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

Precipitation forecasts from numerical weather prediction systems can contain systematic errors in both occurrence and magnitude, limiting their usefulness for hydrological applications. This study evaluated two operational precipitation forecasting systems against two observational reference datasets across three basins in São Paulo State, Brazil, and assessed statistical post-processing for dry/wet occurrence and precipitation magnitude. Four logistic-regression-based methods were evaluated for occurrence correction, while Quantile Delta Mapping (QDM) was applied to precipitation amounts. Occurrence correction was assessed using POD, FAR, CSI, and ACC, with dry days defined as the event. Changes in individual classifications were evaluated using the exact McNemar test, while changes in categorical metrics across seven lead times were assessed using exact paired permutation tests and bootstrap 95% confidence intervals. A descriptive multi-metric ranking was used to compare correction methods. QDM was evaluated using RMSE skill score and KGE. Occurrence correction modified categorical performance, with effects depending on forecast–observation pairing, basin, and lead time. The McNemar test identified significant classification changes after Holm adjustment in some configurations, whereas the permutation tests did not provide evidence of systematic metric improvement. HBLR-AR1 showed the most balanced overall performance, whereas LR-Seasonal was the least consistent method. QDM improved RMSE skill, particularly at longer lead times, but did not consistently improve KGE. Overall, correction effectiveness depended on forecast–observation discrepancies.