Smart Subsampling of Climate Models Boosts Decadal Rainfall Forecasts for France

Predicting how much rain will fall over a specific region several years from now has long been one of the most stubborn problems in climate science. Seasonal forecasts lose their edge after about a year, while long-term climate projections deliberately ignore the year-to-year and decade-to-decade wiggles of natural variability. Yet for water managers, farmers, and energy producers, the decadal horizon is precisely where planning decisions are made. A new study published in Earth System Dynamics by Joanne Couallier of SUEZ and the University of Bordeaux’s EPOC laboratory, together with Didier Swingedouw, Charlotte Sakarovitch, Théodore Raymond, and Ramdane Alkama, now presents a five-step framework that delivers skillful, high-resolution decadal precipitation forecasts over France by exploiting the predictability of large-scale atmospheric and oceanic circulation patterns.

The core insight behind the approach is that regional rainfall is not random noise. It is strongly steered by a handful of recurring pressure patterns over the North Atlantic, most famously the North Atlantic Oscillation, or NAO, which describes the seesaw in sea-level pressure between the Azores High and the Icelandic Low. When the NAO is in one phase, the jet stream and storm tracks shift, drenching some parts of Europe and leaving others dry. Because the ocean evolves far more slowly than the atmosphere, these slowly varying modes carry a degree of memory that initialized decadal climate predictions, which start from the observed state of the ocean and atmosphere and run forward for ten years, can in principle capture. Previous work by Smith and colleagues in 2020 showed that the NAO is far more predictable than climate models superficially suggest, and that selecting model simulations whose NAO matches a skillful NAO forecast can dramatically improve winter rainfall predictions over Europe.

The new study builds on a 2026 refinement by Alkama and colleagues that fixes a subtle but damaging timing flaw in earlier NAO forecasts. When multi-model decadal hindcasts were averaged across consecutive start dates, a systematic two-year lag crept into the predictions. By restricting the averaging to hindcast years that actually fall within each target eight-year window, and then rescaling the forecast variance to correct for the well-known signal-to-noise paradox, in which ensemble means underestimate the true amplitude of climate signals, the boosted winter NAO forecast improved its anomaly correlation coefficient from 0.58 to 0.83, a striking gain in predictive skill.

But the French team did not stop at the NAO. Using empirical orthogonal function analysis and wavelet analysis of the high-resolution SAFRAN observational dataset, which provides precipitation over France on an 8-kilometer grid going back to 1958, they identified the dominant modes of rainfall variability in both winter and summer. Winter precipitation turned out to be governed by two main patterns: a spatially coherent mode explaining just over half the variance, tied to the West Europe Pressure Anomaly index, and a north-south dipole linked to the winter NAO. Wavelet analysis revealed statistically significant periodicities of roughly six, eight, and twenty-one years, hinting at an oceanic imprint on French rainfall. In summer, the leading mode correlated strongly with the Summer North Atlantic Oscillation, while a newly defined Mediterranean-Scandinavia index captured the second dipole mode.

The heart of the method is a subsampling procedure applied to a 32-member uninitialized ensemble of extended historical simulations from the IPSL-CM6A-LR climate model. For each eight-year window, the boosted decadal forecast of the chosen index is compared with the index values simulated by each of the 32 members, and the five members whose large-scale circulation best matches the prediction are selected. These five simulations then provide the precipitation forecast, on the logic that their atmospheric states are dynamically consistent with the predicted phase of the NAO or other index. Crucially, because the members come from an uninitialized ensemble, the method sidesteps the initial drift that plagues directly initialized decadal predictions in their first years. For summer forecasts, the team combined the Summer NAO with an unadjusted Atlantic Multidecadal Variability index, retaining the externally forced warming trend in North Atlantic sea surface temperatures so that the selection could constrain both internal variability and the forced signal.

The results are impressive for winter. Subsampling on the winter NAO produced significant anomaly correlation coefficients across 78 percent of French grid cells, covering nearly all of mainland France except the Mediterranean coast, with a median correlation of 0.54. Residual correlation analysis confirmed that this skill goes beyond what the forced, externally driven response alone can explain, and probabilistic skill scores beat the full uninitialized ensemble over 61 percent of the territory. Around Rouen in northern France, correlations with observed eight-year precipitation anomalies improved from 0.68 to 0.83 compared with the uninitialized ensemble mean. Winter forecasts based on the West Europe Pressure Anomaly were also skillful, though less so, and the authors trace this paradox to a northward shift of the pressure centers in the IPSL model, which pushes the simulated WEPA influence toward Britain and Ireland rather than France.

Summer proved harder, as expected. Summer rainfall over France is shaped more by localized convective processes that coarse-resolution models struggle to represent, and the summer pressure indices fluctuate within a much narrower range of roughly one hectopascal, compared with four for the winter NAO. Subsampling on the Summer NAO alone yielded significant skill over only 12 percent of the country. But adding the Atlantic Multidecadal Variability index as a joint constraint nearly quadrupled the skillful area to 48 percent of France, with a median correlation of 0.32, and delivered significant probabilistic gains over 39 percent of grid cells. In the Rouen region, summer correlations jumped from essentially zero to 0.41. The oceanic index acts as a filter, weeding out ensemble members whose sea surface temperature state is dynamically inconsistent with the predicted phase, even though the index itself is not used as a direct rainfall predictor.

To make the forecasts usable at the local scale, the team applied the CDF-t statistical downscaling method, a quantile-mapping technique that adjusts the distribution of model precipitation to match observations while preserving large-scale variability. This transformed the roughly 150-kilometer model output into bias-corrected forecasts on the 8-kilometer SAFRAN grid, the same resolution used by French hydrological services and impact studies. Skill evaluations against both SAFRAN and the ERA5 reanalysis, using block bootstrap significance testing and false discovery rate correction, showed that the subsampled winter forecasts substantially outperform the raw multi-model DCPP ensemble mean, whose skill over France was largely confined to the Atlantic coast.

The framework is deliberately designed to be transferable. The authors emphasize that the same five-step workflow, identifying relevant indices, boosting their decadal forecasts, subsampling a large ensemble, downscaling, and evaluating skill, could be applied to other regions and other variables, provided suitable large-scale predictors and uninitialized ensembles exist. Sensitivity tests showed the method remains meaningful with five-year rather than eight-year averages, and that five members is an optimal sub-ensemble size balancing correlation against uncertainty estimation. Limitations remain, including the coarse resolution of the uninitialized ensemble, the single-model design, and the weaker predictability of some summer indices, but perfect-prediction experiments suggest that any improvement in index forecasts would translate directly into better regional rainfall predictions.

For a country where rainfall governs hydropower output, agricultural planning, and the management of increasingly stressed water resources, the arrival of skillful decadal precipitation forecasts at municipal scale is more than an academic milestone. It offers water managers and climate services a genuinely new tool: a physically grounded, decade-ahead outlook that captures the rhythm of natural Atlantic variability rather than merely the long march of greenhouse warming. As larger ensembles and improved index predictions become available, this dynamically consistent subsampling strategy could well become a standard component of operational decadal climate prediction across Europe and beyond.

Subject of Research: Decadal precipitation forecasting over France using dynamically consistent subsampling of climate model ensembles based on predicted North Atlantic circulation indices

Article Title: Toward robust fine-scale decadal precipitation forecasts through dynamically consistent subsampling

Article References: Toward robust fine-scale decadal precipitation forecasts through dynamically consistent subsampling. (n.d.). https://doi.org/10.5194/esd-17-1177-2026

Image Credits: AI Generated

DOI: 10.5194/esd-17-1177-2026

Keywords: decadal prediction, precipitation, North Atlantic Oscillation, Atlantic Multidecadal Variability, subsampling, France, climate services, downscaling, sea-level pressure, CMIP6, hydroclimate, water resources