Bias Correction Doubled European AI Weather Model’s Subseasonal Forecast Skill

Microsoft researchers and collaborators developed a machine-learning method that improved weather forecasts two to six weeks ahead, including results that doubled the subseasonal skill of a European forecasting system’s AI model.

The method, called probabilistic bias correction, learns from historical probabilistic forecasts and adjusts predictions from existing models. It was applied to forecasts from the European Centre for Medium-Range Weather Forecasts, which runs a global weather forecasting system.

For the center’s AI Forecasting System, probabilistic bias correction doubled subseasonal forecast skill. When applied to an operationally debiased dynamical model, it improved 91% of pressure targets, 92% of temperature targets and 98% of precipitation targets.

Subseasonal forecasts cover the weeks between medium-range forecasts and seasonal outlooks. The European center’s AI Weather Quest evaluates forecasts for days 19–25 and 26–32, periods when errors and biases can build over time.

The research paper was posted to arXiv on April 17, 2026. Probabilistic bias correction is a correction layer for existing models.