- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
Stock-JEPA learns equity representations relative to a simple financial prior. A point-in-time Ridge model predicts return and risk statistics, which are projected into a latent anchor. A context encoder then predicts a revision to that anchor using historical observations, with separate losses and EMA targets for the two branches. A frozen representation is evaluated through a return readout. The paper gives a conditional-expectation argument for optimal squared-error reduction with a fixed target representation, and evaluates the method on broad Chinese and U.S. equity panels. It reports the highest mean RankIC and net Sharpe among the compared methods in both markets, with readout and representation ablations and a controlled signal-recovery experiment.
Critical strengths
-
The prior-relative objective is a sensible way to bring financial structure into predictive representation learning. The two branches have clear roles, and the separate losses prevent the revision loss from directly changing the anchor. The theoretical result is a standard conditional-projection identity rather than a new generalization guarantee, but the paper explains its scope carefully: the target encoder is fixed in the analysis, and lower latent prediction risk need not imply better trading performance. This makes the motivation easier to assess.
-
The empirical support goes beyond the main comparison table. In the matched MLP ablation, adding revisions improves RankIC over state, prior, and anchor inputs by 2.02 and 3.67 percentage points in China and the U.S.; revisions also beat equally sized context features (Appendix I). The controlled simulation provides useful evidence of complementary signal recovery, with explicit caveats about probe capacity and variation across worlds. I also appreciate the chronological splits, target-window purging, inclusion of delisted securities, next-open execution, and detailed accounting for overlapping portfolios. These experiments make the representation contribution reasonably convincing.
Critical weaknesses
-
The headline baseline comparison does not fully isolate the representation-learning contribution. Appendix E trains the baselines on 2014–2021, while Appendix F.2 fits the test-period financial prior using eligible history available before that period. As written, the latter can include 2022–2023 outcomes that the baseline models do not use for fitting. That is a legitimate point-in-time procedure, but it gives the proposed pipeline a more recent supervised component. The JEPA baselines also use linear readouts, whereas Stock-JEPA uses an MLP with additional prior and state inputs. The matched readout and Ordinary-JEPA controls address part of this concern, and I regard them as stronger evidence than the large gaps to the JEPA baselines in Table 1. However, Appendix J compares fixed-budget continuations from the same S2 checkpoint. It does not establish the same ordering when each mechanism is trained from scratch on real data. A common prior-update schedule and matched nonlinear readouts would make the main comparison much more persuasive. The incremental contribution of separate objectives and EMA is also less uniform than the overall pipeline gain: Joint-online has the best CN RankIC and US IC in Table 6.
-
The robustness of the reported gains needs clearer quantification. Three optimization seeds describe training variability, but do not measure uncertainty across market histories. Consecutive 21-day targets and overlapping portfolios are temporally dependent, so paired date-level tests that preserve this dependence are needed to assess the margins over the strongest baselines. The market-state breakdown is useful, but still reuses the same test period. The portfolio results also rely on assumptions that matter for the broad U.S. universe. Appendix G explicitly omits stock-borrow availability and fees, funding, market impact, and execution restrictions, while using a fixed 5-bps charge. I appreciate that these limitations are disclosed; nevertheless, net Sharpe here is net of the modeled trading fee, not all implementation costs. Liquidity/borrowability filters and cost sensitivity would help show whether the result is driven by difficult-to-trade names. The complete-future-window filtering in Appendix B also needs to be reconciled with point-in-time portfolio eligibility, especially near delistings or data gaps.
Questions for authors
-
What exact date cutoff is used to fit the prior supplied to the 2024–2026 test period? Does it include 2022–2023 outcomes? If so, how do the strongest baselines perform with a comparable pre-test update, or how does Stock-JEPA perform with the prior frozen at the end of 2021? For JEPA comparisons, please also match the MLP readout and access to state/prior features.
-
Does the Separate-EMA advantage persist when the real-data representation variants are trained independently from scratch, rather than continued from a common S2 checkpoint? Please report paired uncertainty for the RankIC and portfolio differences using date blocks long enough to account for the overlapping 21-day horizon. This would clarify which gains are consistent across methods and which are small relative to variation.
-
Is portfolio eligibility determined entirely from information available at the forecast date, independently of whether a complete finite 21-day target will later exist? Please explain the handling of delistings and interrupted price histories. How sensitive are the U.S. results to a liquid, borrowable universe and higher trading/borrow costs?
4 comments
Response to Reviewer m3na
Q3. Does Separate-EMA remain effective when all variants are trained independently from scratch?
- To test the training schemes under independent initialization, we train all four variants from scratch with matched data splits, training budgets, and downstream evaluation settings.
| Market | Training scheme | RankIC (%) | Net annualized return (%) | Net Sharpe |
|---|---|---|---|---|
| CN | Separate-EMA | |||
| CN | Separate-online | |||
| CN | Joint-online | |||
| CN | Ordinary-JEPA | |||
| US | Separate-EMA | |||
| US | Separate-online | |||
| US | Joint-online | |||
| US | Ordinary-JEPA |
Results are mean ± standard deviation. All four variants in each market are newly trained independently from scratch.
- Separate-EMA achieves the highest mean in four of the six market–metric combinations. Its CN/US RankIC gains are 1.87/4.03 percentage points over Ordinary-JEPA and 1.77/2.14 percentage points over Separate-online. Against both variants, all three metric means are higher in both markets. Relative to Joint-online, mean net annualized return is higher by 9.46/7.55 percentage points.
- These comparisons support the combined prior-relative design and the incremental value of EMA within the separate-objective scheme. The two elements serve complementary roles: separate objectives prevent direct revision-loss updates to the anchor, keeping its task focused on prediction from the prior; EMA smooths the anchor-projector updates and provides a gradually updated reference for revision learning.
Response to Reviewer m3na
Q4. How do paired uncertainty estimates account for overlapping 21-day horizons?
- To complement the results with uncertainty across test dates, we use 10,000 paired studentized stationary-bootstrap replications for Stock-JEPA minus the original JEPA baselines (Table 1). All methods and seeds share the resampled date indices. Metrics are computed separately for each seed and then averaged. SE is the HAC standard error estimated with 21 lags.
| Market | Baseline | RankIC (HAC SE) | 95% CI | Net Sharpe (HAC SE) | 95% CI |
|---|---|---|---|---|---|
| CN | T-JEPA | 9.30 (1.74) | [5.31, 13.29] | 1.15 (0.42) | [0.23, 2.07] |
| CN | TS-JEPA | 7.13 (2.09) | [3.06, 11.19] | 1.29 (0.43) | [0.38, 2.20] |
| CN | Fin-JEPA | 9.61 (1.69) | [5.69, 13.52] | 0.84 (0.33) | [0.08, 1.59] |
| US | T-JEPA | 16.73 (2.26) | [11.41, 22.05] | 1.16 (0.52) | [−0.06, 2.37] |
| US | TS-JEPA | 13.46 (1.53) | [10.61, 16.30] | 1.54 (0.43) | [0.76, 2.31] |
| US | Fin-JEPA | 12.27 (1.37) | [9.48, 15.05] | 1.61 (0.54) | [0.38, 2.84] |
RankIC differences are in percentage points; Sharpe differences are dimensionless.
- All six RankIC differences have pointwise 95% confidence intervals above zero; five of the six Net Sharpe differences also have intervals above zero. These results provide statistical support for consistent ranking gains and broad portfolio-performance improvements over the original JEPA baselines.
Q5. How are portfolio eligibility and delistings handled, and how sensitive are results to trading constraints and costs?
- Regarding these practical constraints, thank you for highlighting their implications for portfolio management. They are important considerations for live trading and the eventual deployment of our work. Our paper focuses on a method for financial representation learning. Developing and evaluating portfolio management under these constraints is an important next step for deeper application, and we regard it as a future research direction.
We would appreciate it if you could confirm whether your concerns have been addressed. We would be happy to engage in further discussion.
Does our response address your concerns?
Dear Reviewer m3na,
Thank you again for your detailed feedback on information alignment, independent training, and the robustness of the evaluation.
In our responses, we clarified the prior-fitting cutoff and provided matched JEPA comparisons using the same updated prior, historical summaries, and MLP readout settings. We also trained all four representation variants independently from scratch and reported paired uncertainty estimates against the original JEPA baselines while accounting for temporal dependence. In addition, we discussed practical portfolio implementation constraints as a future research direction.
Could you please take a look at our responses when you have a chance? We would appreciate your feedback on whether the additional experiments and explanations address your concerns, and particularly on any remaining points that would benefit from further clarification. If your concerns have been addressed, we would be grateful if you could reconsider your assessment. We would be happy to continue the discussion.
Best regards,
Authors of Submission 20922
Response to Reviewer m3na
Thank you for recognizing the prior-relative formulation and the supporting representation analyses. We have added matched comparisons, independently trained variants, and paired uncertainty estimates.
Q1. What is the prior’s fitting cutoff, and how is information recency matched before testing?
Q2. Are the MLP readout and state/prior inputs matched across JEPA methods?
Results are mean ± standard deviation over three runs.
[1] Wang. Fin-JEPA: Joint-embedding predictive representation learning for financial time series. SSRN. 2026.
[2] Ennadir et al. Joint embeddings go temporal. NeurIPS 2024 Workshop. 2024.
[3] Thimonier et al. T-JEPA: Augmentation-free self-supervised learning for tabular data. ICLR. 2025.