Core Path · L6.1.06

Extreme value theory: block maxima

14 minExtreme value theoryGEVBlock maximaReturn levelsProfile likelihood

By the end you will be able to

  1. 1Explain why the largest value in a block follows a GEV distribution whatever the daily distribution, and what the shape ξ says about the tail.
  2. 2Fit a GEV to monthly, quarterly and yearly maxima of real losses by maximum likelihood (M-D08: Maximum likelihood estimation), with standard errors.
  3. 3Compute a return level for T years and its 95% interval by the delta method and by profile likelihood, and say why they differ.
  4. 4Show how the estimate moves with the block length and with a single extreme block, and choose blocks with that trade-off in mind.

Why it matters

On 24 June 2016 GBP/USD's fixing fell 8.16%. How rare was that? A normal fitted to all days says practically never; L6.1.04: GARCH models's GARCH-t calls it rare but possible. Extreme value theory gives a third answer by ignoring everything except the extremes: fit the worst day of each month and it puts that loss at about once in 590 years. Fit the worst day of each year and it says once in 47. Both fits are honest. This lesson is about why they differ, and what that means for a 1-in-200-year number.