A Microsoft Research-led team published a study in Nature on October 14, 2026, introducing probabilistic bias correction (PBC), a machine learning framework the authors report improves weather forecasts two to six weeks ahead, alongside an open-source release of the code.
The study states that although modern dynamical and AI systems deliver accurate forecasts out to roughly two weeks, skill drops precipitously at subseasonal timescales of two to six weeks ahead because of compounding errors, systemic model biases and the chaotic nature of the atmosphere. The authors present PBC as a framework that reduces systematic error by learning to correct a model’s historical probabilistic forecasts, operating directly on probability distributions rather than raw observations.
Hannah Guan and Soukayna Mouatadid contributed equally to the work, and Guan performed her research as a Microsoft Research intern; Judah Cohen of MIT and Lester Mackey of Microsoft Research are the corresponding authors. The 13-author team lists affiliations with Microsoft Research, Microsoft, MIT, Harvard College, the University of Toronto, IMPA, Janus Research Group, Rhiza Research and ECMWF. Nature records the paper as received on April 17, 2026 and accepted on September 4, 2026.
How PBC Works and How It Was Evaluated
Given a raw ensemble of deterministic forecasts, PBC first converts the ensemble into an initial probabilistic forecast measured against climatological quantile bins. Two machine learning models then apply complementary corrections in parallel: Debias++ perturbs the forecast with location-, date- and quantile-specific adjustments learned by minimizing probabilistic forecasting error over adaptively selected training periods, while Persistence++ blends the forecast with lagged observations and climatology to capture recent weather trends and day-of-year effects. The two corrections are projected onto the space of valid cumulative distribution functions and averaged into the final PBC prediction.
The team evaluated forecasts using the conventions of ECMWF’s AI Weather Quest competition: average 2-meter temperature and total precipitation over land plus average mean sea-level pressure on a global 1.5-degree grid with five quintile bins, at week-3 (days 19–25) and week-4 (days 26–32) lead times, scored with the ranked probability skill score (RPSS) against a climatological baseline. Ground truth was 1979–2026 ERA5 reanalysis regridded to 1.5-degree resolution, and the default test set covered non-leap-day Monday and Friday target dates in 2016–2024. The authors state PBC is lightweight enough to be retrained incrementally each day using only the reanalysis and model forecast data observable at each issuance date, mimicking operational deployment.
Skill Gains Across Dynamical and AI Models
The paper describes ECMWF’s dynamical ensemble as widely considered the most accurate weather model in the world and the gold standard for benchmarking subseasonal skill. Applied to that ensemble over 2016–2024, the study reports that PBC tripled the raw model’s positive precipitation skill of 0.023 RPSS and converted its negative-skill temperature and pressure forecasts into skillful predictions. Against ECMWF’s operationally debiased output, the authors report that PBC raised temperature RPSS by 26–37% and pressure RPSS by 16–32%, with gains in 98% of grid cells for precipitation, 91–92% for temperature and 89–91% for sea-level pressure.
The study reports that PBC also improved AI-based systems. Applied to AIFS-SUBS, ECMWF’s deep learning model for subseasonal prediction, PBC doubled the model’s mean RPSS of 0.03 across all variables and lead times over the 2025 test year. Applied to POET, a hybrid system that post-processes the ECMWF ensemble with convolutional and transformer neural networks, PBC added 2024 gains of 68–98% for precipitation, 15–23% for mean sea-level pressure and 5–7% for temperature, according to the paper. The authors further report that PBC-ECMWF showed pronounced skill gains over the FuXi-S2S deep learning model across the 2017–2021 test years, with 99–100% of grid cells improving for temperature, 97–99% for precipitation and 96–100% for mean sea-level pressure.
Competition Results and Extreme-Event Forecasts
To benchmark real-time performance, the team entered an ensemble of PBC-ECMWF and PBC-POET named MicroDuet into ECMWF’s 2025 AI Weather Quest competition. The paper reports that MicroDuet placed first for all weather variables and lead times in the inaugural September–November season, ahead of the debiased dynamical models of six operational forecasting centers (ECMWF, NOAA, ECCC, HMCR, KMA and JMA), a six-model dynamical multimodel ensemble, three versions of AIFS-SUBS and the forecasting systems of 34 teams worldwide, and that it also led both lead times in the December–February season. Across both seasons, the authors report, MicroDuet was the only submission to significantly outperform debiased ECMWF (P
ECMWF described the same competition in a March 13, 2026 science blog, stating that 42 teams submitted forecasts under 74 uniquely named models across the first two competitive periods and that the MicroEnsemble team’s post-processing model MicroDuet consistently led the leaderboard and significantly outperformed ECMWF’s Integrated Forecasting System for all variables and lead times.
The paper also reports that in November 2025, MicroDuet’s real-time week-4 temperature forecast anticipated a cold air outbreak in the eastern United States in December 2025 that debiased ECMWF missed.
For extreme-event forecasting, the authors trained PBC on semidecile (5% quantile) increments, defining extreme highs as events verifying above the 95th percentile and extreme lows as those below the 5th percentile, and measured skill with the Brier skill score. For week-3 extreme highs over 2016–2024, the study reports PBC boosted operationally debiased ECMWF skill by 51% for temperature and 39% for pressure, with grid-cell gains of 96% for precipitation, 89% for sea-level pressure and 85% for temperature.
Across 2,435 flood events catalogued by the Global Disaster Alert and Coordination System (GDACS) between 2016 and 2026, the authors report PBC boosted week-3 flood forecast skill by 76–92% while operational debiasing reduced raw ECMWF’s baseline skill of 0.04–0.06 toward near zero. For 88 severe floods, which GDACS defines as events causing more than 100 deaths or 80,000 displacements, the paper reports PBC enhanced severe flood forecasting skill by 86–150%. The authors state that these probabilistic skill gains translate into more accurate prediction of extreme events and have the potential to improve agricultural planning, energy management and disaster preparedness in vulnerable communities.
Open-Source Code and Research Context
The PBC code is available on GitHub under an MIT license and archived on Zenodo, with replication code for the study; the repository states the Python codebase was tested on Rocky Linux 8.10 with Python 3.13.3. An authors’ preprint was submitted to arXiv on April 17, 2026 and last revised on September 4, 2026, and Microsoft Research’s publication page dates the work to April 2026.
In the paper’s discussion, the team describes PBC as targeting predictable errors rather than attempting to relearn atmospheric dynamics, framing it as a post-processing tool rather than a replacement for input model improvement and as a low-overhead strategy for meteorological centers, including those in resource-constrained settings. Funding disclosures list FAPERJ support for Orenstein, a US National Science Foundation grant for Cohen, and EU Destination Earth funding for the ECMWF authors; the competing-interests declaration states that several authors are Microsoft employees or own Microsoft stock.
Credit: Source link
