An AI that fixes weather forecasts weeks ahead

What is it?

A weather forecast is easy to trust for tomorrow. It is hard to trust for next month. Small errors pile up as a model steps forward in time, and the atmosphere is chaotic, so tiny differences grow.

Scientists call the two-to-six-week range “subseasonal”. The paper calls it a “predictability desert”.

A team from Microsoft, MIT, Harvard, ECMWF (the European weather centre) and other places built a fix. It does not replace the weather model. It sits on top and learns the model’s habits from past forecasts.

The method is called probabilistic bias correction. “Bias” means a mistake that keeps pointing the same way. “Probabilistic” means the forecast gives chances, such as a 40 in 100 chance of a wetter than normal week, not one answer.

How well did it work?

It entered a contest run by ECMWF, called the AI Weather Quest. In the first season, from September to November 2025, it placed first for every weather variable and every lead time. It beat forecasts from six national weather centres and every other team.

2x

the skill of ECMWF's own AI forecast model, once corrected

For rainfall, the corrected forecast beat the plain climate average. The models from all six weather centres did not, after their usual fixes. The authors say this is where their gains were biggest.

The paper was first posted on arXiv on 17 April 2026 and updated on 4 September 2026. MIT News reported it on 14 October 2026.

Why should you care?

Farmers decide when to sow and when to protect crops. Power companies plan for cold spells. Relief teams stock supplies. A good forecast a few weeks out gives them time that a two-day warning does not.

It also shows a useful idea about AI. A model does not have to do everything. A small AI that learns from past mistakes can make an older tool better.

The authors have released their code as open source, so other weather services can use it.

What’s the catch?

• The scores are small. A forecast two to six weeks out is still far less sure than tomorrow’s.
• The results come from the authors’ own tests and one contest season.
• This is a research paper. It is not a forecast you can look up today.

Try it

Pick something you guess often, like how long your bus takes. Write down your guess and the real time for ten days. Do you miss in the same direction each time? If so, you have found a bias. Correcting for it is the same idea the paper uses.

Sources

• Advancing Subseasonal Forecasting with Machine Learning, Guan, Mouatadid, Orenstein, Cohen, Mackey and others, arXiv, version of 4 September 2026
• Researchers improve weather forecasts looking two to six weeks out, MIT News, 14 October 2026