Parametric Approaches (II):
Extreme Value
Extreme Value Theory
Estimating extreme values is important because they can
be very costly. The challenge is that because they are rare,
many have not even been observed. Thus, it is difficult to
model them.
Extreme value theory (EVT) can be used to model
extreme events in financial markets and to compute VaR
as well as ES.
© Kaplan, Inc. 1
1
Extreme Value Theory (cont.)
Extreme value theorems provide a template for estimating
the parameters used to describe extreme movements.
One approach for estimating parameters is the Fisher-
Tippett theorem. According to this theorem, as the sample
size n gets larger, the distribution of extremes converges to
the generalized extreme value (GEV) distribution.
© Kaplan, Inc. 2
Extreme Value Theory (cont.)
x – μ
–1/ξ
F(X | ξ,μ,σ) = exp – 1+ ξ × if ξ ≠ 0
σ
x – μ
F(X | ξ,μ,σ) = exp –exp if ξ = 0
σ
x – μ
1+ ξ × > 0
σ
© Kaplan, Inc. 3
2
Extreme Value Theory (cont.)
1. If ξ > 0, the GEV becomes a Frechet distribution, and the
tails are “heavy”—as is the case for the t-distribution and
Pareto distributions.
2. If ξ = 0, the GEV becomes a Gumbel distribution, and the
tails are “light”—as is the case for the normal and
lognormal distributions.
3. If ξ < 0, the GEV becomes a Weibull distribution, and the
tails are “lighter” than a normal distribution.
© Kaplan, Inc. 4
Peaks-Over-Threshold Approach
The peaks-over-threshold (POT) approach is an
application of EVT to the distribution of excess losses over
a high threshold.
It requires fewer parameters than approaches based on
extreme value theorems.
It assumes that observations beyond the threshold follow a
generalized Pareto distribution (GPD).
© Kaplan, Inc. 5
3
Generalized Pareto Distribution
The GPD exhibits a curve that dips below the normal
distribution prior to the tail.
It then moves above the normal distribution until it reaches
the extreme tail.
Because all distributions of excess losses converge to the
GPD, it is the natural model for excess losses.
© Kaplan, Inc. 6
Generalized Pareto Distribution (cont.)
When applying the GPD, the researcher must choose a
threshold.
There is a tradeoff because the threshold must be high
enough so that the GPD applies, but it must be low
enough so there are sufficient observations above the
threshold to estimate the parameters.
© Kaplan, Inc. 7
4
Tradeoff in Setting the Threshold
β n
–ξ
VaR = u + (1– confidence level) – 1
ξ Nu
VaR β – ξu
ES = +
1– ξ 1– ξ
© Kaplan, Inc. 8
GARP 2018 Market Risk Q1 (also 2017 Q1)
An investment bank with an active position in commodity futures is using the peaks-over-threshold (POT)
methodology for estimating VaR and ES at the 99% confidence level. The bank’s risk managers have set a
threshold level to evaluate excess losses. The choice of the threshold, they argue, is suitable and consistent
with the finding that 5.00% of the observations are in excess of the threshold value. The risk managers have
concluded that the position’s VaR using the POT measure is 4.45%. The VaR estimate incorporates the
following assumptions generated from the managers’ empirical analysis:
Parameter Symbol Value
Loss threshold u 3
Number of observations N 740
Number of observations that exceed threshold n 37
Scale ß 0.75
Shape (tail index) e 0.22
Given the VaR value and the parameter assumptions, which of the following is correct?
A. Increasing the value of the tail index lowers both the ES and the VaR.
B. Increasing the loss threshold level increases both the ES and the VaR.
C. The value of ES is 4.57%.
D. The value of ES is 5.71%.
© Kaplan, Inc. 9
5
GARP 2018 Q1—Answer
ES calculation
VaR β − εu 4.45 0.75 − 0.22 × 3
ES = + ES = + = 5.82%
1− ε 1− ε 1 − 0.22 1 − 0.22
ES tail index a.k.a. the shape/tail parameter
The higher the value, the fatter the tail, the larger the ES value
ES loss threshold level
Think of this like a significance level (e.g., 5% or 1%)
The lower the threshold, the further out into the distribution tail, the higher the ES value
Tutor Tip: if you knew this last point, there was no need to do the calculation!
Correct answer: B
© Kaplan, Inc. 10
GEV vs. POT
1. GEV uses block max/min.
2. POT requires a choice of threshold.
3. Data may help someone decide which approach is easier
from an estimation perspective.
© Kaplan, Inc. 11
6
Multivariate EVT
Multivariate EVT is important because many extreme values are
dependent on each other, and elliptical distribution analysis and
correlations are not useful in the modeling of extreme values for
multivariate distributions.
Modeling multivariate extremes requires the use of copulas.
Given that more than one random variable is involved, modeling
these extremes can be challenging due to the rarity of multiple
extreme values occurring at the same time.
© Kaplan, Inc. 12
Sample Exam Question
Which of the following statements regarding extreme value
theory (EVT) is false?
a. Traditional approaches for estimating VaR may fail to
properly account for fat tails.
b. In contrast to the historical simulation approach, EVT only
considers the tail behavior of the return distribution.
c. When smoothing the tail of the distribution, EVT ignores
extreme events.
d. EVT is able to model ES, which is sometimes referred to as
conditional VaR.
© Kaplan, Inc. 13
7
Sample Exam Question – Answer
Answer: c
EVT does not ignore extreme events and losses (usually
referred to as outliers) in the data set.
© Kaplan, Inc. 14