A study published in Nature on October 14 presents Probabilistic Bias Correction, a method that learns a weather system’s systematic errors from historical forecasts and observations. It aims to improve probabilities over the difficult two-to-six-week horizon, not promise the exact weather on a particular day six weeks ahead.
Applied to ECMWF systems, PBC doubles what the authors describe as the AI Forecasting System’s still-modest subseasonal skill. For the already-debiased dynamical model, the study reports improvement for 91% of pressure targets, 92% of temperature targets and 98% of precipitation targets. These shares of improved targets are not percentages of forecasts that became correct.
MicroDuet combines PBC corrections and ranked first across the evaluated variables and lead times in the inaugural 2025 AI Weather Quest season. This result belongs to a specific probabilistic evaluation; doubling a low initial score neither doubles every weather capability nor guarantees each extreme event. Microsoft’s storm examples do not replace that evaluation.
PBC code is public. Microsoft describes testing with Kenyan forecasters and an S2S model in private preview for selected customers, distinct from general commercial availability. For research teams, the useful path is to compare the correction against existing calibration methods on periods and regions genuinely separated from training.