Microsoft's PBC machine learning bias-correction method published in Nature boosts 2-6 week forecasts
A Microsoft Research team led by Lester Mackey published a paper in Nature on the Probabilistic Bias Correction (PBC) framework for subseasonal weather forecasting, using a lightweight machine learning method to correct systematic, predictable biases in both physics-based models and AI forecasting systems at the 2–6 week range. The work won a global competition, sweeping the ECMWF AI weather competition; the paper is open access with model downloads available.
Confirmed
- Existing physics and AI models are fairly accurate within 2 weeks, but the 2–6 week range becomes a reliability blind spot due to error accumulation and systematic bias, with reliability dropping sharply after about 10 days
- PBC corrects predictable long-range errors in both physics-based models and AI forecasting systems, as a lightweight ML approach
- Per the post, PBC roughly doubles 3–4 week forecast skill and took first place in a global competition (including the ECMWF AI weather competition)
- The paper was led by former Microsoft Research interns Hannah Guan and Soukayna Mouatadid, with co-authors including Paulo Orenstein; Lester Mackey shared the full author list
- The paper was published open access in Nature, with model downloads provided
Why it matters
- Farmers, resource managers, communities, and disaster responders often need to make critical decisions weeks in advance, and improvements in 2–6 week forecasts directly serve these needs
- The work shows that ML-based bias correction, as a lightweight post-processing step, can meaningfully extend effective forecast horizons without rebuilding forecasting systems
2026-10-14 ~ 2026-10-15 · 10 related posts
Primary sources
- [source] Lightweight ML method fixes predictable errors in long-range weather forecasts — LesterMackey · 2026-10-14
- [source] Microsoft's PBC doubles 3-4 week forecast skill, sweeps ECMWF AI Weather Quest — LesterMackey · 2026-10-14
- [source] Microsoft reveals author lineup behind Nature extended-range weather paper — LesterMackey · 2026-10-14
- Microsoft's PBC framework wins global competition for 2-6 week weather forecasts — LesterMackey · 2026-10-14
- ML subseasonal forecast model published in Nature, open access with downloadable model — LesterMackey · 2026-10-15
- Microsoft Research's Probabilistic Bias Correction wins top honors in global extended weather forecasting competition — LesterMackey · 2026-10-15
- ML subseasonal weather forecasting paper published in Nature — LesterMackey · 2026-10-15
- Microsoft open-sources ML tool that corrects weather model bias, extending reliable forecasts to 2-6 weeks — tomwarren · 2026-10-15
2 near-duplicate retellings: LesterMackey · LesterMackey