Correcting the probabilities in an existing weather forecast can improve its outlook three to four weeks ahead. In a Nature paper published on 14 October 2026, Guan and colleagues describe a machine learning approach that learns from past forecast errors and adjusts the distribution of possible outcomes.

For the AI forecasting system AIFS-SUBS, the authors report that the correction doubled mean probability skill in their 2025 evaluation, from a rounded baseline score of 0.03. That is a gain from a modest starting point on a scale measured against historical climate patterns. It does not mean the system became twice as likely to predict the weather correctly.

What a forecast three to four weeks ahead predicts

The study examines a part of forecasting called subseasonal: the outlook several weeks into the future. Its week-three target covers days 19–25 after the forecast starts; week four covers days 26–32.

Each target summarizes a whole week. The researchers evaluate average near-surface air temperature, total precipitation and average sea-level pressure on a global grid with 1.5-degree spacing. These are broad geographic outlooks, not predictions of the weather on a particular afternoon in a particular neighbourhood.

Temperature and precipitation are evaluated only in grid cells that are at least half land. The precipitation evaluation also excludes very dry cells where even the boundary below the wettest fifth of historical outcomes is zero. The stated land restriction does not apply to pressure.

The paper’s forecasting setup sorts outcomes into five equally frequent historical categories, or quintiles. For temperature, for example, those categories run from the coolest fifth to the warmest fifth of the local historical distribution. The forecast assigns probabilities to them.

That makes an outlook useful for asking whether a week is more likely than usual to be warm or wet.

How Debias++ and Persistence++ adjust the probabilities

An ensemble forecast supplies multiple possible weather outcomes. Their distribution provides a way to estimate probabilities, including how much uncertainty surrounds the forecast.

The study’s probabilistic bias correction, or PBC, works on that distribution. A forecast can have recurring errors in its central estimate, its spread or its shape. Correcting only the average can leave useful information in the rest of the distribution untreated.

As described in the paper’s mechanism and Figure 1, PBC combines two approaches:

  • Debias++ learns adjustments from historical probabilistic forecast errors. Those adjustments vary with location, calendar date and position within the distribution.
  • Persistence++ blends the forecast with earlier observations and climatology, the historical climate reference. This lets recent conditions and the time of year contribute to the outlook.

The method ensures the adjusted cumulative probabilities form valid distributions, then averages the corrections. It can change both which outcomes look most likely and how probability is distributed around them.

The underlying weather model still produces the atmospheric forecast. PBC adds a learned correction to its probabilities.

What “doubling skill” means when the baseline is 0.03

The headline AIFS-SUBS result comes from a 2025 evaluation covering 179 target dates. In Figure 3 and the accompanying analysis, the authors report a doubling of mean ranked probability skill score, or RPSS, across the evaluated variables and lead times from a rounded baseline of 0.03.

RPSS compares a forecast’s probability errors with those of a climatological reference. Here, that reference assigns equal probability to each of the five historical categories.

The underlying ranked probability score assesses errors in cumulative probabilities across the ordered categories. RPSS expresses improvement relative to the reference:

  • Zero means the forecast matches climatology under the score.
  • A positive score means it performs better.
  • A negative score means it performs worse.
  • One represents perfect probabilities under the score.

A baseline of 0.03 therefore puts the reported doubling in perspective: the initial mean skill was only a little above the reference. Doubling that modest advantage can be scientifically useful while leaving substantial uncertainty.

Averaging skill scores across targets does not necessarily produce the same result as pooling their underlying errors before calculating a score.

The verification reference was ERA5 reanalysis, which combines observations with modelling to reconstruct past weather. Each grid value should therefore be understood as a reanalysis estimate rather than a direct local measurement.

Which evaluations support the result

The paper includes several distinct evaluations. Its retrospective analysis of ECMWF dynamical forecasts covers 2016–2024. The AIFS-SUBS result above belongs to the separate 2025 evaluation.

The authors also describe real-time competition evidence for MicroDuet, a combination of corrected ECMWF and POET systems. That is evidence about a different combined system; it does not turn the AIFS-SUBS evaluation into a prospective test.

The results also contain limits. Figure 2 shows that improvements were not universal across locations. In a component analysis, Persistence++ alone did not significantly improve operationally debiased ECMWF pressure forecasts at week four. That finding concerns one component and forecast target, rather than the full correction method across all targets.

An aggregate improvement cannot establish that every region, season or future model version will benefit. Higher RPSS also does not, on its own, establish that all reported probabilities are well calibrated, meaning that outcomes occur as often as their assigned probabilities suggest.

How a weekly outlook could inform a decision

A hypothetical energy planner could use an increased probability of a warmer-than-usual week to prepare alternative demand scenarios. The appropriate action would depend on local forecast performance, the cost of preparing and the cost of being wrong.

That is a possible use of a probabilistic weekly outlook. The study evaluates predictive scores; it does not establish changes in users’ decisions, costs, crop yields or lives saved. Before using a corrected outlook for a consequential decision, a user would still need evidence for the relevant location, target and decision threshold.