The possibility of future changes in extreme events fosters specific diagnoses of global and regional climate model experiments. The rarity of extremes and the limited period of model simulations, however, impose a statistical uncertainty, which limits the chance to detect a change between simulations for present and future climate. In this study the statistical limits for detecting a change in extremes is quantified in terms of a probability of detection (power of the statistical test). This quantity is evaluated as a function of the rarity of events, the length of simulation period and the magnitude of the change. Two diagnostic methods for extremes are considered: the block maximum method (generalized extreme value distribution), and the peek over threshold method (generalized Pareto distribution). The probability of detection is calculated by means of statistical simulation using pairs of surrogate extreme event samples with a prescribed change.

The results pinpoint to the strong limitation for detecting changes with currently available simulation periods. In the case of the block maximum method using extremes from 30-year model simulations, a doubling/halving of the return period can be detected with a probability of 60% for events with a 10-year return period. However this probability drops to 10% for events with a 50-year return period. The results have implications on the choice of diagnostic method and the definition of extreme event thresholds, and they point to the importance of longer simulation periods or ensemble simulations.