Scoring rule inference for spatial statistics based on cross-validation

Abstract: Although scoring rules are traditionally aimed at model evaluation, they have also successfully been used for model inference. We propose parameter inference of spatial models through a leave-one-out cross-validation approach (LOOS), where the predictive ability is optimised instead of the likelihood. The approach is studied for different Gaussian spatial models. For Gaussian models with sparse precision matrices, such as spatial Markov models, the approach results in fast computations compared to the likelihood approach. Moreover, the approach allows affecting the robustness to outliers and sensitivity to non-stationarity. Applying the LOOS to ERA5 temperature reanalysis data for the contiguous United States and the average July temperature for the years 1940 to 2023 resulted in estimates with better predictive performance than the maximum likelihood in a fraction of the computation time.

machine learningprobabilitystatistics theory

Audience: researchers in the discipline

Gothenburg statistics seminar

Series comments: Gothenburg statistics seminar is open to the interested public, everybody is welcome. It usually takes place in MVL14 (http://maps.chalmers.se/#05137ad7-4d34-45e2-9d14-7f970517e2b60, see specific talk). Speakers are asked to prepare material for 35 minutes excluding questions from the audience.