NatureSource publication:
Loading content…
Subseasonal forecasts, two to six weeks ahead, have long hovered near the climatological average because errors compound, and operational dynamical ensembles and even state-of-the-art probabilistic AI routinely underperform climatology at longer subseasonal leads. Probabilistic bias correction uses machine learning to fix the forecast distribution itself rather than only a single point estimate. It doubles the subseasonal skill of an AI forecasting system and beats the operationally debiased dynamical model for 91% of pressure, 92% of temperature and 98% of precipitation targets; in the 2025 real-time competition its global forecasts ranked first for all variables and lead times.
Subseasonal error no longer has to be accepted as is; it can be learned and corrected. Probabilistic bias correction treats a model's probabilistic forecast as the object to be fixed, adjusting the bias of the forecast distribution directly. The prior operational route compared deterministic predictions against model hindcast quantiles, so correction happened only at the level of point forecasts, and precipitation skill often sat near zero or below. In a global 2016-2024 evaluation, probabilistic bias correction tripled the dynamical ensemble's positive precipitation skill and turned its negative temperature and sea-level-pressure skill positive; against operationally debiased dynamical forecasts it raised temperature and pressure skill by 26-37% and 16-32%, with 98% of precipitation, 91-92% of temperature and 89-91% of pressure grid cells improving.
The correction acts on the centre, spread and shape of the probability distribution, so it can address systematic errors in an ensemble. Point correction can only shift a forecast value; it cannot change how the forecast expresses uncertainty. The framework trains two lightweight models in parallel: Debias++ learns location-, date- and quantile-specific corrections, while Persistence++ regresses the forecast on recent observations and climatology; both are projected onto valid cumulative distribution functions and averaged. It doubled the mean skill score of an AI forecasting system and lifted precipitation skill of the hybrid POET system by 68-98%.
The corrected forecasts are built for operational use and extend to extreme-event warning. Earlier subseasonal extreme-event warning often relied on case-specific post-processing rather than a general tool spanning variables and models. In the inaugural 2025 AI Weather Quest season, MicroDuet, an ensemble of corrected dynamical and hybrid forecasts, ranked first for all variables and lead times, ahead of debiased dynamical forecasts from six operational centres, their multimodel ensemble, three versions of an AI forecasting system and 34 teams; on GDACS flood events it raised severe-flood forecast skill by 86-150%.
The framework is positioned as a post-processing tool that corrects systematic error in an existing model, not a replacement for the model's simulation of atmospheric dynamics. Evaluation covers 2-m temperature, precipitation and mean sea-level pressure on a global 1.5-degree grid at weeks 3 and 4, using ERA5 reanalysis as ground truth. It suits meteorological operations and researchers who want to raise the skill of existing products without rebuilding an ensemble.
Skill gains vary with input model, variable and lead time, and the lowest raw skill, and largest headroom, appears for precipitation and extreme low events. Correction depends on the availability of historical forecasts and observations and must be retrained as input models are upgraded. Extreme-event and flood assessments rest on specific quantile definitions and the GDACS event catalogue, so their conclusions are limited to the reported event sets.