Microsoft and MIT unveil AI methods to sharpen two-to-six-week forecasts
Archived — this story has rotated out of today’s deck. It is kept here in full.
The gist
Microsoft Research and MIT researchers unveiled AI techniques that improve two-to-six-week weather forecasts.
The window matters for farming, wildfire and energy planning, where current models fall short.
Background
Microsoft Research is Microsoft's research organization, and MIT is the Massachusetts Institute of Technology; both had teams working on the two-to-six-week forecasting gap. Today's forecasts are highly dependable out to about two weeks, and longer-term climate forecasts are increasingly accurate, but models struggle in the sub-seasonal window between two and six weeks because of compounding errors and complex interactions between global earth systems and local weather. Better predictions in that window would help with farming, managing wildfires, allocating water and energy, and preparing for weather extremes. The work is tied to the European Centre for Medium-Range Weather Forecasts (ECMWF), the organization behind a leading operational sub-seasonal model and the AI Forecasting System (AIFS), which launched a forecasting competition in 2025.
How it unfolded
- 2025 -08ECMWF kicked off the AI Weather Quest, a real-time competition aimed at improving probabilistic sub-seasonal forecasting.
- Oct 14, 2026Microsoft Research announced Probabilistic Bias Correction (PBC), a framework for mid-range weather forecasts.
- Oct 15, 2026MIT News reported a machine-learning technique, published in Nature, that improves two-to-six-week forecasts.
Who’s saying what
- Microsoft
- Microsoft Research said PBC learns how a chosen model tends to err and corrects those errors, boosting ECMWF's operationally debiased sub-seasonal model for 98% of precipitation, 92% of temperature and 91% of pressure targets and doubling the accuracy of ECMWF's AI Forecasting System.
- Microsoft
- Microsoft Research said its MicroEnsemble team's PBC-based entry, MicroDuet, placed first for every weather variable and lead time in the inaugural three-month season, beating six leading operational forecasting centers and systems from 34 teams worldwide.
- Expert
- MIT News said the machine-learning model, applied to a gold-standard forecasting model, doubled forecasting accuracy over the two-to-six-week time frame and placed first in a prominent international forecasting competition.