Extreme Value Theory

EVT

• Extreme value theory can be used to investigate the properties of the

right tail of the empirical distribution of a variable x. (If we interested
in the left tail we consider the variable –x.)
• We first choose a level u somewhat in the right tail of the distribution
• We then use Gnedenko’s result which shows that for a wide class of
distributions as u increases the probability distribution that v lies
between u and u+y conditional that it is greater than u tends to a
generalized Pareto distribution
EVT
• Generalized Pareto Distribution
• This has two parameters x (the shape parameter) and b (the scale
parameter)
• The cumulative distribution is
1 / 
  
1  1  y
  
EVT
• Estimating the two parameters x (the shape parameter) and b (the
scale parameter) using Maximum Log Likelihood Estimate
• The observations, xi, are sorted in descending order. Suppose that
there are nu observations greater than u
• We choose x and b to maximize
nu 1  ( v i  u ) 
1 /  1

 ln  
1  
 
i 1

    

EVT & Power Law

Our estimator for the cumulative probabilit y that the

variable v is greater th an 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 th erefore explains why the power law

holds so widely
 VAR and Expected Shortfall using EVT

The estimate of VaR when the confidence level

is q is obtained by solving

1 / 
nu  VaR  u 
q  1 
1   

n   
It is

  n 



VaR  u    (1  q )   1
   nu  
 