Emphasizing Reliability in Member-by-Member Postprocessing of Temperature Forecasts

Abstract Member-by-member postprocessing (MBMP), a nonparametric method that undertakes bias and dispersion correction on individual ensemble members, has emerged as a promising approach. Traditionally, MBMP variants have relied on regression for bias correction, a technique that does not take into account type-1 conditional bias, i.e., reliability. This study introduces novel approaches to implement MBMP that seek to improve forecast quality by focusing on ensemble reliability rather than accuracy during the bias-correction process. A new evaluation metric is proposed, and an innovative multiobjective combination of metrics is implemented during coefficient estimation. This is tested on daily air temperature forecasts with lead times of 2, 5, and 9 days over 44 watersheds in Quebec, Canada. Results demonstrate that higher ensemble forecast reliability is achieved when it is emphasized during the bias-correction step compared to other MBMP variants.