Advancing Subseasonal Forecasting with Machine Learning

The European Centre for Medium-Range Weather Forecasts (ECMF) has introduced the machine learning framework PBC, which significantly reduces systematic errors by correcting historical probability forecasts. This framework is integrated into ECMF’s leading dynamic and AI models, doubling the forecasting skills at sub-seasonal scales (2–6 weeks). It also improves the forecasting skills for pressure, temperature, and precipitation targets to 91%, 92%, and 98%, respectively, using the operational bias-corrected dynamic model. In the ECMWF 2025 real-time forecasting competition, the global forecasting system using PBC ranked first in all weather variables and lead times, outperforming six operational forecasting centers, international dynamic multi-model ensembles, ECMF AI systems, and 34 teams worldwide. These advancements help to more accurately predict extreme events, thereby improving agricultural planning, energy management, and disaster preparedness in vulnerable communities.