Market Risk VaR:

Historical Simulation
Approach

N. Gershun
 Historical Simulation
• Collect data on the daily movements in all
market variables.
• The first simulation trial assumes that the
percentage changes in all market variables are
as on the first day
• The second simulation trial assumes that the
percentage changes in all market variables are
as on the second day
• and so on
 Historical Simulation
continued
• Suppose we use n days of historical data with
today being day n
• Let vi be the value of a variable on day i
• There are n-1 simulation trials
• Translate the historical experience of the
market factors into percentage changes
• The ith trial assumes that the value of the
market variable tomorrow (i.e., on day n+1) is
vi
vn
vi 1
 Historical Simulation
continued
• Rank the n-1 resulting values
• VaR is the required percentile rank
 Example of Historical
Simulation
• Assume a one-day holding period and 5%
probability
• Suppose that a portfolio has two assets, a one-
year T-bill and a 30-year T-bond
• First, gather the 100 days of market info
Date T-Bond Value % Change T-Bill Value % Change
12/31/10 102 - 97 -
12/30/10 100 2.00% 98 -1.02%
12/29/10 97 3.09% 98 0.00%
: : : : :
: : : : :
9/12/10 103 -2.91% 96 2.08%
9/11/10 103 0.00% 97 -1.03%
 Example of Historical
Simulation cont.
• Apply all changes to the current value of assets
in the portfolio

• T-bond value = 102 x % change

T-bill value = 97 x % change
T-Bond Modeled T-Bill Modeled Portfolio

Date % Change Value % Change Value Value

12/31/10 2.00% 104.04 -1.02% 96.01 200.05
12/30/10 3.09% 105.15 0.00% 97.00 202.15
: : : : : :
: : : : : :
9/12/10 -2.91% 99.03 2.08% 99.02 198.05
9/11/10 0.00% 102.00 -1.03% 96.00 198.00
 Example of Historical
Simulation cont.
• Rank the resulting 100 portfolio values

• The 5th lowest portfolio value is the VaR

Rank Date Value
1 11/12/10 195.45
2 12/1/10 196.24
3 10/17/10 197.13
4 10/13/10 197.60
5 9/11/10 198.00
: : :
: : :
99 12/8/10 202.15
100 9/25/10 203.00
Notes on Historical Simulation
• Historical simulation is relatively easy to do:
Only requires knowing the market factors and
having the historical information

• Correlations between the market factors are

implicit in this method because we are using
historical information

• In our example, short bonds and long bonds

would typically move in the same direction
 Accuracy
Suppose that x is the qth quantile of the loss
distribution when it is estimated from n
observations. The standard error of x is
1 q (1  q )
f ( x) n

where f(x) is an estimate of the probability density

of the loss at the qth quantile calculated by
assuming a probability distribution for the loss
 Example
• We are interested in estimating the 99 percentile from
500 observations
• We estimated f(x) by approximating the actual
empirical distribution with a normal distribution mean
zero and standard deviation $10 million
• Using Excel, the 99 percentile of the approximating
distribution is NORMINV(0.99,0,10) = 23.26 and the
value of f(x) is NORMDIST(23.26,0,10,FALSE)=0.0027
• The estimate of the standard error is therefore

1 0.01 0.99
  1.67
0.0027 500
 Example (cont.)
• Suppose that we estimated the 99th
percentile using historical simulation as
$25M
• Using our estimate of standard error, the
95% confidence interval is:

25-1.96×1.67<VaR<25+1.96×1.67
That is:
Prob($21.7<VaR>$28.3) = 95%
Extension 1
 Extension 2
• Use a volatility updating scheme and adjust the
percentage change observed on day i for a
market variable for the differences between
volatility on day i and current volatility
• Value of market variable under ith scenario
becomes

vi 1  (vi  vi 1 ) n 1 /  i
vn
vi 1

– Where n+1 is the current estimate of the volatility of

the market variable and i is the volatility estimated
at the end of day i-1
 Extreme Value Theory
• Extreme value theory can be used to investigate
the properties of the right tail of the empirical
distribution of a variable x. (If we are interested
in the left tail we consider the variable –x.)

• We then use Gnedenko’s result which shows that

the tails of a wide class of distributions share
common properties.
 Extreme Value Theory
• Suppose F(*) is a the cumulative distribution
function of the losses on a portfolio.

• We first choose a level u in the right tail of the

distribution of losses on the portfolio

• The probability that the particular loss lies

between u and u +y (y>0) is
F(u+y) – F(u)
• The probability that the loss is greater than u is:
1-F(u)
Extreme Value Theory
 Extreme Value Theory
• Gnedenko’s result shows that for a wide class of
distributions, Fu(y) coverges a Generalized
Pareto Distribution

17
 Generalized Pareto Distribution
(GPD)
• GDP has two parameters  (the shape parameter)
and  (the scale parameter)
• The cumulative distribution is
1 /ξ
 ξ 
cdf  F(y)  1  1  y 
 β 
• The probability density function
1 1
 1    ξy  ξ
pdf  f(y)    1   
 β    β 
 Generalized Pareto Distribution
•  = 0 if the underlying variable is normal
•  increases as tails of the distribution become heavier
• For most financial data >0 and is between 0.1 and 0.4

1.0
fx(x)

=+0.5
0

0.5
=-0.5

0.0
0 1 2 3 4
/
Generalized Pareto Distribution
(cont).
• G.P.D. is appropriate distribution for
independent observations of excesses
over defined thresholds
• GPD can be used to predict extreme
portfolio losses
Maximum Likelihood Estimator
• The observations, i, are sorted in descending
order. Suppose that there are nu observations
greater than u
• We choose  and  to maximize

nu  1  (v  u )  1 /  1 
 ln  1  i
 
i 1      

21
 Tail Probabilities
Our estimator for the cumulative probability
that the variable  is greater than x is
1 / 
nu  x u
  
 
1
n 

Setting u    we see that this correspond s to the power law

Prob(v  x)  Kx -
where
1 / 
nu   1
K   
n   

Extreme Value Theory therefore explains why

the power law holds so widely
Estimating VaR Using Extreme
Value Theory
The estimate of VaR at the confidence level q
is obtained by solving
1 / 
nu  VaR  u 
q  1 1   
n   
It is

  n  
VaR  u   (1  q )  1
  nu  
 Estimating Expected Shortfall
Using Extreme Value Theory
The estimate of ES, provided that the losses exceed the
VaR, at the confidence level q, is given by:

VaR  β  ξu
ES 
1 ξ
 Example
• Consider an example in the beginning of the
lecture. Suppose that u= 4 and nu = 20. That is
there are 20 scenarios out of total of 100 where
the loss is greater than 4.
• Suppose that the maximum likelihood
estimation results in = 34 and = 0.39
• The VaR with the 99% confidence limit is
 Example
• The VaR with the 99% confidence limit is


34 100 
0.39


VaR  4   (1  0.99)  1
0.39 
  20  

 197.25