Extreme events evaluation using CRPS distributions

Verification of ensemble forecasts for extreme events remains a challenging question. The general public as well as the media naturely pay particular attention on extreme events and conclude about the global predictive performance of ensembles, which are often unskillful when they are needed. Ashing classical verification tools to focus on such events can lead to unexpected behaviors. To square up these effects, thresholded and weighted scoring rules have been developed. Most of them use derivations of the Continuous Ranked Probability Score (CRPS). However, some properties of the CRPS for extreme events generate undesirable effects on the quality of verification. Using theoretical arguments and simulation examples, we illustrate some pitfalls of conventional verification tools and propose a different direction to assess ensemble forecasts using extreme value theory, considering proper scores as random variables.