Study shows machine learning improves forecasts two to six weeks ahead

Researchers describe a machine-learning correction technique in Nature that doubled a forecasting model's predictive power two to six weeks ahead. The winning entry in an ECMWF competition last year combined ECMWF's model with another model, then applied the correction layer. The technique learns from historical forecasting errors and adjusts the probabilities assigned to possible weather outcomes. Tests also showed statistically significant improvements in predicting extreme weather and better land forecasts across most regions. The resulting model is open source and does not require a supercomputer to run; Kenya's meteorological office has expressed interest in implementing it.