A new technique significantly improves weather forecasting beyond two weeks. Researchers developed a machine-learning model that learns to correct errors in existing forecasting models to make their output more accurate over the two-to-six-week time frame, a key window for emergency planning.