Researchers at Microsoft and MIT published a open-source machine learning tool named Probabilistic Bias Correction in Nature on October 14, 2026. Designed to correct systemic biases in subseasonal weather forecasts two to six weeks out, the tool doubled the accuracy of ECMWF's AI forecast model and improved flood prediction skill up to 150%. Running efficiently on standard CPUs, the software enables resource-constrained nations like Kenya to improve local disaster warnings.
Probabilistic Bias Correction tool development and design
- ▪The Probabilistic Bias Correction framework uses two machine learning models, Debias++ and Persistence++, to correct probability distributions of existing dynamical, artificial intelligence, or hybrid weather forecast models.
- ▪Lester Mackey of Microsoft Research New England and Judah Cohen of the Massachusetts Institute of Technology developed the open-source Probabilistic Bias Correction tool to reduce systematic errors in weather forecasting models.
Performance and accuracy gains in testing
- ▪Applying the Probabilistic Bias Correction tool doubled the subseasonal forecasting accuracy of the European Centre for Medium-Range Weather Forecasts artificial intelligence forecasting system in 2025 testing.
- ▪Applying the Probabilistic Bias Correction tool to the European Centre for Medium-Range Weather Forecasts numerical model improved forecast skill for 98% of precipitation and 91% to 92% of temperature targets.
- ▪In the European Centre for Medium-Range Weather Forecasts real-time AI Weather Quest competition, the Probabilistic Bias Correction tool produced the most accurate global subseasonal forecasts across all variables.
- ▪For forecasts three weeks ahead, the Probabilistic Bias Correction tool significantly improved the European Centre for Medium-Range Weather Forecasts numerical model's predictions of extreme heat, extreme precipitation, and sea level pressure.
Real-world applications and flood forecasting
- ▪The Kenya Meteorological Service Authority and Rhiza Research began testing the Probabilistic Bias Correction tool to improve monsoon and rainy season early warnings ahead of El Niño.
- ▪Testing the Probabilistic Bias Correction tool on historical floods over the prior decade improved flood prediction skill by 76% to 92% overall, and by 86% to 150% for severe floods.
Accessibility and computational requirements
- ▪The lightweight open-source design of the Probabilistic Bias Correction tool enables meteorologists in resource-constrained developing nations to train and configure subseasonal forecasts locally without expensive infrastructure.
- ▪The Probabilistic Bias Correction software runs on central processing units rather than supercomputers or high-power graphics processing units required by physics-based models or generative artificial intelligence.
Challenges of subseasonal weather forecasting
- ▪Traditional subseasonal forecasting models routinely underperform climatological averages for temperature, precipitation, and sea-level pressure outside the tropics due to complex atmospheric and oceanic interdependencies.
- ▪Subseasonal weather forecasting, which predicts weather conditions two to six weeks in advance, suffers from small initial errors that compound over successive simulation time steps.
- ▪Beyond a two-week horizon, small variations in atmospheric conditions cause single deterministic weather forecasts to become unreliable, requiring meteorologists to use probabilistic forecasting ensembles.
Story comments
Loading comments…