MIT-Led Study in Nature: Weather Forecasts That Beat Climatological Averages Two to Six Weeks Out, for the First Time
An MIT-led research team has developed a machine-learning model that learns to correct errors in existing weather forecasts, and which, according to the researchers, doubles forecasting accuracy in the two-to-six-week window. The study is published in Nature, and according to MIT's press release it is the first time anyone has managed to beat climatological averages in this interval — an area meteorologists call the "predictability desert." According to the press release, the method also won first place in an international forecasting competition.
The results come from the press release and secondary coverage, not from an independent review of the article — but they point to something concrete: not a new AI model replacing weather forecasting, but a postprocessing layer that fixes the errors in the models we already trust.
What's new
The researchers, with MIT scientist Judah Cohen on the team, developed a machine-learning model that learns to correct errors in existing forecasting models in order to improve their output in the two-to-six-week window. According to MIT's press release, carried by Phys.org, the model doubled forecasting accuracy when applied to a "gold standard" model, and, per the press release, took first place in a prominent international competition.
In the Nature paper, the researchers describe, according to the press release, how the approach makes it possible to produce forecasts that "for the first time beat climatological averages between two and six weeks." The exact metrics behind "doubled accuracy," however, are not specified in the available coverage, and AIMag has not independently verified the numbers itself.
The competition context comes from Popular Science, which writes that Cohen's team submitted three versions of the model to ECMWF's forecasting competitions — reportedly the first entry with postprocessing capabilities. Neither the name of the competition nor which model was postprocessed is mentioned in the available sources.
Why this particular window is so hard
Two to six weeks ahead is what meteorologists call subseasonal forecasting, and it has long been considered an almost impossible in-between zone: too far out for the classic one-to-ten-day forecasts, too close for seasonal patterns to be used. Popular Science relays Cohen's characterization of the territory as a "predictability desert" — a desert landscape for predictability. Here, models run up against slow-moving, highly variable forces, such as this year's super El Niño, according to Popular Science.
Cohen traces the theoretical two-week limit to Edward Lorenz's chaos theory from the 1970s. According to the press release, Cohen explains Lorenz's position as follows: that there is a theoretical two-week limit beyond which we simply cannot predict the weather, because the noise in the system grows so much that it overwhelms any signal you start with. This idea, Cohen continues, "was influential in the field and may have held people back from exploring forecasting beyond two weeks" — but, he adds, "I never really believed that completely."
The distinction matters for understanding what the team has actually done: they are not claiming to have overturned chaos theory. The weather state as it will be in two weeks remains unpredictable in detail. What they show is that there are slow patterns — such as El Niño cycles — that can be captured better if you correct the systematic errors the models already make.
The mechanism: postprocessing, not a new model
The technique is called postprocessing. Instead of building an entirely new weather simulation, the team runs the AI model behind existing forecasts: according to Popular Science, the AI improves by comparing its inaccuracies with existing historical data, and then adjusts accordingly. The model thus learns which errors a given forecasting model typically makes in this window — and corrects for them.
It is a modest technical intervention with potentially large consequences: dynamical models, based on physics, and modern AI models both break down in the two-to-six-week window. A postprocessing model that knows where they usually go wrong can extract value from both. And because the method builds on the models we already have, in principle no new supercomputer or wholesale rebuilding of the forecasting apparatus is needed to adopt it.
Who is behind it
The study has two lead authors: Hannah Guan, a student at Harvard College, and Soukayna Mouatadid, a postdoctoral researcher at the University of Toronto. Also on the author list, according to the press release, are Genevieve Flaspohler; Paulo Orenstein (associate professor at Brazil's National Institute for Pure and Applied Mathematics); Jeremy Berman (the data company DTN); Jakob Schloer and Joshua Talib (both at the European Centre for Medium-Range Weather Forecasts, ECMWF); and Alex Lu, Haiyu Dong, Jonathan Weyn, Zekun Ni and Lester Mackey — all researchers at Microsoft.
It is worth noting a discrepancy in the coverage: Popular Science frames the collaboration as a project between "Cohen and his colleague Lester Mackey," while MIT's press release lists a larger, cross-institutional team with Guan and Mouatadid as lead authors. The sources do not provide grounds for determining which description best captures the division of labor, so both are relayed here.
Why it matters
According to the press release, the motivation is not merely academic. In February 2021, Texas was hit by three severe winter storms, leading to the worst energy infrastructure failure in the state's history. At least 246 people died as a result of supply failures, which led to shortages of food, water, heat and power. Longer lead time — two to six weeks instead of days — could, according to the press release, have given authorities and energy companies time to prepare.
Cohen also points out, according to Popular Science, that the improvement applies to extreme weather events, not just average conditions:
"It also improves forecasts of extreme weather events, whether that's flooding, cold spells or something else. Predicting both is very important, but extreme weather may be the most impactful part of this for society."
Concretely, that could potentially mean something for farmers planning planting and harvesting, energy companies that must balance consumption and production weeks in advance, and authorities that need advance warning of floods or cold spells.
Open questions
Several things remain unresolved. The competition the team won, according to MIT's press release, is not named in the sources — Popular Science writes that the entries were submitted to ECMWF's forecasting competitions, but which specific competition and organizer applied is unclear. Nor is it stated which "gold standard" model was postprocessed, although ECMWF's presence on the author list suggests the answer. The numbers behind "doubled accuracy" are not specified, and it is unclear how robust the results are across regions, seasons and weather phenomena. It is also not stated what operational use in a national forecasting apparatus would require in terms of validation, infrastructure and updating over time.
But the main impression from the available documentation is this: the most important advance in long-range forecasting may not necessarily come from bigger AI models alone, but from systematically training machines to see where our models go wrong — and correcting it.
AIMag has not independently verified the results in the Nature article; the figures and the claims about doubling and first place come from MIT's press release and Popular Science's coverage.