Gradient boosting machine for enhancing global forecast data products of seasonal precipitation: Hirmand river basin experience

Study region The Hirmand (Helmand) River Basin (HRB) shared by Afghanistan, Iran, and Pakistan. Study focus This study uses daily precipitation data from 20 meteorological stations located in Iran and Afghanistan for the period 2009–2020. The performance of six global seasonal climate forecast products (CMCC, DWD, ECCC, ECMWF, MF, and UKMO) was assessed, and three machine learning–based models were applied for bias correction across the HRB. The models include 1) the Daily Gradient Boosting Machine ( D -GBM), which corrects forecasts on a daily scale, 2) the Monthly GBM (M-GBM), which corrects forecasts on a monthly scale, 3) and the Combined GBM (C-GBM), which integrates all forecast products simultaneously. Model performance was evaluated using PBIAS, ME, RMSE, POD, FAR and CSI. New hydrological insights for the region Raw forecast analysis showed that UKMO provided the most accurate predictions during winter and spring, while DWD and ECMWF performed best in summer and autumn. In contrast, MF and ECCC tended to overestimate precipitation in several seasons. Bias correction substantially improved model accuracy. D -GBM performed best in wetter seasons, while M-GBM was more effective in drier months. Moreover, C-GBM outperformed D -GBM in many cases, especially for daily-scale assessments, demonstrating superior overall performance, stability, and suitability for operational forecasting purposes. The enhanced forecast accuracy achieved through these methods can support more reliable seasonal planning in the HRB, aiding informed water allocation and strengthening transboundary water cooperation.