Chapter 21 Regression in the presence of extremes: extreme value regression
Caution: in a highly developmental stage! See Section 1.1.
The problem with estimating extreme quantiles in the “usual” sense:
Here is a histogram of 100 observations generated from a Student’s t(1) distribution (it’s heavy-tailed):
## `stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
Here are estimates of high and low quantiles, compared to the actual. You can see the discrepency grows quickly. Extreme-low quantiles are too high, whereas extreme-high quantiles are too low.
As a rule of thumb, it’s best to stay below \(\tau=0.95\) or above \(\tau=0.05\). If you really want estimates of these extreme quantiles, you’ll need to turn to Extreme Value Theory to make an assumption on the tail of the distribution of the data. One common approach is to fit a generalized Pareto distribution to the upper portion of the data, after which you can extract high quantiles.