金融研究arXiv cs.LG7 Oct, 12:00 pm SGTSIGNAL 9481C8

Stock-JEPA: Prior-Anchored Latent Revision Representation Learning in Equity Markets

ORIGINAL / STOCK-JEPA: Prior-Anchored Latent Revision Representation Learning in Equity Markets

This study proposes Stock-JEPA, a joint-embedding predictive framework that uses multi-horizon return and risk statistics from a low-complexity financial model as anchors in the target encoder's latent space, and estimates predictable displacement from the anchor via a context-conditioned revision predictor. Its novelty lies in constraining black-box representation learning with structural priors while retaining an interpretable reference; the authors report that on China and

01 ABSTRACT

Fact: Stock-JEPA adopts a joint-embedding predictive framework: a low-complexity financial model generates multi-horizon return and risk statistics, a prior projector maps them as anchors into the target encoder's latent space, and a context-conditioned revision predictor estimates predictable displacement of the future representation from the anchor; separate losses update the anchor and revision branches, and a downstream readout is trained on frozen representations and evaluated through cross-sectional ranking and portfolio performance. The authors theoretically prove that the optimal revision reduces the prior anchor's expected squared error for the same future representation by exactly the expected squared norm of the revision vector. Author's view: this design combines the interpretability of classic linear financial models with deep models' ability to capture non-linear signals, and experiments show it outperforms 13 strong baselines on large-scale China and U.S. equity universes, with ablation and representation analyses demonstrating the value of learned revisions for equity market representation learning.

02 KEY FINDINGS

Stock-JEPA uses multi-horizon return and risk statistics from a low-complexity financial model as fixed priors, projected into the target encoder's latent space as anchors.
A context-conditioned revision predictor estimates predictable displacement of the future representation from the anchor; the anchor and revision branches are updated with separate losses, and a downstream readout evaluates cross-sectional ranking and portfolio performance on frozen representations.
The authors prove that the optimal revision reduces the prior anchor's expected squared error by the expected squared norm of the revision vector, and report outperformance over 13 strong baselines on five metrics across China and U.S. equity universes.

03 RELEVANCE

Equity market data have a low signal-to-noise ratio; black-box deep models can overfit noise and lack explicit economic structure, while classic linear financial models are interpretable but fail to capture non-linear signals. Stock-JEPA offers a path to combine structural financial priors with deep representation learning, potentially improving representation robustness and interpretability and providing a reference for feature construction and portfolio evaluation in quantitative research.

04 EVIDENCE BOUNDARY

Evidence boundary: the source is only an arXiv abstract, which does not specify the 13 baseline names, the definitions of the five evaluation metrics, the sample period for the two equity universes, transaction costs, price limits and liquidity constraints, data leakage controls, or statistical significance tests and multiple-comparison corrections. The theoretical proof concerns the reduction in expected squared error for the same future representation, and its non-negativity comes from the expected squared norm of the revision vector; it does not directly guarantee improved portfolio returns or risk-adjusted performance. The experimental conclusions are based on the authors' chosen China and U.S. equity universes without independent replication, and cross-market, cross-period stability, capacity constraints, and trading feasibility are not discussed. Downstream evaluation uses frozen representations plus a readout, so the influence of readout capacity and training details on results is not ruled out. The framework depends on a low-complexity financial model to produce priors, and prior selection bias and parameter sensitivity are not addressed in the abstract.

Return to the primary source

AI GENERATED SUMMARY / DISCOVERED BY ARXIV CS.LG

Read original ↗