Assessing Predictive Distributions: A Diagnostic Approach

 

Tilmann Gneiting and Adrian E. Raftery

 

University of Washington

Department of Statistics

 

This talk discusses verification measures for continuous probabilistic forecasts that take the form of a probability distribution function (pdf) or predictive distribution. A diagnostic approach is proposed that addresses both calibration (reliability) and refinement (variability) through plots and summary measures. The notions of probabilistic calibration and exceedance calibration are introduced, and special emphasis is put on ramifications of the rank histogram and multicategory reliability diagram. The various diagnostic tools are applied to assess and compare methods of forming predictive distributions for wind speed based on the University of Washington MM5 mesoscale short-range ensemble prediction system.