Disclosure: this article contains affiliate links to online courses. If you sign up through them, DIGITAL WOXTER TECHNEWS may earn a commission at no extra cost to you. We choose courses for their content, not for payments.
Weather forecasts are great for tomorrow and pretty good for next week, but ask about the weather three weeks from now and most models shrug. A new AI tool, described in a study in Nature, makes those far-off forecasts clearly better, and it does it without replacing a single weather model.
Why it matters: the gap between “next week” and “next season” is where farmers decide when to plant, power companies prepare for cold snaps and emergency services get ready for floods. A few extra weeks of warning about a heatwave or a deep freeze can be worth a lot, and Nature’s news team highlights that the tool improves early warnings of heat, cold, heavy rain and flooding.
A spell-checker for forecasts
The method is called Probabilistic Bias Correction, or PBC. It is a “post-processing” step: it takes the output of an existing forecast model and fixes it, a bit like a spell-checker going over a text someone else wrote. According to the Science Media Centre Spain summary, it learns from three things: the model’s past mistakes, recent weather trends and the typical climate of each place. Then it adjusts both the most likely outcome and how uncertain that outcome is.
The best part is how light it is. As MIT’s write-up on Phys.org explains, the tool can be bolted onto any forecasting model, updates itself as new data arrives, does not need a supercomputer and is open source.
What the tests showed
The team, led by Harvard undergraduate Hannah Guan and University of Toronto postdoc Soukayna Mouatadid, with MIT’s Judah Cohen and colleagues from the European Centre for Medium-Range Weather Forecasts (ECMWF), Microsoft and others, entered ECMWF’s 2025 AI forecasting competition and won it. Their winning entry combined ECMWF’s model with a second model and applied the correction on top.
- About twice as good at 2-6 weeks. Applied to ECMWF’s model, the correction roughly doubled its predictive skill in that range, both sources report.
- Better than “the usual weather”. Long-range forecasts are judged against climatology, which simply means what the weather is normally like at that time of year. “We consistently beat climatology averages with our models,” Cohen told MIT News.
- Better warnings of extremes. The improvement in predicting extreme events was statistically significant, something ECMWF’s own non-AI correction did not achieve.
The fine print
Independent experts are positive but careful. Froila M. Palmeiro of the Institute of Geosciences (CSIC-UCM) in Madrid calls it a significant step forward, but notes that the gains are concentrated in the tropics and are limited in Europe. Robert Monjo of the Foundation for Climate Research warns that some rainfall gains look huge only because the starting skill was below zero, and that the tool cannot fix a base model that misses a rare situation completely. Matthew Chantry, who leads machine learning at ECMWF, says turning research like this into an everyday operational service is very challenging. All of these comments were collected by the Science Media Centre Spain.
So no, your phone app will not tell you whether it rains on your birthday next month. What it may soon get is a more honest “chance of a warmer, wetter week than usual” a month ahead, which is exactly the kind of signal planners need.
What happens next
The code is public, so national weather services can test it on their own models. Kenya’s meteorological department has already asked about using it to forecast its monsoon season for farmers, according to MIT. AI is also changing how science itself gets checked, as we saw when an AI wrote 700 maths papers.
Learn it yourself
Want to go further? These online courses cover the skills behind this story:
- AI For Everyone Coursera: Andrew Ng's non-technical DeepLearning.AI course: what machine learning can and can't do, in plain language, for any background.
See the course on Coursera
Read next
- AI wrote 700 maths papers: what Lean proof checks really mean
- How close are we to AGI? Where things really stand in 2026
- AI drone with one camera beats world champion pilots
The week's best tech stories, free
One email a week with the stories that matter, explained simply. No spam.
Sources
- Advancing subseasonal forecasting with machine learning – Guan, Mouatadid, Orenstein, Cohen et al., Nature, 14 October 2026
- An AI tool improves weather forecasts up to a month ahead – Nature, 21 October 2026
- AI-enhanced weather model doubles predictive power two to six weeks out – Phys.org / MIT News, 14 October 2026
- An artificial intelligence method improves weather forecasts up to six weeks in advance (expert reactions) – Science Media Centre Spain, 14 October 2026
