Evaluating Rainfall Forecast Skill in Numerical Weather Prediction Models and the Effects of Bias Correction on the PCJ (Piracicaba-Capivari-Jundiaí) River Basins in Brazil

This study evaluates the performance of two high-resolution precipitation forecasting systems—the SIMEPAR operational forecast (FCST-SIM, 5 km) and the ECMWF Integrated Forecasting System (FCST-ENS, ~11 km)—over the PCJ River Basin, São Paulo, Brazil. Forecasts were validated against local rain gauge observations and the Brazilian Daily Weather Gridded Data (BR-DWGD) dataset to assess their suitability as inputs for hydrological models. Performance was evaluated at three strategic control stations (Atibaia, Valinhos, and Buenópolis) during the 2019–2024 period using categorical verification metrics. FCST-SIM consistently outperformed FCST-ENS, particularly during the first forecast lead times, exhibiting higher event-detection skill across all sub-basins. To improve forecast performance, two bias-correction approaches were investigated: Quantile Delta Mapping (QDM) and dry/wet occurrence correction based on frequentist and Bayesian logistic regression models. QDM generally reduced forecast errors, especially at longer lead times, although improvements in Kling–Gupta Efficiency were limited and varied among basins, seasons, and observational datasets. The occurrence-correction approach produced the largest gains in forecast skill, increasing the Critical Success Index from 0.637 to 0.767 (approximately 20%), but still shows its sensibility to the main characteristics of the forecasting model. Overall, the results demonstrate that high-resolution precipitation forecasts can provide valuable inputs for hydrological forecasting in the PCJ Basin when combined with appropriate bias-correction techniques, while highlighting opportunities for further methodological improvements.