MIT team doubles accuracy of two-to-six-week weather forecasts using machine learning to correct errors

A research team from MIT published a paper in Nature, proposing a machine learning post-processing layer that learns errors from existing forecast models using historical data and corrects their outputs, doubling forecast accuracy for the two-to-six-week range. The model ranked first in the ECMWF seasonal forecasting competition, becoming the first to consistently exceed climatology averages within this time frame, and significantly improving predictions for extreme weather events such as floods and cold waves; the model is open-source, low-cost, and can be attached to any forecast model, with the Kenya Meteorological Department already in contact to use it for improving monsoon season forecasts.