A comparison between raw ensemble output, Bayesian model averaging and logistic regression using ECMWF ensemble precipitation reforecasts

Using a 20-year ECMWF ensemble reforecast data set of total precipitation and a 20-year data set of a dense network of precipitation stations in the Netherlands, a comparison is made between the raw ensemble output and two primary ensemble post-processing methods, namely Bayesian model averaging (BMA) and logistic regression (LR).

A previous study has indicated these methods to be successful in calibrating multi-model ensemble forecasts of precipitation for a single forecast projection, with BMA giving better estimates of the probability of high-precipitation events than LR. However, a more elaborate comparison between these methods has not yet been made, and this study does so for single-model ensemble reforecasts of precipitation, namely from the ECMWF ensemble prediction system (EPS) with a maximum lead time of 10 days and a T255 spectral truncation. The following preliminary conclusions can be drawn. The raw EPS output turns out to be generally well calibrated up to 6 forecast days, if compared to the mean 24-h precipitation sum in 1° x 1° grid boxes. However, if the raw EPS forecasts are used as a proxy for the maximum 24-h precipitation sum in each of these grid boxes, both BMA and LR improve the raw EPS output substantially for at least the first 3 forecast days. It is investigated whether the difference in skill between BMA and LR is statistically significant. Finally, it would be interesting to extend this study to other relevant predictands like wind (gusts) and/or to repeat it for an other EPS.