Tail Skill and Ensemble-Configuration Sensitivity in Six-Hour Precipitation Postprocessing

Pengxin Lin, Kaicheng Huang, Wanfeng Ding 

Statistical postprocessing supports probabilistic precipitation forecasting. Comparing methods requires separating forecast performance from the effects of case selection, ensemble configuration, and scoring convention. We develop a paired diagnostic workflow that separates observation conditioning from aggregation, decomposes model and raw-ensemble responses to member reduction, and examines exceedance reliability. We apply it to semilocal censored shifted gamma regression (M2) on EUPPBench, using 1997–2016 training data and 2017–2018 forecasts at 116 European stations. At +24 h, the primary difference in threshold-weighted continuous ranked probability score between M2 and the adjusted raw ensemble is -0.002656 mm, with no statistically significant difference under paired bootstrap inference. The secondary full-distribution score improves by 5.91% against the same reference. Conditioning on observed active dates shifts the tail contrast by -0.005076 mm, compared with -0.000877 mm from subsequent equal-episode weighting. Reducing inputs to control plus ten members worsens the fixed model's tail score, with the comparative shift changing sign between empirical and adjusted raw-ensemble scoring. Reliability analysis locates underprediction in the 0.4–0.8 probability bins despite lower mean Brier scores. The diagnostics also identify interannual instability and member-configuration sensitivity at +48 h, linking aggregate scores to distinct evaluation effects and calibration errors.