Uncertainty of Rainfall Forecasts for Impact-Based Flood Warning in the Cagayan River Basin, Philippines

Abstract. Effective impact-based flood early warning requires not only information on when heavy rainfall may occur, but also a calibrated estimate of the uncertainty in the resulting impacts. This study develops a framework for translating multi-centre sub-seasonal to seasonal (S2S) rainfall forecasts into probabilistic, municipality-level impact-based early-warning information for staged flood preparedness in the Cagayan River Basin, Philippines. Daily Gamma-kernel Bayesian model averaging is first applied to ensemble forecasts from ECMWF, NCEP, and UKMO to generate continuous predictive distributions of basin-mean rainfall. Leave-one-year-out verification for 2015–2025 identifies the ECMWF+NCEP+UKMO combination as the most robust tested input, with useful daily rainfall information mainly retained up to approximately lead days 5–6. The daily predictive samples are then accumulated into rolling seven-day rainfall distributions, because flood impacts in the basin are more closely related to multi-day rainfall than to isolated daily totals. Threshold-based probability recalibration improves the reliability of seven-day exceedance probabilities, raising Brier Skill Scores from −0.02 to +0.08 at 100 mm per seven days and from −0.07 to +0.02 at 150 mm per seven days relative to a monthly climatological baseline. The recalibrated seven-day rainfall distributions are subsequently propagated through municipality-level rainfall–damage functions to estimate probabilistic impacts on affected population, building damage, rice damage, and maize damage. An application to Typhoon Ulysses in November 2020 demonstrates how forecast-state-dependent impact intervals evolve as the event approaches and how municipalities with potentially large impacts can be prioritised. The results show that calibrated S2S rainfall probabilities can support uncertainty-aware, impact-based flood preparedness, while also highlighting limitations related to lead-time skill, basin-mean rainfall representation, upper-tail rainfall coverage, and the validation of rainfall–damage functions.