Chapter 20: Extreme value theory Flashcards
Why tail events may be incorrectly modelled STICO
• Statistical Distribution of extreme data may differ from smaller loss data
• True data is more skewed and leptokurtic
• Influence of bulk of data on parameter estimates
• Changes in data features over time (heteroskedasticity and structural breaks)
• Origin of the event differs from standard event
Description of GEV BITSS
Block maxima sample X_m=max(X_1,X_2,…,X_n)
IID assumed
Tend to a particular distribution with an increase in observations
Standardized CDFs exist to solve for
Shape parameter is gamma
Weibull, γ<0 RUN
o Reinsured losses
o Upper bound
o Natural events may it
EVT Gumbel, γ=0 FEN
o Fat tails
o Exponentially falling
o No Bounds
Fréchet γ>0 HPF
o Heavy tails with increased variance
o Power laws in the tails
o Financial loss events
Fitting a GEV distribution SMDC MET
- Subdivide data into blocks
- Maximum for each block calculated
a. Max observation in each block (return level)
b. Exceeding some set level (return period)
c. Trade off between information and variance - Distribution selection based on data behaviour
- Calculate parameters of GEV distribution with method of moments or MLE
Character of GPD AM PULE
Asymptotic property – excess distribution of standardised maxima converges to GPD
Mean excess loss function used for fitting
Pareto GPD: If gamma > 0
Upper bound when γ<0
Lower bound γ>= 0
Exponential GPD: If gamma = 0
Choosing a threshold for the distribution LASS BRS
Linearity between mean excess loss
function and threshold – lower point chosen
Asymptotic properties applied to increase accuracy
Slope of mean excess loss functions determines value of the shape parameter γ
Subjectivity introduced in selecting the slope level
Bias to quality trade-off
Reliability of the data
Scanty data
Parameterisation of the GDP function MIRS
• MLE of MOM can be used
• IID assumed for all observations
• Random amount of variables exceed threshold
• Shortcomings of the approach should be considered