Microsoft/MIT/ECMWF Nature paper: ML bias correction sharply improves 2-6 week forecasts

Microsoft Research, MIT, and ECMWF jointly published a study in Nature (Guan et al., Nature 658, 951–957) applying machine-learning-based probabilistic bias correction to subseasonal weather forecasts of 2-6 weeks, achieving record accuracy gains in this notoriously difficult middle range and improving predictions of extreme weather events.

Confirmed

• The method is not a new forecast model but a post-processing layer on existing numerical systems: AI learns the forecast system's systematic biases from historical data, reasons about causal relationships, and continuously calibrates outputs against real-world data — essentially building an "intelligent verification system."
• Nature's research briefing notes the tool delivers accurate forecasts of precipitation, temperature, and pressure in the most challenging subseasonal window, 3-5 weeks ahead, while also improving physical model outputs.
• Posts claim the work breaks through Lorenz's classic predictability limit, with accuracy multiplying versus prior levels.
• Microsoft CEO Satya Nadella shared the study, saying it could help tackle extreme weather, manage natural resources, and inform agricultural decisions; he also argued future gains will come not just from stronger models but from the verification and calibration systems around them.
• The authors note the 2-6 week range has long been a blind spot for traditional forecasting, yet reliable predictions there have real value for emergency planning (the post cites a 2021 emergency scenario as an example).

Why it matters

• Subseasonal forecasting has long been capped by the predictability limits of chaotic systems; this work shows "post-processing + AI calibration" can significantly extend effective forecast horizons without rebuilding forecast models, opening a weeks-early decision window for extreme weather response, agriculture, and natural resource management.
• Nadella's remarks suggest Microsoft sees "intelligent verification around models" as a key direction for AI in science.

2026-10-22 ~ 2026-10-22 · 5 related posts

• [source] AI tool extends weather forecasts to 3–5 weeks ahead, Nature paper reports — LesterMackey · 2026-10-22
• [source] Microsoft, MIT and ECMWF's Nature paper doubles 2-6 week forecast accuracy with open-source CPU bias correction — LesterMackey · 2026-10-22
• [source] Nadella touts Microsoft's Nature paper on ML forecasts for the 2-6 week window — LesterMackey · 2026-10-22
• Microsoft's Nature research uses ML to sharpen weather forecasts 2-6 weeks out — LesterMackey · 2026-10-22
• MIT ML technique breaks records for 2-6 week weather forecasts by learning to correct model errors — LesterMackey · 2026-10-22