Microsoft PBC Weather Forecasting Improves Models Without Replacing Them

Microsoft has released a machine-learning framework that targets weather’s stubborn two-to-six-week forecasting gap. The Microsoft PBC weather forecasting approach does not replace established models. Instead, it learns from their recurring errors and adjusts the probabilities they produce.

That distinction creates the central tension around the October 14 release. Technology companies have spent years building increasingly ambitious AI weather models. Microsoft is arguing that a lightweight correction layer can extract more value from models that already exist.

The framework is called Probabilistic Bias Correction, or PBC. Microsoft and its research partners published the method in a peer-reviewed study. They also released its implementation as open-source software.

Microsoft says PBC improved forecasts from physics-based, AI, and hybrid systems. Its strongest evidence includes historical evaluations and a year of real-time competition results. Those findings are substantial, but they do not eliminate the underlying unpredictability of weather weeks ahead.

The practical question is therefore narrower than whether AI can “solve” weather forecasting. It is whether better-calibrated probabilities can help people make safer decisions before local details become predictable.

What Microsoft PBC Weather Forecasting Changed

Microsoft’s important move is a new way to improve existing forecasts, not another standalone weather model.

Traditional numerical weather prediction starts with current observations. It then uses mathematical equations to simulate how the atmosphere and oceans will evolve.

Newer AI systems learn relationships from large historical datasets. They can generate forecasts much faster than many physics-based systems. Both approaches still lose useful information as errors accumulate over time.

PBC operates after another model has produced its forecast. It analyzes that model’s historical predictions, identifies systematic errors, and adjusts its future probability distributions.

That last point matters. Ordinary bias correction often adjusts a predicted value, such as expected rainfall. PBC adjusts the probabilities assigned to different possible outcomes.

A forecast might assign probabilities across five historical categories, or quintiles. Those categories range from unusually dry conditions to unusually wet conditions.

The framework receives deterministic forecasts, recent observations, and climatological thresholds. It then generates a calibrated probability distribution for each forecast target.

Microsoft’s researchers built PBC from two complementary components. Debias++ learns corrections tied to location, date, and probability category. Persistence++ incorporates recent conditions and normal seasonal patterns.

The system combines their outputs after ensuring that the resulting probabilities form a valid distribution. That combination seeks to preserve useful signals while reducing predictable errors.

The research summary describes PBC as a lightweight layer compatible with dynamical, AI, and hybrid models. It can also be incrementally retrained as new forecast and observation data arrive.

The researchers tested weekly average temperature, total precipitation, and mean sea-level pressure. Their principal evaluation covered days 19 through 25 and days 26 through 32.

Those periods fall inside subseasonal forecasting, the difficult interval between conventional weather prediction and seasonal climate outlooks. Decisions made at this horizon often concern risks, not exact daily conditions.

A farmer does not need an exact rainfall total five weeks ahead to begin planning. A reliable probability of abnormally dry weather can influence seed, labor, and irrigation decisions.

Emergency managers can use the same logic. An elevated probability of unusual heat, cold, or precipitation supports early preparation, even without a precise local forecast.

Microsoft has also released the open-source code under an MIT license. That allows forecasting organizations to inspect the implementation and test it against their own models.

This openness makes PBC easier to challenge than a closed commercial demonstration. It also lowers the barrier for smaller research teams that already possess archived forecasts and observations.

The release does not give those teams a complete forecasting operation. They still need dependable input models, historical records, computing infrastructure, and meteorological expertise.

PBC changes one part of that system. It offers a reusable method for extracting better-calibrated probabilities from forecast machinery that organizations already trust.

Why Fixing Forecasts Can Beat Replacing Them

PBC shifts the AI weather race from model replacement toward measurable improvements across several forecasting systems.

Subseasonal weather prediction remains difficult because the atmosphere is chaotic. Small errors in a model’s starting state grow as the forecast reaches further forward.

Slower forces can provide some guidance during this period. Ocean temperatures, soil moisture, snow cover, and broad atmospheric patterns all influence future conditions.

However, those signals rarely determine local weather with certainty. A useful system must represent uncertainty instead of producing a single confident outcome.

Microsoft’s method builds around that limitation. It does not claim to recover information that no forecast contains. It searches for useful signals inside forecasts that also contain systematic errors.

The Nature paper evaluates PBC with ranked probability skill score, or RPSS. This metric measures whether a probabilistic forecast improves upon a historical climatology baseline.

A positive score indicates that the forecast adds value over normal conditions for that place and date. A negative score means climatology would have performed better.

Across tests covering 2016 through 2024, PBC tripled the positive precipitation skill of the raw ECMWF dynamical ensemble. It also turned negative global temperature and pressure skill into positive results.

The comparison with ECMWF’s operational debiasing method was more demanding. PBC produced precipitation improvements across 98 percent of evaluated grid cells.

Temperature improved across 91 to 92 percent of cells. Mean sea-level pressure improved across 89 to 91 percent, depending on the forecast window.

The researchers reported temperature skill gains of 26 to 37 percent over the operationally corrected ECMWF forecasts. Pressure gains ranged from 16 to 32 percent.

These percentages describe improvements over modest baseline skill scores. They should not be interpreted as equivalent percentage increases in everyday forecast accuracy.

PBC also improved an AI model rather than only correcting physics-based output. Applied to ECMWF’s subseasonal AI Forecasting System, it doubled the model’s average RPSS of 0.03 during 2025 testing.

That result supports Microsoft’s broader claim that PBC is model-agnostic. A correction layer can remain useful even when the underlying forecast already comes from machine learning.

The method also improved POET, a hybrid system that applies neural networks and transformers to an ECMWF ensemble. POET had therefore already received sophisticated post-processing.

PBC increased POET’s precipitation skill by 68 to 98 percent during the 2024 evaluation. Pressure improved by 15 to 23 percent, while temperature improved by 5 to 7 percent.

These differences reveal an important pattern. PBC delivered its largest relative gains where the input model’s baseline skill remained weakest.

That is useful, but it requires careful interpretation. Doubling a small score can represent genuine progress without creating a dependable local warning.

The framework’s value comes from combining models rather than declaring one universal winner. Forecast centers can preserve trusted dynamical systems while adding learned corrections.

That pathway carries lower operational risk than replacing an entire forecasting pipeline. It also lets teams compare the corrected output directly with their established products.

The Real Contest Is Correction Versus Replacement

Microsoft is challenging the assumption that the best AI weather strategy begins with building a larger prediction model.

Google DeepMind’s GenCast illustrates the replacement side of this contest. GenCast generates probabilistic forecasts by learning global atmospheric behavior from historical data.

Its published probabilistic weather model outperformed ECMWF’s ensemble system across many medium-range benchmarks. The model focuses on forecasts reaching about 15 days ahead.

Other research has pushed machine learning deeper into the subseasonal period. FuXi-S2S uses decades of reanalysis data to predict global conditions several weeks forward.

Published FuXi-S2S results reported improvements over conventional subseasonal models. Its researchers also extended skillful Madden-Julian Oscillation predictions from 30 days to 36 days.

These systems matter because they seek to generate the forecast itself. They compete directly with numerical weather prediction at particular horizons and variables.

PBC takes a different position. It assumes that strong forecast models will continue to coexist, then improves the probability distributions they produce.

This correction strategy offers several practical advantages. Forecast organizations do not need to discard investments in dynamical models, observation pipelines, or expert workflows.

They can also apply the same correction framework to multiple systems. That flexibility matters because no single model performs best for every variable, region, and horizon.

PBC’s MicroDuet entry demonstrated this strategy in ECMWF’s AI Weather Quest. MicroDuet combined PBC-corrected ECMWF forecasts with PBC-corrected POET forecasts.

The real-time benchmark asks teams to submit weekly global predictions. Its main evaluation windows cover days 19 through 25 and days 26 through 32.

The competition attracted more than 250 participants across over 55 teams. Those participants submitted forecasts from more than 100 AI and machine-learning models.

During the initial September through November 2025 season, MicroDuet ranked first across every evaluated variable and lead time. It competed against 34 teams and six operational forecasting centers.

Those centers included ECMWF, NOAA, Environment and Climate Change Canada, and agencies from the United Kingdom, South Korea, and Japan.

Microsoft says MicroDuet also achieved the highest overall results across a full year of real-time forecasting. Real-time performance matters because it reduces the risk of tailoring a model to completed events.

The result still does not settle the correction-versus-replacement contest. MicroDuet relies on the underlying models that replacement systems seek to outperform.

A correction layer cannot create a useful signal when every input misses an event. It can improve calibration only when historical relationships remain relevant.

The strongest interpretation is therefore not that PBC makes large forecast models unnecessary. It makes their errors a reusable source of training data.

That could pressure model developers to treat calibration as a central capability. Raw model performance alone becomes less decisive when an inexpensive post-processing layer produces material gains.

Operational weather agencies face another kind of pressure. They must decide whether established debiasing procedures still justify their place in production.

PBC reportedly outperformed conventional debiasing across most evaluated grid cells. Agencies can now reproduce those comparisons using open code and their preferred observations.

This approach also gives smaller institutions a plausible entry point. Building a global foundation model remains difficult, while correcting an accessible forecast archive is a narrower engineering task.

The result is a more plural weather ecosystem. Physics models, AI models, and hybrid systems remain competitors, yet each can also become an input to the same correction framework.

Strong Benchmarks Do Not Guarantee Local Warning Skill

The headline results are credible evidence of progress, but they do not establish reliable local forecasts for every extreme event.

The first limitation concerns the metric itself. RPSS evaluates probability distributions against climatology across many places and dates.

A model can improve its average RPSS while still missing a dangerous storm. It can also assign better probabilities without identifying an event’s exact timing or location.

Global averages can hide regional weakness. Performance across broad temperature fields may not translate into equally useful predictions for rainfall extremes.

Precipitation is especially difficult because it varies sharply across short distances. Local terrain, convection, and moisture transport complicate predictions weeks in advance.

The Nature evaluation uses a global grid with cells spanning 1.5 degrees of latitude and longitude. That scale is useful for broad anomalies, not neighborhood-level weather.

Its primary targets are weekly averages or totals. Those targets can support planning, but they differ from forecasts for a particular storm on a particular day.

The paper’s historical evaluation also relies on ERA5 reanalysis as ground truth. Reanalysis combines observations with modeling to create a consistent global record.

That dataset is widely used, but it is not identical to every local observation network. Sparse measurements can introduce additional uncertainty in some regions.

The researchers addressed another common weakness through real-time competition. MicroDuet’s results were not limited to a retrospective test selected after outcomes became known.

Even so, one year is a short operational record. Rare events require longer evaluation because the most consequential failures occur infrequently.

The Microsoft PBC weather forecasting story also includes vivid storm examples. Microsoft says MicroDuet anticipated two major United States winter storms 19 to 26 days ahead.

Those cases help explain the framework’s potential. They should not be treated as proof that PBC consistently predicts individual storms at that lead time.

Forecasts weeks ahead usually identify an elevated probability of a broad anomaly. They do not resemble the detailed storm track shown closer to landfall.

Microsoft’s claim that PBC doubled AIFS-SUBS skill needs similar context. The paper reports that the corrected system doubled an average RPSS of 0.03.

That improvement is statistically meaningful within the study. However, the starting score was modest, and forecast usefulness depends on each decision’s costs and thresholds.

An independent expert response published by Spain’s Science Media Centre emphasized this distinction. Doubling limited skill does not automatically produce a useful operational forecast.

That caution does not invalidate PBC. It identifies the next standard that Microsoft and forecasting agencies must meet.

Users need reliability diagrams, regional breakdowns, and performance during high-impact events. They also need comparisons against the forecasts already used in real decisions.

False alarms carry costs. A farmer who changes planting plans based on a forecast can lose income if the predicted conditions never arrive.

Missed events also matter. Emergency agencies may overestimate a system’s value if strong average scores obscure weak performance on dangerous extremes.

PBC’s design is transparent enough to support this scrutiny. The paper explains the inputs, evaluation targets, and component models.

Its code release creates another path for independent testing. Researchers can change regions, observations, variables, and probability bins without relying on Microsoft’s product interface.

That is more valuable than a broad claim of superior accuracy. Replication will show where the framework travels well and where it needs local adaptation.

Kenya’s Rainy Season Is the Practical Test

PBC becomes meaningful when calibrated probabilities change a decision, not when they merely improve a global score.

Microsoft and its partners are testing the framework with Kenya’s national weather service. The work focuses on forecasts during one of the country’s two rainy seasons.

Many smallholder farms in the region depend directly on rainfall. Irrigation is limited, which makes planting and fertilizing decisions sensitive to expected conditions.

Farmers must decide which seeds to buy and when to plant. They also need to weigh the risks of dry spells, heavy rain, and damaging downpours.

These choices match the strengths of probabilistic subseasonal weather forecasts. A weekly probability distribution can inform planning without promising an exact daily outcome.

Rhiza Research, a nonprofit working on AI weather forecasting, is supporting the effort. It collaborates with Kenyan forecasters and research groups in the United States and Germany.

Microsoft says PBC is helping the team assess probabilities for unusually heavy or unusually light rainfall. Forecasters then use those results while preparing weekly updates.

Phillip Okello, a principal meteorologist with Kenya’s weather service, described the difference in concrete terms. He said ordinary methods mainly adjust predicted rainfall amounts.

PBC instead corrects the probabilities themselves. According to Okello, it improved forecasts across the Greater Horn of Africa more than other tested methods.

His next priority also exposes the remaining gap. The forecasts must become sharper for heavy rainfall and dry spells because those outcomes matter most to farmers.

That statement is more informative than a general endorsement. It recognizes meaningful improvement while identifying the hardest unresolved targets.

A forecast’s communication channel matters as much as its model score. Probabilities must reach users early, in language they understand, with guidance tied to actual choices.

A five-week forecast has little value when delivered after seed purchases. A technically accurate probability can also fail when users interpret it as certainty.

This is where the cost argument becomes relevant. Microsoft presents PBC as a computationally efficient framework that builds on existing forecast outputs.

The public repository says a demonstration can run quickly on conventional research hardware. That example does not measure the full cost of an operational forecasting service.

Organizations still need to download data, maintain pipelines, validate forecasts, and communicate them responsibly. Those operational costs can exceed the correction model’s computing requirements.

Yet the lightweight design can still widen access. A national service may adapt PBC without training a global AI weather model from scratch.

Local adaptation also gives meteorologists more control. They can optimize against observations and errors that matter within their own territory.

This model of adoption differs from purchasing a finished global forecast product. It treats local agencies as technical partners rather than passive recipients.

Microsoft has a commercial path as well. Its own subseasonal-to-seasonal model provides weekly probabilistic guidance reaching six weeks ahead.

The company says that model runs in real time and is entering private preview with selected customers. Microsoft Weather already supplies forecasts across several Microsoft products.

That creates a useful tension between open research and commercial delivery. PBC’s code is public, while a complete operational service includes proprietary data, infrastructure, and support.

Independent deployment will show how much value resides in the framework itself. Commercial adoption will test whether organizations will trust it for consequential planning.

Kenya’s pilot is therefore more than a showcase. It asks whether improved calibration survives contact with sparse data, local extremes, and human decision-making.

What to Watch Next for Two-to-Six-Week Forecasts

The next evidence should come from independent replication, extreme-event performance, and sustained operational use.

The first signal is replication outside Microsoft’s research network. Forecasting agencies and university teams can now run the published code against their own archives.

Researchers should report both successes and failures. Regional comparisons will reveal whether gains remain broad across different climates and observation systems.

The most persuasive tests will use held-out dates and real-time forecasts. That design limits the opportunity to tune the framework after seeing the weather.

Results should also compare PBC with each organization’s current bias-correction process. Beating an uncorrected model is useful, but operational debiasing is the stronger baseline.

Consistent independent gains would strengthen Microsoft’s central argument. Mixed regional results would suggest that PBC requires more local tuning than its broad design implies.

The second signal is extreme-event skill. Average temperature, pressure, and precipitation scores cannot fully represent the risks facing farmers and emergency managers.

Evaluators should separate ordinary anomalies from dangerous tails. They should measure heavy rainfall, persistent drought, severe cold, and prolonged heat.

Those studies need to report false alarms alongside successful warnings. A system that detects more hazards by constantly raising probabilities may create little practical value.

Lead time also needs careful handling. A useful week-three signal may disappear by week four, while different variables can decay at different rates.

Microsoft PBC weather forecasting will be most consequential if it improves rare-event probabilities without destroying calibration. That standard is harder than raising an average global score.

The third signal is sustained operational adoption. Microsoft’s private preview will show whether energy, agriculture, logistics, and public agencies can incorporate the output into routine decisions.

Adoption should not be measured by customer announcements alone. The stronger evidence will be repeated use, documented forecast verification, and decisions linked to defined probability thresholds.

Kenya offers an early test of that progression. Forecasters are already evaluating PBC during a live rainy season and identifying where its predictions need greater sharpness.

ECMWF’s continuing competition provides another public test. MicroDuet must maintain its position as new teams, new AI models, and additional forecast periods enter the benchmark.

Competitors will also adopt similar post-processing ideas. If correction becomes standard, PBC may influence forecasting even when another implementation eventually performs better.

That outcome would confirm the strategic shift behind Microsoft’s release. The important contest would no longer be physics against AI or Microsoft against Google.

It would become a contest over complete forecasting systems. Those systems combine observations, dynamical models, learned models, calibration layers, and expert communication.

Readers should therefore resist both easy conclusions. PBC has not solved the fundamental unpredictability that limits long-range weather forecasts.

It has produced peer-reviewed, reproducible evidence that existing models leave correctable information on the table. Its real-time competition results make that evidence harder to dismiss.

The next question is practical: which decisions improve when these probabilities reach people two to six weeks before the weather arrives?

Researchers can test the code against local archives. Forecast providers can publish regional and extreme-event results. Decision-makers can define what accuracy would justify action.

That process will determine whether Microsoft PBC weather forecasting becomes a widely used operational layer or remains an impressive research benchmark.