Better forecasts 3 to 4 weeks out.
Microsoft's AI fix won a global contest.

What happened?

On October 14, 2026, a team led by Microsoft Research and MIT published a machine learning method called probabilistic bias correction, or PBC, in Nature Nature ↗. It improves weather forecasts two to six weeks ahead, and an ensemble using it ranked first in ECMWF's 2025 AI Weather Quest Nature ↗ Gizmodo ↗.

Weather forecasts are reliable for about two weeks. Beyond that, small differences in today's conditions grow into very different outcomes, and errors pile up with every step a model takes forward in time Gizmodo ↗. This two to six week window is called the subseasonal range, and it matters for farmers, power grids and disaster planners.

PBC does not replace weather models. It sits on top of them as a post processing step, learning from years of past forecasts and what actually happened, then correcting each model's systematic mistakes Nature ↗ Unite.AI ↗. The authors say it works on both traditional physics based models and newer AI models Nature ↗ Gizmodo ↗.

The team tested it on forecasts from the European Centre for Medium Range Weather Forecasts (ECMWF), including its AI model AIFS SUBS and its hybrid model POET Nature ↗. In the live AI Weather Quest from September to November 2025, the team's ensemble, called MicroDuet, ranked first for every variable and lead time, ahead of six operational weather centers, ECMWF's own AI models and 34 competing teams Nature ↗ Unite.AI ↗.

The work was led by Lester Mackey of Microsoft Research and Judah Cohen of MIT, with co first authors Hannah Guan and Soukayna Mouatadid Unite.AI ↗. The 13 authors come from Microsoft, MIT, Harvard, the University of Toronto, ECMWF and other groups Unite.AI ↗.

How does the AI work?

A forecast model usually runs many times with slightly different starting points, producing an ensemble of possible futures. PBC turns that ensemble into probabilities for whether temperature, rain or pressure will land above, near or below normal Unite.AI ↗.

Figures are from the Nature paper and are relative gains over low starting skill Nature ↗ Unite.AI ↗.

It runs two machine learning models side by side. One, called Debias++, makes adjustments specific to each location, date and range of values. The other, Persistence++, blends in recent observations and long term averages. PBC then combines the two into a single forecast Unite.AI ↗. It is light enough to retrain every day using only data available on the forecast date Unite.AI ↗.

Judah Cohen told Gizmodo that older correction methods sometimes made extreme weather forecasts worse, while this one improved them Gizmodo ↗. The tool does not need a supercomputer, which could help countries with limited forecasting budgets Gizmodo ↗.

Which fields could this affect?

PBC is a research tool with open code. These connections are our assessment.

What has been checked?

The evidence is a peer reviewed, open access Nature paper, live competition results run by ECMWF, open source code and news coverage. Leapscope reviewed these sources; we did not rerun any forecasts.

Evidence status: Peer reviewed paper. Stage: Verified. Peer reviewed in Nature and ranked first in a live ECMWF competition; not yet used by weather agencies.

From contest winner to daily forecasts

1. Watch the next contest rounds.The team also led the December to February round, so later seasons will show if the lead holds Unite.AI ↗.
2. Look for agency adoption.Kenya's weather service is testing it for monsoon forecasts Gizmodo ↗.
3. Check independent tests.Open code means other groups can try it on their own models GitHub ↗.

This is our suggested way to follow the story, not a promised timetable.

Can I use it today?

The code is free on GitHub GitHub ↗. It is a tool for forecasters and researchers, not a weather app.

A few things you might be wondering