Chenyu Dong and Gianmarco Mengaldo of the National University of Singapore report that S2S-JEPA scored higher than ECMWF’s extended-range ensemble on week-six probability forecasts for near-surface temperature and precipitation. The comparison covered 102 past forecast starts; ECMWF retained the higher score for pressure at week six and for all three variables at week three.
Subseasonal-to-seasonal (S2S) forecasting looks roughly two weeks to two months ahead. S2S-JEPA uses a joint-embedding predictive architecture (JEPA), which learns to predict a compressed representation of the atmosphere instead of rebuilding every grid detail. The researchers first trained it to predict one week ahead, then repeatedly fed each predicted weekly representation back as the next starting point, advancing through week six. During this second stage, they compared each predicted step with the representation of that observed week.
A decoder translates each predicted representation into a single weekly forecast. For probabilities, a separate diffusion model generates 50 possible outcomes around that forecast. The team turns those members into five probability bins: each bin records what share of the 50 outcomes falls within a temperature, precipitation or pressure range.
The model used 79 atmospheric, land and ocean variables from ERA5 reanalysis at 1.5-degree resolution; training covered 1979–2018, validation 2019–2021. The test used the 102 Wednesday starts shared with ECMWF’s 101-member real-time extended-range forecast, from December 2023 to November 2025. For the fair ranked probability skill score (RPSS), each model’s five ranges came from its own 2004–2023 hindcasts. The score compares probabilities with a uniform baseline that assigns 20% to each range; higher is better. At week six, S2S-JEPA scored 0.043 for temperature versus ECMWF’s 0.031, and 0.015 for precipitation versus 0.005. The paper’s bootstrap test marks both differences as significant; precipitation also scored higher at week five, 0.018 versus 0.008.
For single-value forecasts, the researchers measured temporal correlation of weekly anomalies (TCC). Temperature scores at weeks five and six were 0.238 and 0.230 for S2S-JEPA, above ECMWF’s ensemble-mean scores of 0.202 and 0.181; both gains were significant. Precipitation’s single-value scores remained below ECMWF’s at weeks three through six. ECMWF also retained higher probability scores for pressure through week six, although the week-five and week-six gaps were not significant.
The paper says S2S-JEPA remains substantially less skillful than ECMWF at weeks one to three. Its results are retrospective hindcasts, not live forecasts. The authors report training and inference on one GPU but do not provide a controlled cost-per-forecast comparison. Independent replication and evaluation in live forecast cycles are needed to establish operational value.
Sources
- Chenyu Dong and Gianmarco Mengaldo, “S2S-JEPA: Predicting the Predictable at Subseasonal-to-Seasonal Timescales”, arXiv:2610.03106v1, Oct. 2, 2026. Preprint; CC BY 4.0. Figure 1 is reproduced unmodified under that license.
- ECMWF AI Weather Quest, the benchmark context cited by the paper.
- ECMWF sub-seasonal forecast overview, including its operational forecast horizon and weekly outlook format.
- Creative Commons Attribution 4.0 International, the paper’s declared license.