Researchers improve weather forecasts looking two to six weeks out

A team including MIT researchers has built a machine-learning add-on that sharply improves weather forecasts two to six weeks ahead, a span known as subseasonal forecasting and nicknamed the "predictability desert." MIT News reports that the work, described in Nature, produced forecasts that for the first time beat long-term climate averages in that window. Co-authors come from MIT, Harvard, the University of Toronto, Microsoft, the data company DTN and the European Center for Medium-Range Weather Forecasts (ECMWF).

The tool, called Probabilistic Bias Correction, does not forecast on its own. It sits on top of existing models, learns from their historical mistakes and shifts their probabilistic output, updating as new data arrives. Last year the team entered three models using it in ECMWF's seasonal competition. The winner combined ECMWF's model with another and applied the correction, doubling the original's predictive power at that range.

MIT research scientist Judah Cohen says the gain extends to extreme events such as floods and cold spells, a statistically significant improvement he contrasts with ECMWF's own non-AI correction, which worsened extreme-weather predictions. The method is open source and needs a supercomputer only for training, not for running, and it also lifted weaker models to roughly ECMWF's level.

The stakes are practical: MIT News points to the February 2021 Texas storms, which killed at least 246 people, as a case where more warning could have limited damage. Cohen says Kenya's meteorological office has already inquired about using the model to help agriculture through its monsoon season.