Published in *Nature*, the method developed by MIT, Microsoft and their partners corrects systematic errors in the models and, for the first time, outperforms climatological averages.
The "sub-seasonal" deadline, of two to six weeks, remains the poor relation of the weather: too far away for initial conditions to suffice, too close for climate trends alone. A team led by Judah Cohen of MIT's Department of Civil and Environmental Engineering, together with Harvard, the University of Toronto, Microsoft and the European Centre for Medium-Range Weather Forecasts (ECMWF), is making a clear improvement.
Correcting rather than replacing
Rather than building a new model, the researchers designed a machine learning post-processing layer that learns from the history of recurring biases in an existing system and then removes them from its forecasts. Applied to the ECMWF model, it doubles the accuracy over this window and becomes the first to consistently outperform climatological averages, especially for extreme events. It won first place in an international forecasting competition.
Described in Nature in October 2026, the method is open source, inexpensive and can be transposed to other systems. Energy network operators and farmers could thus anticipate cold waves, heat waves or rainfall episodes several weeks in advance.
Read the full article on the original source
Go to the source →