A learned correction layer doubles a gold-standard weather model's accuracy at weeks two to six
Researchers including a team at MIT report in Nature that a machine-learning layer doubled a gold-standard model's forecast accuracy at two to six weeks. Beyond two weeks, energy, farm and emergency planners have had little reliable forecast to use.
The Scientist · Science desk
Farmers, energy industry: Cohen calls forecasts for this range very useful. Emergency officials: forecasts beyond two weeks are unreliable. Texas, February 2021: worst energy failure in state history, at least 246 killed. Energy officials: more time could have mitigated damage.
- capability Farmers Cohen says time-averaged forecasts for this range are already very useful for farmers, claim 13
- capability Energy industry Cohen says they are useful to energy industry members trying to match supply with demand, claim 13
- constraint Emergency response officials Forecasts beyond two weeks are notoriously unreliable, leaving a tighter window, claim 16
- precedent Texas, February 2021 After three winter storms, worst energy failure in state history; at least 246 people killed, claim 14
- exposure Energy officials More time to prepare, such as more trucks of natural gas, could have mitigated damage, claim 15
| Who | How | Kind | Claim |
|---|---|---|---|
| Farmers | Cohen says time-averaged forecasts for this range are already very useful for farmers | capability | 13 |
| Energy industry | Cohen says they are useful to energy industry members trying to match supply with demand | capability | 13 |
| Emergency response officials | Forecasts beyond two weeks are notoriously unreliable, leaving a tighter window | constraint | 16 |
| Texas, February 2021 | After three winter storms, worst energy failure in state history; at least 246 people killed | precedent | 14 |
| Energy officials | More time to prepare, such as more trucks of natural gas, could have mitigated damage | exposure | 15 |
What happened
- The entry placed first in a prominent international forecasting competition, according to the phys.org account of the Nature paper.
- PopSci reports that Cohen's team submitted three versions of the model to ECMWF, which hosts competitions that pit prediction models against one another.
- The phys.org account says the paper shows forecasts that for the first time beat climatologic averages between two and six weeks.
- MIT research scientist Judah Cohen says the approach also improves predictions of extreme events such as flooding and cold spells.
Why it matters
- capability Farmers and energy firms matching supply to demand already use time-averaged forecasts in this range, Cohen says, so a more accurate version would feed planning they already do.
- constraint The layer corrects another model's output, so a forecaster needs a strong underlying model to run it on.
- precedent Cohen's was the first postprocessed entry in the competition, so we'd expect rival groups to enter their own corrected models next time.
The layer sits on top of an existing model. It learns to correct that model's errors [1], and PopSci describes the method as postprocessing: the system compares its inaccuracies with historical data and adjusts from there [6]. The pairing matters because, according to the phys.org account, AI-only models have improved short-range prediction but do worse than dynamical models at longer ranges [9].
The range is hard for a stated reason. Cohen said it is referred to as the "predictability desert": at short range a forecaster can use the initial conditions of the atmosphere, at longer range the slow boundary forces such as El Nino, and the subseasonal range can use neither [10]. "In most cases forecasting in the subseasonal ranges is no different than going to a casino and rolling the dice," he said [11]. The reported doubling of accuracy [2] starts from that low base.
Neither account says which accuracy measure doubled, or on which base model. Nor do they give the number of forecasts scored, the competition's scoring procedure or the size of the extreme-event improvement. A planner needs lead time or a hit rate for a region and season.
A first place is a leaderboard result: the entry outscored the other entrants under the organisers' scoring [3]. Cohen's summary of the broader claim is "we consistently beat climatology averages with our models," he said [7]. A single competition is one test of that claim. Two of the paper's co-authors, Jakob Schloer and Joshua Talib, are researchers at ECMWF [12], the centre that hosts the competition [4].
The Texas freeze that opens the phys.org account killed at least 246 people [14]. The account says more time to prepare, for example by bringing more trucks of natural gas to disrupted areas, could have mitigated the damage [15]. A forecast at two to six weeks would be used to buy that kind of preparation time.
What to watch
- Whether the Nature paper's tables give the scoring metric, the base model and regional results. With those, the doubling could be read as days of lead time.
- Whether other groups enter postprocessed models in the next ECMWF competition and how they place against this one.
- How the layer's extreme-event forecasts perform in a winter shaped by the super El Nino Cohen mentions.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence60
- Adoption15
- Hype gap+20
- Incentives40
- Confidence58
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
A team including MIT researchers developed a machine-learning model that learns to correct errors in existing forecasting models to improve their output over the two-to-six-week time frame.
- [2]
Applied to a gold-standard forecasting model, the machine-learning model doubled forecasting accuracy over the two-to-six-week time frame.
- [3]
The model placed first in a prominent international forecasting competition.
- [4]
Cohen's team submitted three versions of their postprocessing model to the European Center for Medium-Range Weather Forecasts (ECMWF), which hosts seasonal competitions that pit the latest prediction models against one another.
- [5]
In a new Nature paper, the researchers show how their approach enables forecasts that for the first time beat climatologic averages between two and six weeks.
- [6]
With postprocessing, the AI improves by comparing its inaccuracies with existing historical data and then adjusting from there.
- [7]
Cohen said: "we consistently beat climatology averages with our models."
- [8]
Most competing systems have relied on historic climate averages, but this was the first time a group submitted a program with postprocessing capabilities.
- [9]
AI models have shown improvements in predicting weather at short ranges, but they are even worse than dynamical models at longer time ranges.
- [10]
Cohen said the two-to-six-week range is referred to as the 'predictability desert': in the short range one can use initial conditions in the atmosphere, in the longer term slow boundary forces such as El Niño, and the subseasonal range can take advantage of neither.
- [11]
Cohen said: "In most cases forecasting in the subseasonal ranges is no different than going to a casino and rolling the dice."
- [12]
Jakob Schloer and Joshua Talib, researchers with the European Center for Medium-Range Weather Forecasts, are co-authors of the paper.
- [13]
Cohen said time-averaged forecasts for this range are "already very useful for farmers and members of the energy industry trying to match supply with demand."
- [14]
In February 2021, following three severe winter storms, Texas suffered the worst energy infrastructure failure in state history, and at least 246 people were killed.
- [15]
If energy officials had had more time to prepare for the storms, for example by bringing in more trucks of natural gas to disrupted areas, damage from the event could have been mitigated.
- [16]
Weather forecasts beyond two weeks are notoriously unreliable, which leaves a tighter window for energy operators, farmers and emergency response officials.
- [17]
Cohen said the approach "also improves predictions of extreme weather events, whether it's flooding or cold spells or something else."
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
2 independent publishers whose own reporting we read for this story.
- news.mit.eduResearchers improve weather forecasts looking two to six weeks out
1 article · October 14, 2026
- phys.orgAI-enhanced weather model doubles predictive power two to six weeks out
1 article · October 14, 2026
- popsci.comNew weather model may finally offer a reliable six-week forecast
1 article · October 14, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.