Nature published a study of Microsoft Research-associated Probabilistic Bias Correction (PBC) on October 14, 2026. PBC uses machine learning to adjust probability forecasts from existing weather systems; it is a post-processing framework, not a standalone model. In the study’s week-three and week-four evaluations, it improved forecast skill across most of the grid cells tested for precipitation, temperature and mean sea-level pressure. The Nature study
What PBC does to an existing forecast
PBC takes forecasts from a dynamical, AI or hybrid system and learns from that system’s historical predictions and observations. It applies Debias++ and Persistence++ corrections in parallel, then combines the adjusted probability distributions. The result is a revised forecast distribution—not a new weather model generating predictions on its own. Nature describes the method
What the study measured
The study’s main evaluation covered week three, days 19–25 ahead, and week four, days 26–32. Across a global 1.5° latitude–longitude grid, it evaluated average two-meter temperature over land, total precipitation over land and mean sea-level pressure. The researchers grouped probabilities into five quintiles and used ranked probability skill score (RPSS), compared with a climatological baseline; ERA5 reanalysis served as the reference data. The Nature paper details the evaluation.
Against ECMWF’s operationally debiased dynamical forecast, PBC showed skill gains in these shares of evaluated grid cells:
| Evaluated variable | Lead periods | Grid cells with skill gains versus ECMWF’s operationally debiased dynamical forecast |
| Total precipitation over land | Weeks 3 and 4 | 98% |
| Average 2-m temperature over land | Weeks 3 and 4 | 91–92% |
| Mean sea-level pressure | Weeks 3 and 4 | 89–91% |
These percentages describe how widely skill gains appeared across the evaluated grid. They are not probabilities that a forecast for a particular location will be correct. In a separate 2025 test across variables and lead times, PBC doubled the ECMWF AIFS-SUBS system’s mean RPSS, which was 0.03 before correction. Nature reports both results.
MicroDuet ranked first in the 2025 competition season
The PBC-based MicroDuet ensemble ranked first for every evaluated variable and lead time in the inaugural September–November 2025 season of ECMWF’s AI Weather Quest. The paper reports that it outperformed systems from six operational forecasting centers and entries from 34 teams. That is a result from a defined real-time competition season, alongside the study’s other evaluations. The Nature paper reports the competition results.
Microsoft reported early use in Kenya
Microsoft Research says Kenya Meteorological Service Authority forecasters tested PBC and used it to help prepare weekly forecast updates during the 2026 rainy season. Phillip Okello, the service’s principal meteorologist, told Microsoft Research that PBC improved the team’s forecasts more than other methods it tested across the Greater Horn of Africa. He also identified sharper forecasts of heavy rain and dry spells as a priority. Microsoft Research’s account describes this early use.
Open code and a separate S2S preview
Nature and Microsoft say the PBC implementation and experiment code are open source. Microsoft’s October 2026 account separately described its S2S model, which provides weekly probabilistic guidance up to six weeks ahead, as being in private preview for selected customers. Nature Microsoft Research
