Microsoft PBC: Open-Source Bias Correction That Improves 2-to-6-Week Weather Forecasts

Microsoft Research and MIT have released probabilistic bias correction (PBC), a machine learning post-processor that makes existing weather models better at forecasting two to six weeks ahead, and published the work in Nature on October 14, 2026. The code is open source. The Verge reported the release as software that "could make AI-enhanced weather forecasting more affordable", and a detailed summary of the paper lays out the method and numbers.

This post explains what PBC does, what the reported results are, why it is a different kind of AI weather story from the model launches we usually cover, and where to be careful. For comparison with model-side advances, see our coverage of DeepMind's WeatherNext 3 hourly 5 km forecasts.

Quick reference

Why weeks 3 to 6 are hard

Modern physics-based and AI systems forecast accurately out to roughly two weeks. Beyond that, according to the paper's summary, skill "drops precipitously" at subseasonal timescales because of compounding errors, systematic model biases and the chaotic atmosphere. Yet this window matters most for farmers, energy planners and disaster agencies who need to prepare weeks ahead.

PBC attacks one component: systematic bias. If a model is consistently too warm in a region in a given season, that error is predictable and can be learned from the model's own history. The authors describe PBC as targeting predictable errors rather than trying to relearn atmospheric dynamics.

How PBC works

Based on the paper summary:

1. Convert a raw ensemble of deterministic forecasts into an initial probabilistic forecast measured against climatological quantile bins.
2. Apply two learned corrections in parallel. Debias++ adjusts the forecast with location-, date- and quantile-specific corrections learned by minimizing probabilistic forecasting error over adaptively chosen training periods. Persistence++ blends the forecast with lagged observations and climatology to capture recent trends and day-of-year effects.
3. Project both corrections onto valid cumulative distribution functions and average them into the final prediction.

It operates on probability distributions, not raw observations, and can be retrained incrementally each day using only data available at each forecast issuance date, which the authors say mimics operational use.

The evaluation followed ECMWF's AI Weather Quest conventions: 2-meter temperature, total precipitation over land and mean sea-level pressure on a global 1.5-degree grid, five quintile bins, at week-3 and week-4 lead times, scored with the ranked probability skill score (RPSS) against climatology. Ground truth was ERA5 reanalysis, and the default test set covered Monday and Friday target dates in 2016 to 2024.

The reported results

All numbers below are the authors' reported figures as summarized in the coverage we read; we have not reproduced them.

Note the baselines: very low absolute skill. ECMWF's raw precipitation RPSS of 0.023 and AIFS-SUBS's 0.03 are small numbers, so "tripled" and "doubled" are relative gains from a low base. They still matter because subseasonal skill is notoriously hard to earn, but they are not the same as "forecasts are now accurate".

MicroDuet in the AI Weather Quest

The team entered MicroDuet, an ensemble of PBC-ECMWF and PBC-POET, in ECMWF's 2025 AI Weather Quest. The paper reports it placed first for all variables and lead times in the inaugural September to November season, ahead of debiased dynamical models from six operational centers (ECMWF, NOAA, ECCC, HMCR, KMA, JMA), a six-model multimodel ensemble, three versions of AIFS-SUBS and 34 other teams, and led both lead times again in December to February. It was the only submission to significantly outperform debiased ECMWF (P below 0.05) across all variables and lead times. The summary also notes ECMWF's own March 13, 2026 blog describing 42 teams and 74 models across the first two periods, with MicroDuet leading.

One real-time example: in November 2025, MicroDuet's week-4 temperature forecast anticipated a cold air outbreak in the eastern US in December 2025 that debiased ECMWF missed. A single hit is an anecdote, not a statistic.

What people are asking

Is this another AI weather model? No. Models like WeatherNext or AIFS generate forecasts. PBC takes a model's forecast ensemble and its past performance and outputs a corrected probability forecast. Think of it as calibration software: it learns where a model is habitually wrong, by place, date and quantile, and nudges the probabilities.

Why probabilities and quantile bins? Subseasonal forecasts cannot say "it will rain on day 24". They say how likely the week is to be wetter or drier than normal. Scoring against five quintile bins with RPSS rewards well-calibrated odds instead of point guesses, which is the honest form of a long-range forecast.

What does "debiased ECMWF" mean as a baseline? It is ECMWF's operational output after its own bias adjustment, so beating it is a stronger claim than beating raw model output. The paper's headline gains versus that baseline (26 to 37% for temperature RPSS) are the more meaningful figures, while the "tripled" claim is versus raw output.

Who benefits first? National meteorological services without supercomputer budgets, and anyone running a forecast model who has years of archived runs. Energy traders and agricultural planners consume the outputs, but the tool is built for forecast producers.

How should I read the competition result? The AI Weather Quest is run by ECMWF with real-time submissions, which is harder to game than a retrospective test. Leading two seasons is a meaningful signal, though two seasons is a short record and the team's own paper is the source of the details.

Why this matters

• Post-processing is cheap leverage. Better models cost supercomputers and GPUs; correcting their bias costs far less. Coverage says agencies can run PBC on CPUs. That makes it relevant to resource-constrained meteorological centers, which the authors explicitly highlight.
• It complements, not competes with, big weather models. PBC improved both ECMWF's physics ensemble and AI systems. Anyone shipping a forecast model, including Google's WeatherNext, could in principle apply a correction layer on top.
• Extreme events and floods. The authors say gains translate into better prediction of extreme events with potential for agricultural planning, energy management and disaster preparedness.
• Open code lowers the bar. An MIT-licensed repository with replication code lets researchers check the claims.

The geospatial agent story we covered today is a different layer of the same ecosystem: Google's Earth AI Geospatial Reasoning agent calls forecast models as tools, while PBC improves the forecasts those tools return.

Caveats

• Funding and conflicts. The paper discloses that several authors are Microsoft employees or own Microsoft stock. ECMWF authors are also on the paper.
• Low base skill. Relative gains from tiny baselines overstate practical usefulness.
• Benchmark conditions. Results are on a 1.5-degree global grid and specific target dates; local, high-resolution use may differ.
• Secondary reporting. Our numbers come from a summary of the paper and The Verge's report; the Verge page we could fetch showed only the headline and standfirst. Check the Nature paper and the repository before citing precise figures.
• Not a complete weather solution. It corrects bias; it does not fix chaos. Probabilistic weekly outlooks remain uncertain.

How to try it

Microsoft Research maintains a Subseasonal Climate Forecasting project page describing its work on machine learning for forecasts two to six weeks out. The PBC repository is reported to be tested on Rocky Linux 8.10 with Python 3.13.3. A sensible first experiment for a team with forecast archives: take your model's historical ensemble forecasts, apply PBC with the same quantile-bin convention, and compare RPSS against your current debiasing on a held-out year.

Related reading

• HiRO-ACEv2: Ai2's 3 km extreme-weather emulator, a climate-emulation companion released the same day

• WeatherNext 3: hourly 5 km global forecasts

• DeepMind WeatherNext cyclone forecasting

• Google Earth AI geospatial agent

• Planetary Prediction Engine

• Google DeepMind offers 200 European researchers WeatherNext and credits

• Opus 5.5 agents screening magnetic semiconductor candidates

• Official: Microsoft Research subseasonal forecasting project

Details reflect coverage of the Nature paper as of October 14, 2026.