Research watchScienceNew
Machine-learning correction improves two-to-six-week weather forecasts in Nature study
A Nature paper reports that a machine-learning method that corrects systematic errors in ECMWF forecasts improves skill two to six weeks ahead, where forecasts are weakest. Skill at that range remains modest even after correction.
Evidence: 4 sources · 1 official record · 1 independent report · 2 research papers · 3 open questions
What happened
- On 14 October, Nature published "Advancing subseasonal forecasting with machine learning" by Hannah Guan, Soukayna Mouatadid and colleagues.
- The method, probabilistic bias correction, learns from a model's past forecasts to reduce its systematic errors at lead times of two to six weeks.
- It was applied to both ECMWF's physics-based model and its AI forecasting system.
- The authors report that its global forecasts ranked first for all variables and lead times in ECMWF's 2025 real-time subseasonal forecasting competition.
Why it matters
Forecasts several weeks ahead help with planning for heat waves, cold spells, and flooding in farming, energy, and disaster response, but their skill has been low. A cheap correction applied to existing models could be easier to adopt than a new forecasting system.
What is confirmed and what is claimed
Confirmed
- The peer-reviewed paper was published in Nature in October 2026; an earlier version was posted to arXiv as a preprint (Nature; arXiv; MIT News). 1234
- Collaborators include researchers at MIT, Harvard, the University of Toronto, Microsoft, and ECMWF (MIT News). 23
- The method post-processes output from existing forecast models rather than replacing them (Nature abstract; MIT News). 123
Claimed, not independently verified
- The authors report that the method roughly doubles the skill of ECMWF's AI subseasonal forecasts on average and improves the corrected physics-based model on most pressure, temperature, and precipitation targets.
- MIT says the method also improves forecasts of extremes such as floods and cold spells, where ECMWF's existing correction made them worse.
- The team says the model is open source and does not need a supercomputer to run.
Weather forecasts are reliable for about a week or so. Seasonal outlooks, months ahead, draw on slow-moving signals such as ocean temperatures. The stretch between them, two to six weeks out, is called the subseasonal range, and forecasts there have had little skill compared with simply using long-term averages.
The new paper does not build another forecasting model. It adds a machine-learning layer, probabilistic bias correction, that studies how an existing model's past forecasts went wrong and adjusts new forecasts to match. The team applied it to forecasts from ECMWF, the European weather center, including both its traditional physics-based system and its newer AI forecasting system.
Context
AI weather models have drawn attention for matching physics-based forecasts at ranges of days. Gains weeks ahead are harder to judge, because skill is low to begin with: doubling a small score still leaves a forecast that is often wrong. The authors' strongest independent evidence is a first-place result in ECMWF's 2025 real-time competition, where forecasts were scored against weather that had not yet happened.
An ECMWF scientist who commented on the paper for the Science Media Centre works at the center whose models were corrected, a connection worth keeping in mind when reading reactions. See our guide to reading AI-for-science claims and to benchmark claims.
What to watch
- Whether forecasting centers test the method in their own operations.
- Independent evaluations on other models and regions.
- Whether forecasts of extreme events, the most useful case, hold up in future real-time tests.
What we don’t know yet
- Whether ECMWF or national weather services will adopt the method operationally.
- How well the gains hold for regions and seasons outside the evaluation period.
- Exact figures differ between preprint versions; the published paper's numbers are the reference.
Evidence and sources
- Peer-reviewed study Advancing subseasonal forecasting with machine learning
- Primary announcement Researchers improve weather forecasts looking two to six weeks out
- Independent reporting An AI tool improves weather forecasts up to a month ahead
- Preprint (not peer reviewed) Advancing Subseasonal Forecasting with Machine Learning (preprint)