A1345
Title: Prediction limits based on weighted model combinations and CRPS scoring rules
Authors: Paolo Vidoni - University of Udine (Italy) [presenting]
Valentina Mameli - University of Udine (Italy)
Abstract: The purpose is to define prediction limits based on predictive distributions obtained as weighted combinations of elementary distribution functions. In particular, linear and multiplicative combinations of density functions, as well as linear combinations of quantile functions, are considered. A combined model can serve as a useful surrogate for the true, unknown predictive model of the random phenomenon of interest. The weights associated with the individual components are determined by considering the continuous ranked probability score (CRPS) and its weighted extensions. Simulation studies show that, by using an appropriately weighted version of the CRPS that focuses on the tails of the distribution, the estimated combined model provides prediction limits with coverage probabilities close to the target nominal value. The good performance of the approach is further illustrated using real data on athletics' records.
