Microsoft's Probabilistic Bias Correction wins ECMWF AI Weather Quest

Weather forecasting is highly dependable out to about two weeks, and longer-term climate forecasts are increasingly accurate, but the two-to-six-week sub-seasonal window has remained a stubborn gap. Both physics-based models, which simulate the atmosphere with scientific equations, and newer AI models trained on large datasets do well out to roughly 10 days. Beyond that, compounding errors and the complex interactions between global earth systems and local weather cause both to struggle.

Rather than build another competing model, Microsoft and external collaborators focused on correcting predictable errors in existing models. A paper in Nature outlines how PBC boosted the performance of ECMWF's leading operationally debiased sub-seasonal model for 98% of precipitation, 92% of temperature, and 91% of pressure targets, and doubled the accuracy of ECMWF's AI Forecasting System (AIFS). In the AI Weather Quest, the MicroDuet entry applied PBC to ECMWF's dynamical model and to PoET, a hybrid model that uses transformers to improve probabilistic forecast ensembles. MicroDuet anticipated two brutal U.S. winter storms 19-26 days in advance and also flagged a mid-December cold air outbreak that the leading debiased dynamical model missed.

In East Africa, PBC is being piloted with the Kenya Meteorological Service Authority as one of Kenya's two rainy seasons gets underway. Phillip Okello, a KMSA principal meteorologist, said PBC corrects the probabilities themselves rather than just adjusting how much rain a forecast predicts, and that across the Greater Horn of Africa it improved forecasts more than any other method tested. The next step, he said, is sharpening forecasts of heavy rain and dry spells, which matter most to farmers. With PBC now open-source and Microsoft Weather's S2S model in private preview, the framework's reach beyond the lab appears to be widening.