Assessing Entropy-based Bayesian Model Averaging Method for Probabilistic Precipitation Forecasting

Abstract Bayesian Model Averaging (BMA) is a popular ensemble-based post-processing approach where the weighted average of the individual members is used to generate predictive forecasts. As the BMA formulation is based on the law of total probability, possessing the ensemble of forecasts with mutually exclusive and collectively exhaustive properties is one of the main BMA inherent assumptions. Trying to meet these requirements led to the entropy-based BMA (En-BMA) approach. En-BMA uses the entropy-based selection procedure to construct an ensemble of forecasts with the aforementioned characteristics before the BMA implementation. This study aims at investigating the potential of the En-BMA approach for post-processing precipitation forecasts. Some modifications are proposed to make the method more suitable for precipitation forecasting. Considering the 6-hour accumulated precipitation forecasts with lead times of 6 to 24 hours from seven different models, we evaluate the effects of the proposed modifications and comprehensively compare the probabilistic forecasts, derived from the BMA and the modified En-BMA methods in two different watersheds. The results, in general, indicate the advantage of implementing the proposed modifications in the En-BMA structure for possessing more accurate precipitation forecasts. Moreover, the advantage of the modified En-BMA method over BMA in generating predictive precipitation forecasts is demonstrated based on different performance criteria in both watersheds and all forecasting horizons. These outperforming results of the modified En-BMA are more pronounced for large precipitation values, which are particularly important for hydrologic forecasting.