To what extent does the diagnosis of multiple grid-box weather types add value in post-processing ensemble rainfall forecasts?

"ecPoint" is a statistical post-processing technique that anticipates sub-grid variability and estimates biases in numerical weather prediction (NWP) model outputs, de facto downscaling such outputs from grid-box to point-scale. Forecasts are post-processed based on the distribution of errors in short-range forecasts versus rain gauge observations in the calibration period. This talk will focus on the branch of the ecPoint family products that post-processes ECMWF ensemble (ENS) rainfall forecasts. ecPoint applies the "remote calibration" approach, whose concept is based on the fact that the physics of rainfall generation is the same around the globe. The "remote calibration" approach then allows the post-processing of forecasts in location X using past observations in different regions, countries, or continents as far as the rainfall-generating weather type (WT) between location X and the remote locations is similar. "Weather types" are identified at grid-box scale and can be defined with different degrees of precision. For example, one can have a single-WT scenario where every grid-box is post-processed with the same error distribution. Or one can have a multiple-WT scenario where n significantly different error distributions are defined using predictors calculated from NWP variables (e.g. the fraction of convective rainfall in the total precipitation forecasts, the speed of steering winds, CAPE, etc.) or location-related variables (e.g. solar radiation, orography, etc.). Each grid-box forecast is then post-processed with the corresponding error distribution. In spite of verification having shown that multi-WT ecPoint-Rainfall (214 WTs derived from five predictors) provides more reliable and skilful rainfall forecasts than raw ENS (for point-verification, especially in case of extremes, e.g. rainfall greater than 50 mm/12h), such a product requires the investment of resources to manually re-calibrate ecPoint-Rainfall for new model cycles. Although a multiple-WT ecPoint-Rainfall delivers better forecasts, a single-WT product would be cheaper to re-calibrate. So, is it worth maintaining a multiple-WT ecPoint-Rainfall? This talk aims to answer this question.