MIT and Microsoft Researchers Use Machine Learning to Improve Weather Forecasts Two to Six Weeks Out

Researchers from Microsoft Research, MIT and other institutions have developed a machine-learning technique that they say makes weather forecasts more accurate two to six weeks ahead, a range where forecasts have long been weak. The work was published in Nature on 14 October 2026, and the code has been released as open source.

The problem

Today’s weather models are accurate up to about two weeks out. Beyond that, the authors write in their Microsoft Research abstract, “model skill drops precipitously at subseasonal timescales (2–6 weeks ahead)”. These longer forecasts matter for farming, water and energy planning, and disaster preparation.

What the researchers built

The team calls its method probabilistic bias correction (PBC). In the paper’s words, it is “a machine learning framework that substantially reduces systematic error by learning to correct historical probabilistic forecasts.” Rather than replacing existing forecasting models, it sits on top of them and corrects their predictable errors. The authors describe it as “an effective post-processing tool rather than a replacement for input model improvement”.

“What’s nice about our model is it’s an add-on wrapper,” co-author Judah Cohen of MIT told MIT News, adding that the team used it with ECMWF’s model, “but it can be used with any model.” Lester Mackey of Microsoft Research told The Verge the tool can run on ordinary processors (CPUs) rather than specialised AI chips.

What the authors claim

According to the paper, PBC “doubles the modest subseasonal skill of the AI Forecasting System” from the European Centre for Medium-Range Weather Forecasts (ECMWF). It also improved ECMWF’s main forecasting model “for 91% of pressure, 92% of temperature and 98% of precipitation targets.”

The authors say that in ECMWF’s 2025 real-time forecasting competition, their forecasts “were the best for all weather variables and lead times”, beating forecasts from six national forecasting centres and 34 teams worldwide. These are the authors’ own results; several of the authors work for Microsoft.

What this means

• Longer-range forecasts could improve. Farmers and emergency planners could get more useful warnings weeks ahead.
• It is cheap to run. The method works on top of existing models and runs on ordinary computers.
• Anyone can try it. The code is open source, though the results so far are the authors’ own.

Sources

1. Guan, H., Mouatadid, S. et al., “Advancing subseasonal forecasting with machine learning”, Nature, 14 October 2026. https://doi.org/10.1038/s41586-026-11126-x
2. MIT News, “Researchers improve weather forecasts looking two to six weeks out”, Zach Winn, 14 October 2026. https://news.mit.edu/2026/researchers-improve-weather-forecasts-looking-two-six-weeks-out-1014
3. The Verge, “Microsoft has a new machine learning tool to improve weather forecasts”, Justine Calma, 14 October 2026. https://www.theverge.com/science/1011127/ai-weather-forecast-microsoft
4. Microsoft Research, “Enhancing AI and Dynamical Subseasonal Forecasts with Probabilistic Bias Correction”. https://www.microsoft.com/en-us/research/publication/enhancing-ai-and-dynamical-subseasonal-forecasts-with-probabilistic-bias-correction/

Related stories

• Google DeepMind Gives Up to 200 European Researchers Access to Its Frontier Models, Plus $3M in Compute Credits