Probabilistic short-range forecasts of high-precipitation events: optimal decision thresholds and predictability limits

Abstract

Translation of ensemble predictions into high-precipitation warnings is assessed using user-oriented metrics. Short-range probabilistic forecasts are derived from an operational ensemble prediction system using neighbourhood postprocessing and conversion into categorical predictions by decision threshold optimization. Forecast skill is modelled for two different types of users. We investigate the balance between false alarms and missed events and the implications of the scales at which forecast information is communicated. We propose an ensemble-based deterministic forecasting procedure that can be optimized with respect to spatial scale and a frequency ratio between false alarms and missed events. Results show that ensemble predictions objectively outperform the corresponding deterministic control forecasts at low precipitation intensities when an optimal probability threshold is used. The optimal threshold depends on the choice of forecast performance metric, and the superiority of the ensemble prediction over the deterministic control is more apparent at higher precipitation intensities. Thresholds estimated from a short forecast archive are robust with respect to forecast range and season and can be extrapolated for extreme values to estimate severe-weather guidance. Numerical weather forecast value is found to be limited: the highest usable precipitation intensities have return periods of a few years only, with resolution limited to several tens of kilometres. Implied precipitation warnings fall short of common skill requirements for high-impact weather, confirming the importance of human expertise, nowcasting information, and the potential of machine learning approaches. The verification methodology presented here provides a benchmark for high-precipitation forecasts, based on metrics that are relatively easy to compute and explain to non-experts.