An Intuitive Curve-Fit Approach to Probability-Preserving Prediction of Extremes

Abstract

A method is described for predicting extremes values beyond the span of historical data. The method - based on extending a curve fitted to a location- and scale-invariant variation of the double-logarithmic QQ-plot - is simple and intuitive, yet it preserves probability to a good approximation. The procedure is developed on the Generalised Pareto Distribution (GPD), but is applicable to the upper order statistics of a wide class of distributions.