Microsoft researchers and collaborators have developed a machine-learning method that learns from errors in existing models to improve weather forecasts two to six weeks ahead. The approach is being positioned as a possible help for farmers who depend on more reliable medium-range weather outlooks.