Know these first
- L3.7.07: Measuring frequency and severity
Math Atoms
By the end you will be able to
- 1Model every loss above a high threshold with the generalised Pareto distribution, and say how it relates to the GEV of block maxima.
- 2Choose a threshold with the mean excess plot, a stability plot and the Hill estimator, and explain the trade-off.
- 3Compute VaR and ES at 99% and 99.9% from a fitted tail, with a standard error, and compare them with empirical, normal and Student t figures.
- 4Judge what a tail fitted to a few large losses can say, on the bank's operational losses and the insurer's windstorms.
Why it matters
L6.1.06: Extreme value theory: block maxima threw away all but one loss a month. In a calm month the worst day is ordinary; in October 2008 the second and third worst days were bigger than most months' worst. Peaks over threshold keeps every large loss instead. On GBP/USD that is 256 days above 0.92%, enough to read a 99.9% VaR to within about 0.6% either way. On the bank's operational losses it is far fewer, and L3.7.07: Measuring frequency and severity's question comes back sharper: what can a handful of very large losses tell you about the next one?