Nature Study Introduces Machine-Learning Tool That Boosts Subseasonal Forecast Accuracy

A new study published in Nature presents a machine-learning post-processing "wrapper" designed to correct systematic errors in forecast models and improve the accuracy of two-to-six-week subseasonal predictions. In ECMWF's 2025 real-time forecasting competition, the tool produced the most accurate global forecasts across all weather variables and forecast periods, outperforming numerical models from six operational centers, an international model ensemble, ECMWF's own AI model, and systems submitted by 34 teams worldwide.

The tool, called the probabilistic bias correction (PBC) wrapper, was developed by Judah Cohen, a climatologist at MIT, and Lester Mackey, a statistical machine learning researcher at Microsoft, who collaborated on the idea for more than a decade. According to the researchers, the framework is model-agnostic and can be applied to both traditional numerical models and newer AI models. In tests, the PBC wrapper doubled the subseasonal accuracy of ECMWF's leading AI system and improved the numerical model's skill for 91% of pressure, 92% of temperature, and 98% of precipitation targets.

The study also demonstrates improvements to land forecasts in most regions and notable gains in extreme weather forecasting. In tests on real-world floods from the past decade, the tool improved flood prediction skill by 76% to 92%; for severe floods that caused more than 100 deaths or displaced at least 80,000 people, forecasting skill improved by 86% to 150%. The open-source tool does not require a supercomputer to run, and Kenya's meteorological office has already begun experimenting with it to improve monsoon forecasting.