Researchers including a team at MIT developed a machine-learning method that learns from past forecast errors and applies corrections to existing weather models.
The approach was tested for the two-to-six-week period, a stretch when weather forecasts are often difficult. A submission combining the correction with other forecast models placed first in an international competition run by the European Centre for Medium-Range Weather Forecasts.
MIT reports that the approach doubled the original model's predictive power in this range and produced forecasts that beat climatological averages. The researchers also reported better predictions of extreme weather, without giving a numerical improvement in the university article.
The method is described as an open-source add-on, but MIT says possible use by forecasting offices is still being explored rather than already in routine operation.
