A machine-learning correction layer built by a team led from MIT has roughly doubled the accuracy of the European Centre for Medium-Range Weather Forecasts model for forecasts two to six weeks ahead, a range where weather prediction has long struggled. The work appears in Nature as “Advancing Subseasonal Forecasting with Machine Learning,” and MIT News announced it on October 14, 2026. Judah Cohen, a research scientist in MIT’s Department of Civil and Environmental Engineering, describes the span as a predictability desert.
A correction layer, not a new weather model
The method does not replace any forecasting system. It is a postprocessing step that studies a model’s past errors and shifts its future output accordingly, and it keeps adapting as more forecasts and observations arrive, according to the MIT release. Because it sits on top of an existing model, it can in principle be applied to any of them.
The team trained it on MIT’s Engaging supercomputing cluster, but running it does not require a supercomputer. The release describes the tool as open source and low cost, which matters for forecasting agencies without large computing budgets. When the researchers applied it to the European Centre for Medium-Range Weather Forecasts model, accuracy in the two-to-six-week range doubled, and the layer also improved models that other AI methods had already bias-corrected. Land forecasts improved across most regions of the world.
Beating the long-term average
The result that the researchers stress most is not the doubling itself but what it was measured against. According to the release, no one had previously shown that a model predicts weather at this range better than long-term climatological averages, and this approach is the first to do so. Cohen has long doubted the idea, traced to Edward Lorenz’s chaos-theory work in the 1970s, that predictability simply ends at about two weeks, and said he never fully accepted it.
His explanation of why the range is hard is concrete. At two to six weeks, the starting conditions that drive short-range forecasts have faded, while the slow boundary forces that shape seasonal outlooks have not yet taken over. The window falls between the two methods. A submission combining the European model with another model plus this correction layer took first place in an ECMWF forecasting competition.
Why does the baseline matter so much? Climatological averages are the forecaster’s null hypothesis: tell people that a given week will look like the typical version of that week and, at this range, the guess is hard to beat. The release says forecasts beyond two weeks are unreliable, which is why that average has been so difficult to improve on. A system that consistently scores better than the average is therefore delivering information that a calendar alone could not, and the release presents the team’s layer as the first to clear that bar for two-to-six-week forecasts.
Extreme weather at the edges
Cohen’s team reports a statistically significant improvement in predicting extreme events. That finding covers the events forecasters most want to flag early, not only typical conditions, and the release stresses that the gain is statistically significant rather than anecdotal. Hannah Guan, Soukayna Mouatadid and Genevieve Flaspohler are listed as lead authors of the Nature paper, with Cohen as the senior voice quoted in the MIT account. The researchers explain that standard bias correction tends to pull forecasts toward the mean, whereas extremes sit at the edges of the probability distribution, which is why the researchers say a correction that learns from a model’s errors can help where averaging does not.
The motivation is practical. The release points to the 2021 winter storms in Texas, which caused the state’s worst energy infrastructure failure and killed at least 246 people. Forecasts beyond two weeks are unreliable, which gives energy operators, farmers and emergency officials only a narrow planning window.
Who built it and who is asking for it
The lead authors are Hannah Guan of Harvard College, Soukayna Mouatadid of the University of Toronto and Genevieve Flaspohler, who earned her MIT degrees in 2018 and 2022. Paulo Orenstein of Brazil’s National Institute for Pure and Applied Mathematics, Jeremy Berman of DTN, Jakob Schloer and Joshua Talib of the European centre, and Microsoft researchers Alex Lu, Haiyu Dong, Jonathan Weyn, Zekun Ni and Lester Mackey are co-authors. Cohen credits Mackey with the original idea of using machine learning to reduce forecast errors.
Funding came from the National Science Foundation, Brazil’s Carlos Chagas Filho Foundation for Research Support and the European Union, and the NSF funder portal lists related releases. The Kenya Meteorological Department has already contacted the team about using the model to forecast the annual monsoon season, an early sign of interest from a national weather service. Because the layer needs no supercomputer to run, a service of that kind could in principle test it against its own forecasts, though the release does not say whether any agency has adopted it. The paper is reachable through its Nature DOI, while a preprint with the same title and a NeurIPS 2026 listing for the work are also online.
This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.
More from Morning Overview
- NOAA now gives this winter a 75% chance of the strongest El Nino since 1950
- Doctors warn a silent liver disease now affects one in three American adults
- The FBI told the rest of ShinyHunters to surrender after a 24-year-old was arrested
- Two passengers died at an Ohio toll plaza, and the NTSB now wants the cash lanes shut
