Subseasonal forecasting (from two weeks to two months) is often called the 'predictability desert': atmospheric chaos erases memory of initial conditions, while slow ocean and land processes have yet to gain strength. This work sits apart from its neighbors: nearby studies predict El Niño a year and a half ahead, teach neural networks to read past weather for drought, and transplant terrestrial meteorological templates to Mars. Unlike them, the focus here is not on a specific event or planet, but on the large-scale circulation regimes themselves, which serve as the skeleton for all weather. The authors test the hypothesis that predictability arises from different components of the Earth system depending on the forecast lead time, and they do so systematically, with one model per predictor.
From the ERA5 reanalysis, they took daily fields of 500 hPa geopotential height over North America for 1981–2020. After filtering and standardization, they extracted 12 principal components and used k-means clustering to split them into four clusters corresponding to regimes: Pacific Trough, Pacific Ridge, Greenland Maximum, and Alaskan Ridge. A fifth class, 'No Regime', collected days when the pattern did not resemble any cluster. They then trained numerous XGBoost models, each receiving only one variable: from the atmosphere (500 hPa geopotential, outgoing longwave radiation, zonal wind at 200 and 10 hPa), from the ocean (sea surface temperature, heat content at various depths, sea surface height, mixed layer depth, sea ice thickness and concentration), or from land (soil moisture and temperature at four depths, snow depth). All data were converted to weekly standardized anomalies, and they predicted the dominant regime of the week at lead times from 1 to 8 weeks. Hyperparameters were tuned via Bayesian optimization with decade-based cross-validation.
The atmosphere provides the greatest predictability in the first two weeks, after which its contribution becomes comparable to that of the ocean and land. The best ocean-based models rely on upper-ocean heat content rather than sea surface temperature, highlighting the importance of subsurface processes. For land, soil temperature and moisture yield significant predictability, while snow is useful only in certain seasons. Seasonal stratification showed that winter and summer have higher predictability than spring and autumn, with the atmosphere and ocean leading in winter, and models distinguishing the 'No Regime' class well in summer. For the Pacific Trough in autumn and winter, the ocean, especially El Niño signals, is key; for the Pacific Ridge in winter, stratospheric winds and sea ice matter; for the Greenland Maximum in winter, tropical convection and soil conditions play a role. SHAP analysis confirmed that accurate forecasts are linked to specific anomalies: for example, enhanced convection over the central tropical Pacific precedes the Pacific Trough, while zonal wind anomalies in the southern tropical stratosphere precede the Pacific Ridge.
This work shows that the sources of predictability for weather regimes are far more variable across lead times, seasons, and regime types than previously thought. Unlike neighboring studies that either focus on a single component (e.g., El Niño or soil moisture) or use a unified model for all predictors, here each source is evaluated separately, revealing its individual contribution. This is the first systematic decomposition of the atmosphere, ocean, and land contributions to predicting North American regimes at subseasonal timescales using gradient boosting. The results emphasize that improving forecasts requires not just adding more data, but understanding which component of the Earth system matters in a given situation.
The topic could evolve toward testing causal relationships: do stratospheric wind anomalies actually cause the Pacific Ridge, or are they a common response to another process? Including data from numerical Earth system models would be a promising way to check whether they reproduce the same predictability sources. It is also important to study how these sources will change under climate warming, as circulation regimes may shift or change frequency.
The results directly affect operational subseasonal weather and climate risk forecasting in North America, as well as related fields such as hydrology, agriculture, and energy, where decisions depend on expected temperature and precipitation anomalies.
Immediate next steps include verifying the identified relationships through causal analysis and model experiments, as well as expanding the predictor set with data on other components, such as the cryosphere or atmospheric composition.
In atmospheric sciences, an open question remains: how far does the predictability of large-scale circulation extend, and what physical mechanisms sustain it? This work directly links predictability to specific processes (El Niño, the Madden–Julian Oscillation, the stratospheric polar vortex, soil moisture), but it does not prove causality. This leaves open the problem of separating true sources of predictability from correlated signals.
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The authors systematically investigated the predictability of four weather regimes over North America at lead times of 1–8 weeks, training hundreds of XGBoost models with single predictors from the atmosphere, ocean, and land. They show that the atmosphere dominates during the first two weeks, after which the contributions of the ocean and land become comparable, with key sources of predictability depending on season and regime.
Where its strength is
A strong point is the systematic comparison of contributions from different Earth system components using a unified methodology, which makes it possible to identify which predictors are most informative for each regime and in which seasons.
Where it could go next
Worth trying
Why this may matter
Understanding the sources of weather regime predictability could improve subseasonal forecasts, which is important for agriculture, energy, and risk management.
Nearby works in our archive
The archive is sparse around this work: the closest review is only 0.53, against a usual 0.64. It looks like few have drilled nearby.
The frequency of the 'No Regime' class varies from 18% in winter to 44% in summer, making summer forecasts particularly difficult: the atmosphere is more often in a transitional state without a pronounced large-scale structure.
In Kim Stanley Robinson's novel 'Forty Signs of Rain', scientists attempt to predict abrupt changes in weather regimes due to a shutdown of the Gulf Stream. This study resonates with that idea: it shows which slow components of the Earth system might serve as early indicators of circulation regime shifts.