MIT's ML post-processing doubles forecast skill at two-to-six weeks, published in Nature

Probabilistic bias correction (PBC) is a machine-learning post-processing layer that a joint team from MIT, the European Centre for Medium-Range Weather Forecasts (ECMWF) and Microsoft published in Nature on October 14, 2026, in a paper titled “Advancing subseasonal forecasting with machine learning”. Wrapped around an existing forecast model, PBC learns the model’s systematic errors from history and adjusts its output as new observations arrive. The paper reports it doubles the mean ranked probability skill score (RPSS) of ECMWF’s AI Forecasting System across the two-to-six-week “subseasonal” range, to 0.03 — a band long nicknamed the “predictability desert”, where forecasting has usually lost to guessing. Both the paper and the code are open.

Key points

• PBC is a wrapper, not a retrain: it learns from historical errors, adapts as new data arrives, was trained on MIT’s Engaging cluster, and runs without a supercomputer — the repo README records the full conda environment installing in under 3 minutes on an 8-core, 80 GB machine.
• The paper’s numbers: AIFS subseasonal mean RPSS doubled to 0.03; against the de-biased dynamical ECMWF model, 92% of temperature, 91% of pressure and 98% of precipitation grid-point targets improved.
• The team’s MicroDuet entry won ECMWF’s AI Weather Quest for September–November 2025, first on every variable and lead time among 34 teams — while independent experts caution that the gains concentrate in the tropics and precipitation improvements start from a near-zero baseline.

Background

The two-to-six-week stretch is where weather forecasting has historically given up: initial-condition errors amplify through atmospheric chaos until physical models lose to climatology — forecasting the historical average for the date and calling it a prediction. MIT research scientist and co-author Judah Cohen put it bluntly to MIT News: “in most cases forecasting in the subseasonal ranges is no different than going to a casino and rolling the dice.”

The author list spans MIT, Harvard, the University of Toronto, Brazil’s IMPA, ECMWF, the data company DTN and Microsoft, with Hannah Guan and Soukayna Mouatadid as equal-contribution lead authors and Lester Mackey as senior author. AI-assisted science keeps landing on journal front pages this month — DeepMind’s AlphaProof Nexus just solved nine open Erdős problems, published in Science, and Yann LeCun’s H-JEPA world-model paper surfaced days ago. PBC sits in a different niche: a cheap error-correction layer bolted onto models that already run — a better set of tires for the same car.

The facts

Quantitative results below come from the Nature paper (published October 14, 2026, open access; checked October 15):

Target	Variable	Paper’s result	Absolute baseline
AIFS-SUBS (AI model)	all variables, all lead times	mean RPSS doubled	to 0.03 (2025)
De-biased dynamical ECMWF	temperature	skill +26–37%, 92% of targets improved	RPSS 0.07–0.10
De-biased dynamical ECMWF	sea-level pressure	skill +16–32%, 91% of targets improved	RPSS 0.06–0.11
De-biased dynamical ECMWF	precipitation	positive skill roughly tripled, 98% of targets improved	RPSS 0.023
Extremes	heat, floods	extreme-high temperature skill +51%; flood skill +76–92%, +86–150% severe	—

Verification used the ERA5 reanalysis (1940 to present, regridded to 1.5°). In the AI Weather Quest, MicroDuet — an ensemble of PBC-ECMWF and PBC-POET — took first overall in the September–November 2025 season among 34 teams and was best on every variable at every lead time, ahead of six operational centres including ECMWF, NOAA and the Japan Meteorological Agency. It also won the December–February season among 37 teams, the only entry to beat the dynamical multi-model ensemble on all variables.

Team claims, per the MIT News release: PBC is the first model to consistently beat climatological averages at this range, and extreme-event prediction improved significantly — whereas ECMWF’s own non-AI bias correction made extreme-event prediction worse. Kenya’s meteorological office has approached the team about forecasting the annual monsoon to plan agriculture, and the release points to the February 2021 Texas winter storm — at least 246 deaths and the worst energy-infrastructure failure in state history — as the kind of event a few weeks of warning could have softened by trucking in natural gas in time.

What others say

Science Media Centre España collected comments from four independent experts on publication day; all credit the method, each with a different reservation.

• Froila M. Palmeiro (IGEO, CSIC-UCM): subseasonal prediction with traditional physical models is “currently not very feasible”; the study doubles the performance of the AI-based forecasting system and “improves it by more than 90% in traditional physical models” — but the gains concentrate in the tropical belt and remain “quite limited in Europe”.
• Darío Redolat (Foundation for Climate Research): PBC corrects both the most likely outcome and the uncertainty, updating without rerunning the full model — but the method “remains in the experimental phase”, tested mostly on historical reconstructions, and the large percentage improvements deserve caution.
• Robert Monjo (FIC’s director of research and innovation): post-processing cannot fix a base model that fails in rare situations; the validation treats ERA5 as ground truth even though its precipitation is simulated rather than observed, and doubling precipitation skill from a near-zero baseline “resembles dividing by zero” — a doubled skill is not yet a useful forecast.
• Matthew Chantry (ECMWF’s strategic lead for machine learning): turning an idea into an operational process is the most challenging step — while declaring a conflict of interest, since the corrected AI model was developed by his employer.

Practitioners confirming the value while independent experts hold back is the same rhythm we saw when 950 Claude agents reported a CRISPR-like discovery and researchers asked for the experiments first.

How to try it

The code lives on GitHub at mouatadid/pbc (archived on Zenodo, DOI 10.5281/zenodo.22238796), with the POET baseline model at ecmwf-lab/poet. Per the README, one conda command builds the environment (conda env create -f src/setup/environment-aiwqd.yml) — it installed in under 3 minutes on an 8-core, 80 GB Rocky Linux machine — and the repo ships a demo script that generates global climatology forecasts for mean sea-level pressure. The paper itself is open access on Nature’s site.

Our take

For anyone planning on weekly horizons — energy dispatch, planting schedules, disaster preparation — the news is that a range everyone had written off shows usable signal again. The layer is cheap, open source, and bolts onto models that already run, which makes it a realistic option for national weather services that will never own a supercomputer. Kenya asking about monsoon forecasts is the tell.

The honest catch has three layers: an RPSS of 0.03 is still low in absolute terms, and “doubled” is measured from a near-zero base; the gains concentrate in the tropics, so European readers benefit least for now; and a competition win is not an operational system — ECMWF’s own machine-learning lead calls that step the hardest. Before writing “doubles” into any decision, check the variable, the region and the baseline.