Statistical post-processing of weather forecast ensembles: obtaining optimal deterministic and probabilistic predictions at multiple time scales

Weather forecasts are produced by complex numerical models, issued to end users and then updated after a certain period of time, usually at least several hours. During this time, it might become obvious that the current forecasts are somehow flawed and of little use. Nonetheless, they are not changed until being replaced by a new batch from the most recent run of the model. This work proposes a new statistical post-processing method, Rapid Adjustment of Forecast Trajectories, that improves the quality of predictions even after they have been issued and thus increases their potential value to customers. The inherent correlation between errors at different forecast times allows for adjustments being applied to future predictions based on very recent observations. Thus, both fast-developing and systematic forecast errors can be corrected in a flexible and swift manner. It complements other, conventional statistical post-processing and results in a significant gain in forecast quality. This novel technique can be applied to any forecast time range, from a few hours to several days and weeks, while being very economical and versatile.