Chapter 13. Extreme Value Theory

Natural Catastrophes
Mad-made Catastrophes
Financial Events (Market Crashes)

Uses:
Pricing
Reserving
Capital Modelling

Model Types:

Statistical/Actuarial:
Using past data, most relevant, can adjust past data for trends, closed form simple solutions
Based on past data, current prejudice without much critical thinking, data may not be complete or sufficient

Physical Tests in a Lab:
Most objective, actual situation, carried out by experts
Only natural catastrophes, results need to be translated to insurance contracts, difficult and expensive to commission

Simulation/Cat Models:
Combined stats and physical, can build up models of extreme events
Requires computing power and is expensive, not closed form (black box), difficult to explain

EVT has two main results:

Distribution of a maxima to determine MPL

Distribution above a set value to price XOL contracts

Normally want to ignore the outliers, but here we want to ignore the main losses and target the tail

Generalised Extreme Value (GEV):
Suitable for finding a maxima and the distribution around it

Adjust for exposure
Possibly non-stationary (does not revert to mean over time)
Paramterise using method of moments (MoM) or max likelihood estimator (MLE)

Generalised Pareto Distribution (GPD):
Used more often than GEV as we are more interested in the distribution than predicting a maxima
Used to model claims over some threshold

Better approximation for high threshold, but less data as it increases
Usually between 90-95 percentile of losses
Scale is like your mean, shape is like the variance

Paramaterise:
Need the probability of exceeding the threshold
Determining threshold requires judgment
Use MLE or MoM to determine shape and scale
Can also use Probability-Weighted Moments
Use the Hill Estimator to determine the threshold u
Use Goodness of Fit to determine if the distribution is correct

Frequency:

Poisson:
Frequency of events per unit time - using data per unit time (normally per unit exposure and predicted that way)
Good where threshold is high and probability is low

Uses:
Catastrophe Modelling - alongside traditional models for a broader spectrum
Reinsurance High Layer Pricing
Capital Modelling - to generate large and cat losses in our DFA
IBNR Reserving
IBNER Reserving

Considerations:

Allow for IBNER in losses
Allow for IBNR in loss counts
Claims Inflation - increases size and frequency
Adjust for Exposure (like premium) - depending on how the exposure measure changes - eg. double the policies or just double the risk of the same policies

Pros:

Mathematical view on quant measure of risk

Describes extremes

Tail is the point of interest, not the centre

Reasonable formula for tail from a priori effects

View to deal with skewness, fat tails, rare events and stress scenarios

Cons:

Difficult to make assumptions (scale and shape, threshold)

Setting threshold is difficult, not easy choice

Assume Independent, Identical

Difficult to explain

Limitations of EVT in General Insurance:

Frequency not Poisson (if variance > mean then using negative binomial)

Fitting to extrapolated data:
Severity - should account for uncertainty and volatility in our IBNER's
Frequency - same as above but for IBNR
Can model for both the IBNER and IBNR

No cap on claims size in EVT

Parameter Error - can increase volatility

Uncertain future inflation - can model inflation

GPD not always best there are other severity models. Use Goodness of Fit

Losses are not Independent and Identically Distributed
Different policy characteristics, details and claims change over time
Severity Correlations (eg. inflation)
Frequency Correlations (eg. legislation changes, economic conditions)