AI forecasting system aims to extend reliable weather predictions to six weeks
A new forecasting approach uses artificial intelligence and postprocessing to improve predictions in the subseasonal range of two to six weeks. The method challenges the long-held two-week predictability limit linked to chaos theory by learning from past errors and adjusting future forecasts. Its developers submitted multiple versions to the European Centre for Medium-Range Weather Forecasts for evaluation.
Meteorologists have long treated about fourteen days as a practical ceiling for reliable forecasts. That boundary traces to Edward Lorenz’s 1970s work on chaos, which showed small atmospheric uncertainties can grow until they swamp the original signal. The two-to-six-week stretch, called subseasonal, has been especially difficult because slow, variable phenomena such as a strong El Niño can disrupt predictions.
Judah Cohen and Lester Mackey built an AI method that uses postprocessing: it compares forecast misses with historical records and corrects later output. Their three submissions to ECMWF’s seasonal contest reportedly doubled accuracy in that window, and a Nature study described gains for extreme events too.
If such accuracy holds in operational use, longer subseasonal outlooks could help farmers plan planting or harvest, energy providers manage demand, and emergency managers prepare for floods or cold snaps. Communities exposed to extreme weather may gain extra lead time, though benefits would depend on how widely the method is adopted and whether its performance remains consistent across regions. It may complement, rather than replace, existing forecasting systems.