B1145
Title: Forecast evaluation of extremes using locally tail-scale invariant scoring rules
Authors: Helga Kristin Olafsdottir - Chalmers University of Technology and University of Gothenburg (Sweden) [presenting]
David Bolin - King Abdullah University of Science and Technology (KAUST) (Saudi Arabia)
Holger Rootzen - Chalmers (Sweden)
Abstract: Statistical analysis of extremes can be used to predict the risk of occurrence of future extreme events, such as large rainfalls or devastating windstorms. Averages of proper scoring rules are commonly used to compare the quality of different probabilistic forecasts. The choice of scoring rules can affect the forecast rankings since the scoring rules reward different features of the forecast. When predicting environmental extremes in a spatial area where the scale of the events varies, one might want to put equal importance on the forecasts at different locations regardless of differences in the prediction uncertainty. Scores possessing this ability are said to be locally scale invariant. For extremes, this can be an unnecessarily strict requirement. Instead, local tail-scale invariance is proposed where the scoring rules are locally scale invariant for large events. A scaled version of the weighted continuous ranked probability score is developed and studied. This score is locally tail-scale invariant and a suitable alternative for scoring extreme value models over areas with varying scales of extreme events. It is shown through both simulations and applications on extreme rainfall models.
