Machine learning sharpens forecasts weeks ahead

A machine-learning method called Probabilistic Bias Correction (PBC) aims to improve weather forecasts two to six weeks ahead by learning from past errors in existing models. Microsoft Research says a Nature paper describes the approach and reports improved performance across 98% of tested precipitation targets, 92% of temperature targets and 91% of pressure targets for an operational forecast model. ^[2]

Microsoft also reports that PBC doubled the accuracy of the European Centre for Medium-Range Weather Forecasts’ AI Forecasting System. The company says a PBC-based system placed first across variables and lead times in an extended-range forecasting competition; these results do not establish that the method will improve every forecast or perform equally well in all regions. ^[2]

The Nature article could not be accessed for independent verification, so the reported results here rely on Microsoft Research’s account. The company describes early real-world testing in Kenya, while noting that further work is needed, including improving forecasts of heavy rain and dry spells. ^[2]