Can machine learning make two-to-six-week weather forecasts useful?
A Nature study found that probabilistic bias correction improved leading dynamical and AI systems and won ECMWF’s 2025 real-time competition. Better forecast scores are promising, but benefits for farms, grids and emergency response have not yet been measured.
The short version
3 key points
- 1.Probabilistic bias correction improved operationally debiased ECMWF forecasts for 91% of pressure, 92% of temperature and 98% of precipitation targets tested in the paper.
- 2.Applied to ECMWF’s AI Forecasting System, the method doubled a modest mean ranked probability skill score of 0.03 across variables and lead times in the 2025 test year.
- 3.The system ranked first across variables and lead times in ECMWF’s 2025 real-time competition, but the study did not measure avoided losses, farm yields, grid reliability or disaster outcomes.
The answer: the forecasts became more skilful, not certain
Researchers report that a machine-learning correction layer improved probabilistic forecasts for weeks three and four, a range often described as a predictability desert. The method beat climatological baselines and, in the European Centre for Medium-Range Weather Forecasts' 2025 real-time AI Weather Quest, ranked first for every tested weather variable and lead time.
The advance is not a claim that weather six weeks away can now be predicted like tomorrow's forecast. The method corrects systematic errors in an existing ensemble and produces probabilities across categories. Skill remained modest in absolute terms, especially for a chaotic atmosphere, and a better probability does not remove uncertainty around a particular farm, city or storm.
For decision-makers, that distinction is central. A slightly better early probability can still be valuable when buying fuel, scheduling irrigation or preparing emergency crews, but only if users receive calibrated uncertainty and know when the signal is too weak to justify action.[1][2]
Probabilistic bias correction learns from past forecast mistakes
The framework, called probabilistic bias correction or PBC, starts with an ensemble of deterministic forecasts. It converts those members into probabilities for climatological bins, then applies two efficient models in parallel. Debias++ learns corrections that vary by location, date and probability bin; Persistence++ blends the forecast with recent observations and seasonal climatology.
The corrected outputs are projected back into valid cumulative probability distributions and averaged. This design aims to alter the centre, spread and shape of the forecast distribution rather than merely shifting one predicted value. Because the layer consumes forecasts from another system, it can be attached to dynamical, AI or hybrid inputs.
The paper focuses on land-surface temperature, land precipitation and mean sea-level pressure on a global 1.5-degree grid. It evaluates week three, days 19 to 25, and week four, days 26 to 32, using ranked probability skill score against climatology. Positive scores mean an improvement over the baseline; negative scores mean the forecast is worse.[1]
The comparison covered archived forecasts and a live competition
For historical evaluation, the team used weekly aggregated ECMWF ensemble forecasts spanning 2015 to 2026 from the international Subseasonal to Seasonal prediction database. The paper reports global results across 2016 to 2024 for raw ECMWF forecasts and their corrections, alongside tests on ECMWF's operational debiasing, its subseasonal AI Forecasting System and a transformer-based hybrid called POET.
PBC improved the operationally debiased dynamical system for 91% of pressure, 92% of temperature and 98% of precipitation targets. On the AI system in 2025, it doubled the mean ranked probability skill score, but the starting average was only 0.03. The relative gain is large while the absolute score reminds readers that this remains a difficult forecasting range.
The real-time 2025 competition adds value because models could not simply be tuned after seeing the evaluation weather. The PBC-based entries outperformed dynamical models from six operational centres, an international multimodel ensemble, ECMWF's AI system and submissions from 34 teams. Competition success is stronger than a retrospective benchmark, though it still evaluates forecasts rather than social outcomes.[1]
Extreme-event results are promising but need operational validation
The paper extends evaluation beyond average conditions. For severe floods in the Global Disaster Alert and Coordination System — events classified orange or red when they involve more than 100 deaths or 80,000 displacements — PBC increased Brier skill over raw ECMWF by 86% to 150%, while operational debiasing matched or degraded the raw system in the cited tests.
That finding suggests conventional corrections can wash forecasts toward the mean and lose information at the extremes. The machine-learning method instead adjusts a full probability distribution. It may therefore preserve rare-event signals that matter to emergency managers.
But the number and geography of severe events, thresholds chosen and decision costs all matter. A higher Brier skill score does not show that an alert reached a vulnerable community, that officials acted, or that losses fell. Prospective pilots should log forecasts, warnings, decisions and consequences so operational value can be tested rather than assumed.[1][2]
Farmers, grids and public agencies need usable probabilities
Two-to-six-week notice can influence seed and fertiliser timing, reservoir releases, wildfire staffing, fuel procurement and maintenance planning. MIT reports that Kenya's meteorological service has expressed interest in applying the model to seasonal rainfall planning. That is an implementation signal, not evidence that the system has already improved Kenyan yields or warnings.
A useful service must translate gridded probabilities into local decisions without hiding uncertainty. Farmers may need thresholds tied to crop stage and cost; grid operators need demand and supply scenarios; emergency planners need false-alarm and missed-event rates. A forecast can be statistically better yet still be unusable if it arrives late, lacks local calibration or creates alert fatigue.
Public agencies should retain human meteorological expertise and compare the corrected output with established systems. The method is an add-on rather than a replacement for physical models, observations and forecaster judgement. Local data gaps and changing climate conditions can also weaken relationships learned from history.[1][2]
Funding and interests need to travel with the forecast
The author group spans Harvard, Toronto, Brazil's Institute for Pure and Applied Mathematics, MIT, Microsoft, ECMWF, DTN and other collaborators. MIT's summary says the work received support from the US National Science Foundation, the Carlos Chagas Filho Foundation for Research Support and the European Union. The paper and code are presented for operational use.
Institutional breadth does not remove the need for independent replication. Several authors work for organisations that build forecasting systems or data services, and deployment may create professional or commercial value. Procurement decisions should therefore require transparent validation on the intended region and task rather than relying on the competition ranking alone.
Open methods and reproducible data paths make that validation more feasible. Meteorological agencies in lower-resource settings will also need compute, observations, training and dependable access to upstream forecasts; a low-cost correction layer cannot by itself close those infrastructure gaps.[1][2]
Evidence limits and what would change the assessment
The study is peer reviewed and combines historical evaluation with a real-time competition, but it still tests a defined set of variables, spatial grids, probability bins and lead windows. Performance can vary by region, season and event type. Reported percentage improvements should not be read as a general doubling of forecast accuracy everywhere.
Confidence would rise with independent replication by national services, multi-year prospective comparisons and calibration checks under climate shifts. Evaluations should report absolute scores and uncertainty, including regions where PBC does not help, rather than relying on aggregate global averages.
The decisive evidence is whether better probabilities improve decisions and outcomes without shifting risk onto people with fewer resources. Trials should measure avoided crop and energy losses, warning reach, false alarms, emergency lead time and distributional effects. Until then, PBC is a strong forecasting result and a credible operational candidate, not proof of safer communities.[1][2]
What this means for people
- Farmers and energy operators could gain more time to act on drought, cold, rainfall and demand risks, but forecasts must be locally calibrated and tied to clear decision thresholds.
- Emergency managers may benefit from earlier probabilities for severe events, while communities still need accessible warnings, trusted messengers and resources to respond.
- Meteorologists remain essential for interpreting uncertainty, checking local conditions and preventing one model score from becoming an automatic public decision.
Global context
The evaluation is global and includes a real-time ECMWF competition, while potential implementation has been discussed with Kenya's meteorological service. Operational capacity, observation networks and decision thresholds vary widely, so global skill does not guarantee equal local benefit.
What the evidence does not yet show
- The method improves probabilistic forecast scores but cannot eliminate atmospheric uncertainty beyond two weeks.
- Large relative gains start from modest absolute subseasonal skill in some comparisons.
- Results cover specified variables, grids, bins and lead windows and may not generalise to every locality or event.
- The paper does not measure warnings, decisions, economic losses, crop yields, grid reliability or deaths prevented.
- Independent operational replication across multiple years and climate regimes is still needed.
What to watch next
- Independent tests by national meteorological services
- Prospective measurement of warning quality and real decisions
- Regional calibration as climate patterns shift
- Whether lower-resource services can adopt the method without new infrastructure gaps
Living evidence record
Impact record IAI-1MNFSQE
Evidence stage
Studied
Confidence
Supported
Reporting basis
Multi-source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
14 October 2026
Source trail
2 direct sources across 2 source types.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Evidence trail
Sources used for this report
Links checked 14 October 2026
Research question, topics and regions
Whether machine-learning post-processing can correct systematic errors in probabilistic weather forecasts two to six weeks ahead
Topics: Weather forecasting · Climate adaptation · Machine learning · Emergency planning · Energy systems · Agriculture
Regions: Global · Europe · United States · Kenya
This report is labelled multi-source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
The Impact Brief · Free
Keep following the stories that matter to you.
Get our weekday briefing on what changed and why it matters. Choose your topics below.
Continue the story
There’s more to this story
Climate & Energy
Can AI forecast Egypt’s groundwater decline?
A peer-reviewed study combines 21 years of satellite water-storage estimates with land-use mapping and XGBoost for Wadi El-Assiuti. The held-out fit is strong, but the study area is smaller than one GRACE cell and the 2030 path assumes recent conditions continue.
9 min · 1 source
Climate & Energy
Can satellite AI show where a town stays cooler?
A peer-reviewed Algeria case study compared four machine-learning methods on seven Landsat 9 scenes. The maps may help frame local heat questions, but the apparent accuracy comes from only 30 held-out proxy points—not field temperatures, human exposure or health outcomes.
10 min · 2 sources
Climate & Energy
Can AI warn of severe storms before radar sees them?
A peer-reviewed East China study combined radar, weather stations and three-dimensional atmospheric forecasts to predict severe convection up to 12 hours ahead. The model improved many rain and reflectivity scores, but extreme gusts, regional transfer and operational warning benefits remain unproven.
11 min · 3 sources
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.