Tricks for improving non-homogeneous regression for probabilistic precipitation forecasts: Perfect predictions, heavy tails, and link functions

Raw ensemble forecasts display large errors in predicting precipitation amounts and its forecast uncertainty, especially in mountainous regions where local effects are often not captured. Therefore, statistical post-processing is typically applied to obtain automatically corrected weather forecasts where precipitation represents one of the most challenging quantities. This study applies the non-homogenous regression framework as a start-of-the-art ensemble post-processing technique to predict a full forecast distribution and improves its forecast performance with three statistical tricks. First of all, a novel split-type approach effectively accounts for perfect ensemble predictions that can occur. Additionally, the statistical model assumes a censored logistic distribution to deal with the heavy tails of precipitation amounts. Finally, the optimization of regression coefficients for the scale parameter is investigated with suitable link-functions. These three refinements are tested for stations in the European Alps for lead-times from +24h to +48h and accumulation periods of 24 and 6 hours. Results highlight an improvement due to a combination of the three statistical tricks against the default post-processing method which does not account for perfect ensemble predictions. Probabilistic forecasts for precipitation amounts as well as the probability of precipitation events could be improved, especially for 6 hour sums.