Microsoft Research built a machine-learning framework called Probabilistic Bias Correction. It learns from past errors in weather models to improve forecasts. PBC won a global competition for sub-seasonal forecasting. It also doubled the accuracy of ECMWF’s AI Forecasting System. PBC improved 98% of precipitation, 92% of temperature, and 91% of pressure targets. It anticipated major U.S. winter storms weeks in advance. PBC is now being tested to improve monsoon forecasts for Kenyan farmers. It helps smallholder farms make planting and fertilizing decisions. The framework works on any existing weather model. It does not replace models but corrects their biases.
What changed
Sub-seasonal weather forecasts were less reliable between two and six weeks.
What it unlocks
Making more accurate two-to-six-week weather forecasts for farming, disaster response, and resource management.
- 98% of precipitation targets improved
- 92% of temperature targets improved
- 91% of pressure targets improved
- 2x accuracy for ECMWF’s AI Forecasting System
Sources