A team including MIT research scientist Judah Cohen and Microsoft's Lester Mackey published a Nature paper on October 14, 2026, describing a machine-learning postprocessing layer that learns to correct errors in existing forecast models over the two-to-six-week range.

According to MIT News, when the tool was applied to a gold-standard forecasting model it doubled forecasting accuracy in that window and placed first in a prominent international forecasting competition.

Gizmodo reports that in the European Centre for Medium-Range Weather Forecasts' 2025 real-time contest, the tool outperformed numerical models from six operational centers, ECMWF's own AI model, and 34 competing teams worldwide.

The two-to-six-week range is known as the 'predictability desert' because it is too far out to exploit initial atmospheric conditions and too near-term to rely on slow boundary forces like El Niño, Cohen told MIT News, invoking Edward Lorenz's theoretical two-week chaos limit.

Until now, no model had beaten simple climatological averages in this window — Cohen says you could 'literally flip a coin and do better' with existing subseasonal forecasts.

Gizmodo adds that even standout AI systems like Google DeepMind's WeatherNext family struggle here, as simulation errors accumulate over successive time steps.

According to Gizmodo, the probabilistic bias correction (PBC) tool doubled the subseasonal accuracy of ECMWF's AI system and improved the numerical model's skill for 92% of temperature, 98% of precipitation, and 91% of pressure targets.

For forecasts made three weeks ahead, it significantly improved prediction of extreme heat, extreme precipitation, and unusually high sea level pressure.

Tested on real floods from the last decade, Gizmodo reports it improved flood prediction skill by 76% to 92%, and by 86% to 150% for severe floods causing more than 100 deaths or displacing at least 80,000 people.

Cohen pointed MIT News to the February 2021 Texas winter storms, which killed at least 246 people amid energy infrastructure failures; energy officials say that two weeks of warning, rather than two days, would have allowed adequate preparation like trucking in natural gas.

Gizmodo notes that previous bias-correction schemes actually made extreme weather forecasts worse by pulling toward the mean, while the PBC tool showed clear improvement — the outcome Cohen calls the most impactful for society.

Farmers could time fertilizing around rain and protect crops from freezes, MIT News reports.

MIT News reports the tool is model-agnostic, open source, low cost, and does not require a supercomputer to run, despite being trained on MIT's Engaging cluster.

Applied to lesser models, it put them on par with ECMWF's state-of-the-art systems, and it improved models already bias-corrected by other AI algorithms. According to Gizmodo, Kenya's meteorological office is already experimenting with it to improve monsoon forecasting for agricultural yields.

Cohen hopes energy and emergency management operators worldwide will adopt it to buy preparation time as climate change drives more extreme weather.