Predicting East Africa's rainfall extremes with calibrated AI post-processing systems

Abstract Accurate and reliable rainfall forecasts are crucial for Early Warning Systems (EWS). Artificial Intelligence approaches, which post-process global physical forecasts, enable low-cost ensemble generation, extending skill to medium-range lead times and offering scalable systems to inform EWS particularly under resource-constrained settings. However, assessing their practical value in representing user-relevant rainfall extremes and guiding decisions for EWS remains critical. We present a low-cost, uncertainty quantification method that calibrates ensemble forecasts to probabilities of exceedances for rainfall extremes. Applied to raw and post-processed rainfall forecasts over East Africa, sub-regional analysis shows the threshold-calibrated, post-processed forecasts better reflecting the observed chances of extremes at higher rainfall thresholds. Case-study results show encouraging added value from post-processed forecasts at longer lead times. Forecast skill decomposed across rainfall and probability thresholds highlights the ranges of rainfall and risk tolerances where most actionable improvement arises, whether using satellite or rain-gauge data as reference. This tailored evaluation helps guide forecast adoption for EWS.