Machine learning bias correction sharpens two-to-six-week weather forecasts, Nature study finds

A machine learning method called probabilistic bias correction, which learns a forecast model’s systematic errors from its own past forecasts and corrects them, doubled the subseasonal skill of the European Centre for Medium-Range Weather Forecasts’ (ECMWF) AI Forecasting System and improved ECMWF’s operationally debiased dynamical model for 91% of pressure, 92% of temperature and 98% of precipitation forecast targets, according to a study published in Nature on Oct. 14 by authors including Microsoft Research scientists. Subseasonal forecasts look two to six weeks ahead, a window the paper ties to agricultural planning, water allocation and disaster preparedness. In ECMWF’s 2025 real-time forecasting competition, the method’s global forecasts were the best for all weather variables and lead times, outperforming the dynamical models of six operational forecasting centres and the systems of 34 teams, the paper says; Microsoft said the Kenya Meteorological Service Authority is testing the method in its weekly forecasts for farmers this rainy season.