Microsoft Research in Nature: Probabilistic Bias Correction doubles ECMWF AI model skill for 2–6 week forecasts, improves 91–98% of operational targets, and won every variable in ECMWF's 2025 real-time contest; open source

Why it matters

【Key points】

Microsoft CEO Satya Nadella on Oct 16 at 21:17 UTC promoted a Microsoft Research paper published in Nature this week, "Advancing subseasonal forecasting with machine learning" (out Oct 14, with MIT). It introduces Probabilistic Bias Correction (PBC): instead of building another weather model, a lightweight ML system learns the systematic errors in an existing model's past probabilistic forecasts and corrects them, aimed at the 2–6 week "subseasonal" range, the "predictability desert" where both physics and AI models lose skill.

Results: applied to ECMWF models, PBC doubles the subseasonal skill of ECMWF's AI Forecasting System (AIFS) and improves the operationally debiased dynamical model for 91% of pressure, 92% of temperature and 98% of precipitation targets. The PBC-based entry MicroDuet placed first for every variable and lead time in ECMWF's 2025 AI Weather Quest real-time competition, beating six operational centers and 34 teams, and flagged two major US winter storms in early 2026 19–26 days ahead. PBC is open source, Kenya's meteorological service is piloting it this rainy season, and Microsoft Weather's subseasonal model is in private preview.

【Verifiable】

Microsoft Research story: https://www.microsoft.com/en-us/research/story/new-ai-framework-bridges-a-critical-gap-in-mid-range-weather-forecasts/

Nature paper: https://www.nature.com/articles/s41586-026-11126-x

Preprint: https://arxiv.org/abs/2604.16238

Nadella post: https://x.com/satyanadella/status/2111204941801128366

【Our read】

1. The approach matters: rather than competing on a new model, it is a generic post-processing layer that fits any model, which makes operational adoption easier.

2. "Doubled" starts from a low base: the paper calls AIFS's subseasonal skill "modest", and 2–6 week forecasts remain far less reliable than 10-day ones.

3. The contest was a real-time blind test, stronger than hindcasts, but covers one year; the storm calls are case studies needing longer checks.