss expected from ‘normal’ market movements

• Central to the definition of value at risk is (a) the holding period and (b) the confidence level

• The units of VaR is the units of the portfolio (e.g. GBP, USD or EUR), but that we could write this as a
percentage of the initial portfolio value (i.e. as a return)

4
The Background to Value at Risk

• VaR was first used in the 1980’s by large financial firms to measure the risk of their trading portfolios

• In 1993, the ‘group of thirty’ (a group of 30 prominent US financiers, bankers and academics)
commissioned a report that recommended the use of VaR for all banks to measure the risk of their
derivatives positions

• VaR gained popularity following the launch by JP Morgan in 1994, of Riskmetrics, an integrated package
to enable companies to measure VaR

• In 1996, the Basle Committee on Banking Supervision required banks to set capital requirements
according to ‘internal models’ of market risk, leaving the choice of model to each bank, subject to certain
stringent criteria. Under BCBS recommendations, the estimated VaR for a bank translates directly into
the capital requirements that the bank must meet. The BCBS recommends that banks hold capital equal to
three times the 99% 10-day VaR

• Depending on the effectiveness of the VaR model used, the bank may be required to increase their capital
requirement

• VaR is now widely used not just among banks, but by other financial and non-financial companies

5
The Different Approaches to Calculating Value at Risk

The Variance-Covariance Approach

• Assumes a particular distribution for the portfolio return distribution, or for the underlying factors that
drive returns, and uses knowledge of that distribution to compute the appropriate quantile

Historical Simulation

• Uses the historical distribution of portfolio returns in order to compute the appropriate quantile

Monte Carlo Simulation

• Assumes a particular distribution for the underlying factors that drive portfolio returns, and then draws
randomly from these in order to simulate the distribution of portfolio returns

Extreme Value Theory

• Recognises that the tails of all fat-tailed distributions have a common shape, which is parameterised by
the tail index, and once this is estimated for a given portfolio, extreme quantiles of the portfolio return
distribution can be calculated

6
The variance-covariance approach

• Consider a portfolio comprising a 1000 USD investment in the S&P500 index

• Assume that continuously compounded (i.e. log) daily returns on the S&P500 are normally distributed

• We can estimate the mean, standard deviation and other statistics of daily returns using historical data

• We could use simple returns, but log returns are more likely to be normally distributed

7
The variance-covariance approach

• The VaR of the portfolio is found by first finding the appropriate percentile of the log return series, i.e.
the log return VaR, under the assumption that log returns are normally distributed

• This tells us that there is only a five percent


chance that the log return on our portfolio
will be less than –1.34%

8
The variance-covariance approach

• We could also compute this by noting that the percentile of a normal variable is the same as the percentile
of a standard normal variable, transformed by the mean and standard deviation of the variable

ln %VaR = −(c(α ) * σ + µ )

• where c(α ) is the α -percent quantile of the


standard normal distribution, σ is the standard
deviation of returns, µ is the mean return

9
The variance-covariance approach

• We can then use this log return to calculate the simple return VaR of our portfolio using the following
formula

%VaR = −(e − ln %VaR − 1)

• This tells us that there is only a five percent


chance that dollar loss on out portfolio will
be more than $13.31, or 1.33% of the value
of the portfolio

10
The Variance Covariance Approach

• It is common to assume that the mean return is equal


to zero, and so that the log return VaR of the portfolio
is given by

ln %VaR = −c(α ) * σ

• To compute VaR over longer horizons, we can use the


result that if returns are serially uncorrelated and
homoscedastic, σ (T ) = T σ and so when µ = 0 ,

VaR(T ) = T * VaR (1)

11
The Variance Covariance Approach

Advantages of the variance covariance approach

• Easy to implement for simple portfolios

• Does not require much historical data

Disadvantages of the variance covariance approach

• Poor approximation for ‘non-linear’ portfolios

• Relies critically on the distributional assumption of normality

12
Historical Simulation

• The variance covariance approach relies heavily on the distributional assumptions about portfolio returns.
An alternative approach is to use historical data to compute the quantile of the distribution of the portfolio
value

• In particular, historical simulation takes the distribution of actual daily returns over the previous period,
applies those returns to the current
portfolio to obtain a simulated Histogram of Daily S&P500 Returns
120
distribution of changes in portfolio value
100

• We can then use the appropriate quantile 80


of this distribution to estimate value at 60
risk
40

• Note that we could use simple returns 20

(because the non-normality of simple 0


returns is irrelevant with historical

%
0%

0%

0%

0%

0%

0%

0%
40

00

60

20

80

40
simulation), but to be consistent with the
.8

.2

.6

.0

.4

.8

.2
0.

1.

1.

2.

2.

3.
-3

-3

-2

-2

-1

-0

-0
previous example, we use log returns

(-) 95% one-day ln% VaR for S&P500 daily returns

13
Historical Simulation

• In order to compute the 95% one-day VaR of


the portfolio, we require the 5th percentile of
this empirical log return distribution, which
we can obtain using Excel’s PERCENTILE
function

• This can then be used to compute the dollar


VaR of the portfolio, as before

• This compares with $VaR of $13.75 using the


variance-covariance approach

14
Historical Simulation

Advantages

• Easy to compute for any (linear or non-linear) portfolio

• Doesn’t impose any assumptions about the distribution of returns

Disadvantages

• Requires large sample of data to estimate the probability of rare events

• Data from the distant past may no longer be relevant

• Cannot be used to estimate the probability of something that hasn’t happened before

15
Back Testing

• Firms have a wide range of models to choose from in their estimation of VaR, and many ways of
implementing each of these models

• Inevitably, some models will work much better than others

• If a model systematically understates the true VaR of a firm’s portfolio, then it will hold insufficient
capital to cover unexpected losses

• A critical component of VaR implementation is the back testing of VaR models

• Back testing involves comparing the forecasts of a VaR model with actual returns, ex post

• The problem is that we do not observe VaR, and so a direct comparison is not possible

• Instead, there are a range of indirect measures of performance that can be used to evaluate the
performance of a VaR model

16
Back Testing

• The simplest measure is the unconditional coverage of the VaR


model

• Suppose that we have estimated VaR each day using the previous
1 year (251 days) of data

• Every day, we have a 95% VaR estimate and an actual return; If


the VaR model is accurate, the actual return should exceed the
estimated VaR on only 5% (i.e. one minus 95%) of days

• We can create a new series, which is a ‘1’


if there is a VaR exception, and a ‘0’ if
there is not

17
Back Testing

• We can use this series to calculate both the number of


exceptions and the percentage of exceptions

• To test whether the unconditional coverage of a VaR


model is equal to the nominal confidence level, we can use
an LR test

• If p is the nominal significance level (one minus the nominal confidence level), N is the number of
exceptions and T is the total number of observations

[ T −N
LRU = 2 ln ((1 − N / T ) /(1 − p) ) (( N / T ) / p ) N ]

18
Back Testing

• Under the null hypothesis of correct unconditional coverage, the LR statistic has a chi-squared
distribution with one degree of freedom; the critical values for the LR statistic can be obtained from excel
using the CHIINV function

• In this case, the LR statistic is less


than the critical value, and so we can
not reject at the 5% significance
level the hypothesis that the VCV
model has correct unconditional
coverage

• In addition to the measures of


unconditional coverage and
independence, there are many other
measures that can be used to evaluate
a VaR model

• Some of these measures evaluate the conditional coverage of the VaR model, in other words whether
VaR exceptions are forecastable. Others evaluate the forecast of the entire density of portfolio returns,
rather than just a single quantile

19
Back Testing

• Back testing results are used by the BCBS in conjunction with the bank’s estimate of VaR in order to set
the bank’s capital requirements

• The benchmark capital requirement is three times the estimated 99% 10-day VaR, but a ‘penalty’ capital
requirement that depends on its unconditional coverage

• Using a back testing sample of 250 days, the capital requirement for a bank is increased from three to up
to four times the estimated VaR depending on the number of exceptions recorded

Exceptions Capital requirement factor

0 to 4 3.00
5 3.40
6 3.50
7 3.65
8 3.75
9 3.85
10 or more 4.00

20
Volatility Clustering

• A particular feature of virtually all short horizon financial asset returns is that they display volatility
clustering: large returns (of either sign) tend to be followed by more large returns; small returns (of either
sign) tend to be followed by more small returns

• Allowing for volatility clustering should improve VaR estimates, since if we know we are currently in a
high volatility period, our VaR estimates should be correspondingly higher

• In order to incorporate volatility clustering into the calculation of VaR, we need a model of time-varying,
or conditional, volatility

• The simplest model of time-varying volatility is a rolling window estimator; more sophisticated models
include EWMA and GARCH

21
Incorporating Volatility Clustering into VaR estimates

• To incorporate volatility clustering into the variance-covariance approach involves replacing the
unconditional estimate of volatility with the conditional volatility from the model

ln %VaRt = −c(α ) * σ t

• To incorporate volatility clustering into historical simulation, we compute the empirical quantile of the
historical standardised distribution (i.e. the distribution of returns scaled by each day’s estimated standard
deviation from the model) and then scale this up by the model’s forecast of tomorrow’s standard
deviation

• To incorporate volatility clustering into Monte Carlo simulation, we simply draw randomly from a pre-
specified distribution with an estimate of the conditional volatility from the model

22
Practical exercise 6

You are a risk management analyst for an investment bank. Your manager wants you to compute the 60%,
90% and 99% VaR for a position in 50,000 December 2009 call options on Apple, with an exercise price of
185.00. The VaR horizon is the maturity of the option contract. She would also like you to compute the VaR
for a portfolio of the underlying shares.

You decide to estimate VaR using Monte Carlo simulation. Monte Carlo simulation proceeds by generating
simulated price paths for the underlying asset, valuing the option portfolio at the VaR horizon for each of
these simulated price paths, and using the distribution of portfolio values across all the simulated price paths
in order to estimate the portfolio VaR.

1. Obtain the option price, associated underlying share price and time to expiry from www.cboe.com.

2. Obtain daily data for the underlying share from finance.yahoo.com. Use this to calculate daily log returns
and estimate daily volatility.

3. To simulate the stock price on day t, St , you decide to use the following process

ln St = ln St −1 + σε t

23
where σ is the daily volatility of log returns and ε is a standard normal variable (i.e. a normally
distributed variable with a zero mean and unit variance). Note that we are assuming that the mean log
return is equal to zero as we are considering only a very short horizon.

We can therefore simulate the stock price by generating a series of standard normal variables, and
applying the above formula. Use the random number generator in Excel (Data/Data Analysis/Random
Number Generation) to generate 1000 realizations of a standard normal random number for each day.

4. Using the simulated values of ε , the initial stock price, S 0 , and your estimate of daily volatility, σ ,
compute the series for ln S t for each realisation. Compute also the series St . Plot all the simulated series St .

5. Using the simulated value of ST (the price of the share at the expiry date of the option), compute the value
of the option, CT at expiry.

6. Using the simulated values of ST and CT , compute the simple return on the share portfolio and the option
portfolio for each realisation. Use these returns to compute the VaR for the share portfolio and the option
portfolio. Plot the empirical distributions of the share and option returns.

24