Latent Dynamics · 2026-10-05
S2S-JEPA: Predicting the Predictable at Subseasonal-to-Seasonal Timescales
Chenyu Dong, Gianmarco Mengaldo
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
S2S-JEPA reformulates subseasonal-to-seasonal forecasting (about 2 weeks to 2 months ahead) as latent-space prediction using a JEPA-style joint-embedding objective to focus on slowly varying, predictable structure. The authors report competitive deterministic skill with ECMWF at weeks 5–6 and probabilistic skill improvements for temperature (and precipitation at longer lead times), while training and running on a single GPU.
Why it matters
The authors target the “predictability desert” beyond two weeks by changing the training objective from reconstructing full grid details to predicting latent representations that capture predictability at S2S timescales. They evaluate on weekly-mean T2M, TP, and MSLP against the ECMWF extended-range ensemble, reporting lead-time-dependent gains (weeks 5–6) with significantly better temperature skill and strong longer-range probabilistic performance.
Method
- Forecasts weekly future latent states (for V=79 variables) rather than physical fields directly, using a JEPA joint-embedding loss in latent space.
- Uses two-stage training: one-step latent prediction with an EMA target encoder (stage 1), then multi-step rollout fine-tuning across K=6 weeks with detached hand-offs (stage 2).
- Produces deterministic forecasts via a latent decoder and probabilistic forecasts via a diffusion-based model that samples residuals around the deterministic prediction.
Limitation
The paper does not release code, weights, or demos in the provided text.
Abstract (from arXiv)
The subseasonal-to-seasonal (S2S) timescale, roughly from two weeks to two months ahead, is a critical forecast window for sectors such as agriculture, energy, and water management. Yet, it is widely known as the `predictability desert'. Recent AI weather models excel up to two weeks ahead but deteriorate beyond, largely because they are trained to predict fine-scale details that are neither predictable nor essential at S2S timescales. We argue that a more physically grounded objective is to forecast only the slowly varying components that remain predictable. Computer vision reached the same conclusion with the Joint-Embedding Predictive Architecture (JEPA), which predicts in latent space, discarding unpredictable details. In this work, we introduce S2S-JEPA, which brings the JEPA paradigm to S2S forecasting. It is tailored to this task through design elements from state-of-the-art AI weather models. S2S-JEPA achieves comparable skill to the gold-standard ECMWF physics-based ensemble and surpasses it on multiple metrics at weeks 5 to 6.
Related papers
- Does Joint-Embedding Predictive Architecture Pretraining Help Time Series Forecasting?
- PrecipJEPA: JEPA-Regularized Future-State Prediction with Motion-Source Rendering for Precipitation Nowcasting
- T-JEPA: A Temporal Joint-Embedding Predictive Architecture for Learning Better Remote Sensing Representations
- BeliefGraph-JEPA: Structured Latent World Models for Action-Conditioned Time Series
- Time-JEPA: Grounding Video World Models with Physical Time Series
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-02 · Indexed by Pub-AI: 2026-10-05 UTC
S2S-JEPA: Predicting the Predictable at Subseasonal-to-Seasonal Timescales
Preprint posted: 2026-10-02 · Original record ↗
- arxiv
- 2610.03106
Explicit source versions (1)
- v1 · 2026-10-02
- Preprint available
Source date: 2026-10-02 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
The subseasonal-to-seasonal (S2S) timescale, roughly from two weeks to two months ahead, is a critical forecast window for sectors such as agriculture, energy, and water management. Yet, it is widely known as the `predictability desert'. Recent AI weather models excel up to two weeks ahead but deteriorate beyond, largely because they are trained to predict fine-scale details that are neither predictable nor essential at S2S timescales. We argue that a more physically grounded objective is to forecast only the slowly varying components that remain predictable. Computer vision reached the same conclusion with the Joint-Embedding Predictive Architecture (JEPA), which predicts in latent space, discarding unpredictable details. In this work, we introduce S2S-JEPA, which brings the JEPA paradigm to S2S forecasting. It is tailored to this task through design elements from state-of-the-art AI weather models. S2S-JEPA achieves comparable skill to the gold-standard ECMWF physics-based ensemble and surpasses it on multiple metrics at weeks 5 to 6.
