Statistical post-processing of daily ensembles enhances the one-month forecast skill of SEAS5 global precipitation

Subseasonal-to-seasonal precipitation forecasts are highly relevant for a wide range of practical applications and risk management, but their skill decreases substantially at longer lead times. Although statistical post-processing is widely regarded as a key step for improving forecast reliability, the skill of post-processed daily precipitation forecasts up to one month remains poorly quantified at a global scale. Here, we evaluate global daily precipitation forecasts for July from 1981 to 2024 generated by the Seasonal Forecasting System version 5 (SEAS5) of the European Centre for Medium-Range Weather Forecasts (ECMWF). The Bernoulli-Gamma-Gaussian (BGG) model calibrates daily ensemble spreads, and the Schaake Shuffle links ensemble members to construct daily trajectories. The performance of BGG-based post-processing is benchmarked against quantile mapping (QM) under rigorous leave-one-year-out cross-validation. Raw SEAS5 forecasts of daily precipitation often exhibit negative skill and underperform climatological forecasts owing to systematic biases and the decay of skill with lead times. These daily biases further compound over lead time, severely eroding the skill of accumulated totals. Statistical post-processing effectively corrects systematic biases, aligns ensemble forecasts with observed distributions, and prevents the accumulation of forecast errors. Relative to QM, the BGG further improves forecast skill by enhancing the reliability of ensemble spreads. The post-processing typically extends the lead time of positive skill to about 7 days for daily amounts and beyond 20 days for accumulated totals, with regional variations and seasonal dependence. Overall, this global evaluation provides practical guidance for post-processing daily precipitation forecasts and serves as a valuable reference for their use in water and agricultural management.