AI Error-Corrector Doubles 2-6 Week Forecast Skill
By @sharedot · · 8 pages
- Science
- Weather
- Machine Learning
A machine-learning 'wrapper' called PBC doubled the accuracy of two-to-six-week forecasts and became the first tool to beat climatology in that range.
A record broken in the 'predictability desert'
A team from MIT, Microsoft, ECMWF and partner institutions has published a Nature paper describing a machine-learning postprocessing layer that corrects errors in existing weather models, called Probabilistic Bias Correction (PBC).
Applied to ECMWF's gold-standard dynamical model, it doubled forecasting accuracy in the two-to-six-week 'subseasonal' range and placed first in a prominent international forecasting competition.
For the first time, the corrected forecasts beat climatological averages in that window — a bar no model had cleared before.
"Up until now, no one had been able to show a model predicts weather better than long-term averages for these subseasonal forecasts," said MIT research scientist and co-author Judah Cohen. "They've pretty much all been worse than that — you could literally flip a coin and do better."
Why the two-to-six-week gap resisted prediction
The two-to-six-week window has long been called the 'predictability desert.' MIT's Edward Lorenz established chaos theory's theoretical two-week limit in the 1970s, beyond which tiny changes in initial conditions overwhelm any predictive signal.
As Cohen puts it, in this window forecasters can neither capitalize on initial conditions like short-range models nor on slow boundary forces like the super El Niño dominating this winter.
"In most cases forecasting in the subseasonal ranges is no different than going to a casino and rolling the dice," Cohen said.
The team's insight was not to build a new model but to quantify and remove the systematic errors models already make — a postprocessing wrapper that adapts as more forecasts and data become available.
The evidence: a global competition win
The Microsoft Research story reports that in the inaugural three-month season of ECMWF's AI Weather Quest, launched in August 2025, the MicroEnsemble team's PBC-based entry, MicroDuet, placed first for every weather variable and lead time, outperforming six leading operational forecasting centers and 34 competing teams worldwide.
Gizmodo reports the tool improved ECMWF's leading numerical model's skill for 91% of pressure, 92% of temperature, and 98% of precipitation targets, and doubled the accuracy of ECMWF's AI Forecasting System.
MicroDuet even anticipated two brutal winter storms that struck the United States in early 2026, 19 to 26 days in advance, per Microsoft.
Extreme weather is where it matters most
The most consequential finding is that PBC improved forecasts of extreme events, which traditional bias correction tends to smooth away. "When the ECMWF developed their own, non-AI, error-correction to their model, it got worse at predicting extreme weather," Cohen told MIT News.
"Traditional bias correction is always trying to get to the mean, whereas extreme weather is on the fringe of probabilities."
Gizmodo reports that PBC improved flood prediction skill by 76% to 92% on real floods from the past decade, and by 86% to 150% for severe floods that killed more than 100 people or displaced at least 80,000.
The stakes echo events like the 2021 Texas winter storms that killed at least 246 people, where more lead time could have let utilities shore up natural gas supplies and infrastructure.
Cheap, open, and already deployed
Unlike physics-based modeling on supercomputers or AI models needing powerful GPUs, PBC runs on CPUs and is available as open-source software, The Verge reports — making AI-enhanced forecasting accessible to resource-strapped agencies in developing countries.
Kenya's Meteorological Service Authority is already testing PBC ahead of an expected extreme rainy season during the record El Niño.
"PBC corrects the probabilities themselves, and across the Greater Horn of Africa it improved our forecasts more than any other method we tested," principal meteorologist Dr. Phillip Okello said, per Microsoft.
Accurate forecasts decide which seeds Kenyan farmers buy, when to plant, and when to fertilize in a region largely lacking irrigation.
What comes next
The team showed the technique improves land forecasts in most regions across the globe and lifts even lesser models to near-ECMWF baseline performance, and it works on models already bias-corrected by other AI algorithms, MIT News reports.
Cohen hopes energy operators and emergency managers will use the extended lead time — energy officials told him that two weeks of warning about an ice storm could be enough to make adequate preparations, versus two days.
Microsoft is also exploring extensions, including its subseasonal-to-seasonal S2S model now in private preview, and says continued collaboration with the forecasting community will drive further applications of the open-source framework.
Sources
- news.mit.edu › Researchers improve weather forecasts looking two to six weeks out
- news.lavx.hu › Machine learning model doubles accuracy of 2-to-6-week weather forecasts
- microsoft.com › New AI framework bridges a critical gap in mid-range weather forecasts
- gizmodo.com › Forget the 10-Day Forecast. AI Just Improved Extreme Weather Warnings Up to 6 Weeks Out
- theverge.com › Microsoft has a new machine learning tool to improve weather forecasts