acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

Learning effective representations helps characterize the structure and dynamics of equity markets from financial data with a low signal-to-noise ratio. Black-box deep models can capture complex patterns but may overfit sample noise and lack explicit economic structure. Meanwhile, classic linear financial models provide interpretable references, but their oversimplified assumptions leave non-linear signals uncaptured. To combine the strengths of these two directions, we propose Stock-JEPA, a joint-embedding predictive framework that learns predictable incremental revisions relative to a point-in-time financial prior. First, we leverage a low-complexity financial model to produce fixed statistics summarizing multi-horizon return and risk. A prior projector then maps these statistics into the target encoder's latent space as an anchor. Second, we design a context-conditioned revision predictor to estimate the future representation's predictable displacement from the anchor. Separate losses update the two branches: the anchor learns from prior statistics, while the revision captures additional predictable information from historical context. Third, we freeze all representation modules and train a downstream readout, evaluating its forecasts through cross-sectional ranking and portfolio performance. Theoretically, we prove that optimal revision reduces the prior anchor's expected squared error for the same future representation by exactly . This nonnegative gain is the expected squared magnitude of the additional signal predictable from historical context. Experimentally, Stock-JEPA outperforms 13 strong baselines across large-scale China and U.S. equity universes on 5 key evaluation metrics. Ablation studies and representation analysis further demonstrate the value of the learned revisions for representation learning in equity markets.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Financial representation learning, Joint-embedding predictive architectures, Structured financial priors

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_m3naOpenReview ↗
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
  1. 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.

  2. 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
  1. 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.

  2. 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
  1. 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.

  2. 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.

  3. 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
AuthorsOpenReview ↗

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?

  • To clarify the timing, neural weights and the initial S1 prior are fitted using only pre-2022 training data, while 2022–2023 data are used for neural checkpoint selection. The pre-test Ridge refit uses eligible outcomes fully realized before 1 January 2024, including eligible 2022–2023 outcomes. The updated prior and selected neural checkpoints then remain fixed throughout testing.
  • A practical advantage of our modular design is that it can incorporate recent, already realized return and risk information through a Ridge refit, without retraining the representation networks. In the new matched experiments reported below, every JEPA baseline’s MLP readout receives exactly the same updated financial-prior statistics as Stock-JEPA. These outputs come from the same pre-test prior fitted on the same eligible historical pool, including fully realized 2022–2023 outcomes. The matched baselines therefore share the same information cutoff and updated prior outputs, extending this recent-information benefit to all compared methods.
  • The strongest non-JEPA baselines do not use the same modular prior-update mechanism. Incorporating recently realized financial information without retraining the representation networks is a practical advantage of the complete Stock-JEPA pipeline. Freezing the prior at the end of 2021 would remove this update benefit, while supplying our prior to other methods would extend this component of our design to them. In the matched JEPA experiments, we explicitly share the updated prior to assess the additional contribution of representation learning under common prior information.

Q2. Are the MLP readout and state/prior inputs matched across JEPA methods?

  • Using the same pre-test prior described above, we conduct the Fin-JEPA [1], TS-JEPA [2], and T-JEPA [3] experiments using the same training setup as Stock-JEPA. Each baseline is trained independently from random initialization. We then freeze its encoder and train an MLP readout with the same hidden width and downstream training protocol. Each readout receives its own model’s representation together with the same historical summaries and financial-prior statistics used by Stock-JEPA.
Model / configurationCN RankIC (%)CN Net SharpeUS RankIC (%)US Net Sharpe
Stock-JEPA (original)11.22±0.2411.22 \pm 0.241.50±0.281.50 \pm 0.2812.96±0.4812.96 \pm 0.481.24±0.201.24 \pm 0.20
Fin-JEPA (matched)9.01±1.269.01 \pm 1.260.79±0.340.79 \pm 0.349.53±1.209.53 \pm 1.200.76±0.600.76 \pm 0.60
TS-JEPA (matched)8.53±1.558.53 \pm 1.551.06±0.261.06 \pm 0.2610.19±0.4810.19 \pm 0.480.67±0.860.67 \pm 0.86
T-JEPA (matched)7.17±0.057.17 \pm 0.050.46±0.110.46 \pm 0.114.37±5.884.37 \pm 5.880.51±0.280.51 \pm 0.28

Results are mean ± standard deviation over three runs.

  • Stock-JEPA achieves the highest mean on all four metrics. Its CN/US gains over the strongest matched baseline in each column are 2.21/2.77 percentage points in RankIC and 0.44/0.48 in Net Sharpe. The advantage therefore persists when the baselines receive the same prior information and historical summaries under matched training settings and MLP readouts.

[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.

AuthorsOpenReview ↗

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.
MarketTraining schemeRankIC (%)Net annualized return (%)Net Sharpe
CNSeparate-EMA11.82±0.3011.82 \pm 0.3055.77±8.2855.77 \pm 8.281.626±0.0941.626 \pm 0.094
CNSeparate-online10.05±1.7310.05 \pm 1.7344.99±7.2944.99 \pm 7.291.262±0.1351.262 \pm 0.135
CNJoint-online11.46±0.6111.46 \pm 0.6146.31±0.7346.31 \pm 0.731.682±0.0261.682 \pm 0.026
CNOrdinary-JEPA9.95±0.689.95 \pm 0.6842.58±7.1142.58 \pm 7.111.177±0.1851.177 \pm 0.185
USSeparate-EMA12.64±1.0512.64 \pm 1.0567.23±8.5567.23 \pm 8.551.462±0.1851.462 \pm 0.185
USSeparate-online10.50±1.0410.50 \pm 1.0442.55±15.0542.55 \pm 15.050.887±0.4340.887 \pm 0.434
USJoint-online12.90±1.3812.90 \pm 1.3859.68±19.6759.68 \pm 19.671.342±0.4191.342 \pm 0.419
USOrdinary-JEPA8.61±0.888.61 \pm 0.8822.10±23.0522.10 \pm 23.050.483±0.3010.483 \pm 0.301

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.
AuthorsOpenReview ↗

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.
MarketBaselineΔ\Delta RankIC (HAC SE)95% CIΔ\Delta Net Sharpe (HAC SE)95% CI
CNT-JEPA9.30 (1.74)[5.31, 13.29]1.15 (0.42)[0.23, 2.07]
CNTS-JEPA7.13 (2.09)[3.06, 11.19]1.29 (0.43)[0.38, 2.20]
CNFin-JEPA9.61 (1.69)[5.69, 13.52]0.84 (0.33)[0.08, 1.59]
UST-JEPA16.73 (2.26)[11.41, 22.05]1.16 (0.52)[−0.06, 2.37]
USTS-JEPA13.46 (1.53)[10.61, 16.30]1.54 (0.43)[0.76, 2.31]
USFin-JEPA12.27 (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.

AuthorsOpenReview ↗

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

Reviewer_1h3kOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

STOCK-JEPA combines financial forecasting priors with JEPA based representation learning for stock return prediction. Previous work has applied JEPA to market-state representations and individual stock time series, while other approaches have explored financial priors and residual learning. STOCK-JEPA combines these ideas through two branches: an anchor that predicts a future latent representation from financial forecasts, and a history-conditioned branch that learns a revision to that anchor. The learned features are then frozen and passed to an MLP to predict returns on 21 day horizons.

The main contribution is the use of separate training objectives for the anchor and revision branches. The paper also derives the reduction in optimal squared prediction error from conditioning on additional information. Experiments on Chinese and U.S. equities report improvements in RankIC and net Sharpe, alongside ablations + synthetic experiments.

Critical strengths
  • The revision features appear to contain useful predictive information. With matched downstream readouts, adding them improves mean RankIC by 2.02 % points in China and 3.67% in the U.S. Appendix I also compares against direct context features, which helps assess whether the learned representation adds anything beyond the original inputs.
  • The evaluation covers two equity markets and includes a reasonable range of financial and representation-learning baselines, ablations, and synthetic experiments. The synthetic results are useful for understanding what each branch might be learning.
  • Separating the financial prior from the history-conditioned revision is a sensible design choice. It makes the two components easier to study individually and gives some interpretability to what would otherwise be a single learned representation.
Critical weaknesses
  • The headline comparisons are not fully controlled The main JEPA baselines use linear readouts, while STOCK-JEPA uses an MLP with additional historical-summary and financial-prior inputs. The financial prior at test time also appears to incorporate more recent labeled history than some of the baseline parameter estimates. These differences make it hard to tell how much of the reported gain comes from the proposed method rather than the readout, inputs, or training setup. I'd like to see comparisons with matched information and model capacity.
  • It's unclear whether learning the revision in latent space is necessary The theoretical result is essentially a property of conditional expectation: adding information cannot increase the minimum achievable squared prediction error. This is true regardless of JEPA or the anchor/revision architecture. It therefore doesn't explain why a learned latent revision should outperform simpler approaches. In particular, I'd like to see comparisons against a direct prior-conditioned JEPA and a model that learns residuals in return space, both trained from scratch under the same conditions. The controlled continuation experiments in Appendix J are helpful, but they don't quite answer this question.
  • The novelty is somewhat limited. Prior-guided representation learning, residual learning, and latent prediction have all been explored before. The contribution here seems to be the way these pieces are combined and trained with separate objectives. That's a reasonable contribution, but the paper could do a better job explaining how it differs from closely related work such as FactorGCL, and which design choices actually matter.
Questions for authors
  • What specifically is gained by predicting a revision in latent space rather than directly learning a nonlinear correction to the financial prior in return space (or some sort of simple equivalent)?
  • What are the paired confidence intervals for RankIC and portfolio-performance differences when accounting for dependence across test dates?
5 comments
AuthorsOpenReview ↗

Response to Reviewer 1h3k

Thank you for the constructive feedback. We clarify information timing, matched comparisons, and model size separately, and provide additional mechanism controls and paired uncertainty estimates.

Q1. Does the test prior use more recent labeled history than the baselines?

  • To clarify the information used at each stage, neural weights and the initial S1 prior are fitted using only pre-2022 training data, while 2022–2023 data are used for neural checkpoint selection. Before testing, the Ridge prior is refitted using eligible outcomes fully realized before January 2024, including those from 2022–2023. The updated prior and selected neural checkpoints then remain fixed throughout testing.
  • A practical advantage of our modular design is that it can incorporate recent, already realized return and risk information through a Ridge refit, without retraining the representation networks. In the new matched experiments reported below, every JEPA baseline’s MLP readout receives exactly the same updated financial-prior statistics as Stock-JEPA. These outputs come from the same pre-test prior fitted on the same eligible historical pool, including fully realized 2022–2023 outcomes. The matched baselines therefore share the same information cutoff and updated prior outputs, extending this recent-information benefit to all compared methods.
AuthorsOpenReview ↗

Response to Reviewer 1h3k

Q2. How do the JEPA baselines perform with matched inputs and MLP readouts?

  • With prior information aligned as described above, we conduct the Fin-JEPA [1], TS-JEPA [2], and T-JEPA [3] experiments using the same training setup as Stock-JEPA. Each baseline is trained independently from random initialization. We then freeze its encoder and train an MLP readout with the same hidden width and downstream training protocol. Each readout receives its own model’s representation together with the same historical summaries and financial-prior statistics used by Stock-JEPA.
Model / configurationCN RankIC (%)CN Net SharpeUS RankIC (%)US Net Sharpe
Stock-JEPA (original)11.22±0.2411.22 \pm 0.241.50±0.281.50 \pm 0.2812.96±0.4812.96 \pm 0.481.24±0.201.24 \pm 0.20
Fin-JEPA (matched)9.01±1.269.01 \pm 1.260.79±0.340.79 \pm 0.349.53±1.209.53 \pm 1.200.76±0.600.76 \pm 0.60
TS-JEPA (matched)8.53±1.558.53 \pm 1.551.06±0.261.06 \pm 0.2610.19±0.4810.19 \pm 0.480.67±0.860.67 \pm 0.86
T-JEPA (matched)7.17±0.057.17 \pm 0.050.46±0.110.46 \pm 0.114.37±5.884.37 \pm 5.880.51±0.280.51 \pm 0.28

Results are mean ± standard deviation over three runs. Stock-JEPA retains the original manuscript results; the matched baselines are newly trained.

  • Stock-JEPA achieves the highest mean on all four metrics. Its gains over the strongest matched baseline in each column are 2.21/2.77 percentage points in RankIC and 0.44/0.48 in Net Sharpe on CN/US. Its advantage therefore persists under matched training settings, auxiliary information, and MLP readouts.
  • Furthermore, Appendix I holds the prior, frozen representation checkpoint, and MLP training protocol fixed. Adding revision features to the historical summaries, prior, and anchor improves RankIC by 2.02/3.67 percentage points on CN/US, supporting the incremental predictive utility of the learned revisions.

[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.

AuthorsOpenReview ↗

Response to Reviewer 1h3k

Q3. How does Stock-JEPA compare in model capacity?

  • To make model size transparent, we preserve the baseline architectures and use the same MLP hidden width. The neural inference parameter counts are approximately 1.014M for Stock-JEPA and 0.706M, 1.771M, and 1.404M for Fin-JEPA, TS-JEPA, and T-JEPA, respectively. These counts exclude the shared fixed prior estimator. The two mechanism controls introduced below additionally match Stock-JEPA’s neural inference parameter count at 1.014M.

Q4. Why learn latent revisions rather than use direct prior conditioning or return-space correction?

  • To test this directly, we add two controls trained from scratch. Prior-conditioned JEPA predicts the complete future latent from history and the prior, without anchor subtraction or a separate anchor loss. Return-residual learns a nonlinear correction to the prior’s 21-day return forecast, training the correction network and return head end-to-end. Both use the same inputs, prior, data splits, and principal training schedule as Stock-JEPA.
MarketModelRankIC (%)Net annualized return (%)Net Sharpe
CNStock-JEPA11.22±0.2411.22 \pm 0.2445.66±5.1245.66 \pm 5.121.500±0.2761.500 \pm 0.276
CNPrior-conditioned JEPA10.61±1.4910.61 \pm 1.4939.12±16.0039.12 \pm 16.001.254±0.3511.254 \pm 0.351
CNReturn-residual7.19±2.517.19 \pm 2.5137.28±4.4837.28 \pm 4.481.019±0.0351.019 \pm 0.035
USStock-JEPA12.96±0.4812.96 \pm 0.4868.20±15.9268.20 \pm 15.921.240±0.2001.240 \pm 0.200
USPrior-conditioned JEPA10.96±0.3710.96 \pm 0.3745.91±10.7645.91 \pm 10.760.802±0.2260.802 \pm 0.226
USReturn-residual6.24±2.686.24 \pm 2.6833.37±12.5733.37 \pm 12.570.601±0.3240.601 \pm 0.324

Results are mean ± standard deviation. Stock-JEPA retains the original submission results; the controls are newly trained.

  • Stock-JEPA achieves higher means on all three metrics in both markets. Its CN/US RankIC gains are 0.61/2.00 percentage points over prior-conditioned JEPA and 4.03/6.72 percentage points over return-residual learning.
  • These controls test direct prior conditioning and direct correction of a scalar return forecast. Our design first learns a vector-valued representation of future information and then fits the return readout. The results support the explicit anchor–revision decomposition. The conditional-expectation analysis characterizes ideal incremental prediction for a fixed future representation; these controls assess the learned design.
Reviewer_CkGyOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

STOCK-JEPA learns future stock representations through a financial-prior anchor and a history-conditioned revision, trained with separate objectives. The resulting frozen features improve stock ranking and portfolio performance on broad Chinese and U.S. equity datasets.

Critical strengths
  1. The anchor/revision decomposition is intuitive, and the separate objectives give the two branches clear roles. The matched readout ablations provide useful evidence that revisions add value beyond the prior and direct context features.
  2. Evaluation covers two large equity universes。 The improvements in mean RankIC and net Sharpe across both markets are encouraging.
Critical weaknesses
  1. I am not fully convinced that the comparisons isolate the proposed mechanism. The main JEPA baselines use linear readouts, while STOCK-JEPA uses an MLP with additional prior and historical-summary inputs. This leaves the advantage under matched training from scratch unresolved.
  2. The financial prior is central to the argument, yet its validation is limited. The implementation uses multi-output Ridge forecasts, and the paper does not systematically test alternative or degraded priors. It remains unclear how much the gains depend on this particular prior.
Questions for authors

Do the gains persist under matched baseline settings and with different or weaker financial priors?

4 comments
AuthorsOpenReview ↗

Response to Reviewer CkGy

Thank you for highlighting the importance of matched comparisons and prior robustness. We have conducted additional experiments to address both concerns.

Q1. Do the gains persist with matched inputs and readouts when trained from scratch?

  • To compare the representations under matched readouts and auxiliary information, we conduct the Fin-JEPA [1], TS-JEPA [2], and T-JEPA [3] experiments using the same training setup as Stock-JEPA. Each baseline is trained independently from random initialization. We then freeze its encoder and train an MLP readout with the same hidden width and downstream training protocol. Each readout receives its own model’s representation together with the same historical summaries and financial-prior statistics used by Stock-JEPA.
Model / configurationCN RankIC (%)CN Net SharpeUS RankIC (%)US Net Sharpe
Stock-JEPA (original)11.22±0.2411.22 \pm 0.241.50±0.281.50 \pm 0.2812.96±0.4812.96 \pm 0.481.24±0.201.24 \pm 0.20
Fin-JEPA (matched)9.01±1.269.01 \pm 1.260.79±0.340.79 \pm 0.349.53±1.209.53 \pm 1.200.76±0.600.76 \pm 0.60
TS-JEPA (matched)8.53±1.558.53 \pm 1.551.06±0.261.06 \pm 0.2610.19±0.4810.19 \pm 0.480.67±0.860.67 \pm 0.86
T-JEPA (matched)7.17±0.057.17 \pm 0.050.46±0.110.46 \pm 0.114.37±5.884.37 \pm 5.880.51±0.280.51 \pm 0.28

Results are mean ± standard deviation over three runs. Stock-JEPA retains the original manuscript results; the matched baseline rows report the new experiments.

  • Stock-JEPA achieves the highest mean on all four metrics. Relative to the strongest matched baseline for each metric, its RankIC is higher by 2.21/2.77 percentage points and its Net Sharpe by 0.44/0.48 on CN/US.
  • This result is consistent with the original readout experiments. The S1+S2 variant uses a linear readout and achieves RankIC of 9.94%/11.89% and Net Sharpe of 1.10/1.00 on CN/US (Table 2 and Appendix H). Appendix I further fixes the prior, frozen representation checkpoint, and MLP training protocol. Adding revisions to the historical summaries, prior, and anchor improves RankIC by 2.02/3.67 percentage points. Together, these comparisons support the incremental predictive utility of the learned revisions.

[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.

AuthorsOpenReview ↗

Response to Reviewer CkGy

Q2. Do the gains persist with alternative or weaker financial priors?

  • To test dependence on the particular prior, we replace the default multi-output linear Ridge model with OLS and a simpler prior based on the previous trading day’s return rank. The remaining model design and evaluation protocol are unchanged.
Stock-JEPA priorCN RankIC (%)CN Net SharpeUS RankIC (%)US Net Sharpe
Ridge (original)11.22±0.2411.22 \pm 0.241.50±0.281.50 \pm 0.2812.96±0.4812.96 \pm 0.481.24±0.201.24 \pm 0.20
OLS11.09±0.2411.09 \pm 0.241.41±0.321.41 \pm 0.3211.25±1.7111.25 \pm 1.710.91±0.230.91 \pm 0.23
Previous-day return rank10.01±0.6610.01 \pm 0.661.16±0.051.16 \pm 0.0510.11±0.3810.11 \pm 0.381.03±0.291.03 \pm 0.29

Results are mean ± standard deviation over three runs.

  • With OLS, Stock-JEPA exceeds all three matched JEPA baselines in mean performance on all four metrics. With the previous-day return-rank prior, it remains ahead on both CN metrics and US Net Sharpe, while US RankIC is close to the strongest matched baseline.
  • As complementary evidence, the original S2-only variant removes the prior branch and achieves RankIC of 6.21%/7.92% and Net Sharpe of 0.85/0.81 on CN/US (Table 2 and Appendix H), exceeding the originally reported JEPA baselines in mean performance (Table 1). Together, these results show that the representation learner captures useful information from history and can benefit from different financial references.

We would appreciate it if you could confirm whether your concerns have been addressed and, if so, reconsider your assessment. We would be happy to engage in further discussion.

AuthorsOpenReview ↗

Does our response address your concerns?

Dear Reviewer CkGy,

Thank you again for your valuable feedback on matched comparisons and the robustness of the financial prior.

In our responses, we added JEPA comparisons with independently trained representations, matched MLP readouts, and the same historical-summary and financial-prior inputs. We also evaluated alternative priors based on OLS and previous-day return ranks. These experiments directly examine whether the gains persist under matched settings and beyond the default Ridge prior.

Could you please take a look at our responses when you have a chance? We would appreciate your feedback on whether the additional evidence addresses your concerns and whether any points need further clarification. If your concerns have been addressed, we would be grateful if you could reconsider your assessment. We would be happy to engage in further discussion.

Best regards,

Authors of Submission 20922

AuthorsOpenReview ↗

Urgent request for reviewer's reply on our responses (Deadline on Nov 18)

Dear Reviewer CkGy,

As the rebuttal deadline is fast approaching, we would greatly appreciate a reply to our responses before the discussion closes. We have spent significant amount of time on carefully going through your feedback, revising the paper and responding to questions. We summarize major points as follows:

  • Matched comparisons trained from scratch. We independently trained Fin-JEPA, TS-JEPA, and T-JEPA under the same training setup as Stock-JEPA, then used matched MLP readouts and identical historical-summary and financial-prior inputs. Stock-JEPA achieves the highest mean RankIC and net Sharpe in both markets. Against the strongest matched baseline for each metric, the RankIC gains are 2.21/2.77 percentage points and the net Sharpe gains are 0.44/0.48 on CN/US.

  • Alternative financial priors. We evaluated OLS and a Ridge prior using only the previous trading day’s return rank. With OLS, Stock-JEPA exceeds all three matched JEPA baselines on all four reported mean metrics. With the rank-based prior, it leads on both CN metrics and US net Sharpe, while US RankIC remains close to the strongest matched baseline (10.11% versus 10.19%). These results support the effectiveness of the framework beyond the default prior specification.

These results are included in Table 3 of the revised manuscript, with implementation details in Appendices B and E.

Could you please review this additional evidence and let us know whether any concerns remain before the discussion closes? If these results address your concerns, we would be grateful if you could reconsider your assessment. We would be happy to clarify any remaining points.

Best regards,

Authors of Submission 20922

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response and Updates for Revision

We thank all reviewers for their valuable feedback. Based on their comments, we have added controlled experiments and clarified the experimental protocols. The revised manuscript has been uploaded, with new and revised text highlighted in blue. We summarize the major updates below:

  • We add matched comparisons with Fin-JEPA, TS-JEPA, and T-JEPA, each trained independently from random initialization. Their MLP readouts use the same hidden width, downstream training protocol, historical summaries, and financial-prior statistics as Stock-JEPA. Stock-JEPA retains the highest mean RankIC and net Sharpe in both markets (lines 486–504, Table 3, and Appendix E). [Reviewers CkGy, 1h3k, m3na]

  • We evaluate alternative financial priors using OLS and a Ridge model whose sole input is the previous trading day’s return rank. These experiments examine sensitivity to regularization and input information, supporting the effectiveness of prior-relative learning beyond the default prior specification. Implementation details are provided in Appendix B (lines 506–510 and Table 3). [Reviewer CkGy]

  • We add Prior-conditioned JEPA and Return-residual controls, trained from scratch, to compare explicit latent revision learning with direct prior conditioning and return-space correction. Stock-JEPA achieves higher mean RankIC and net Sharpe than both alternatives in both markets (lines 512–518, Table 3, and Appendix J.1). [Reviewer 1h3k]

  • We extend the mechanism analysis by independently training Separate-EMA, Separate-online, Joint-online, and Ordinary-JEPA from random initialization under a common protocol and matched downstream evaluation. Separate-EMA achieves the highest mean in four of six market–metric combinations and exceeds Separate-online and Ordinary-JEPA on all six (Appendix J.2 and Table 9). [Reviewer m3na]

  • We add paired uncertainty estimates accounting for temporal dependence, using 10,000 studentized stationary-bootstrap replications, an expected block length of 42 trading days, and HAC standard errors with 21 lags. Against the original JEPA baselines, all six RankIC differences and five of six net Sharpe differences have pointwise 95% confidence intervals above zero (lines 521–532, Table 4, and Appendix G). [Reviewers 1h3k, m3na]

  • We clarify the temporal separation of prior estimation, neural-network training, and validation-based checkpoint selection. Appendix F.2 specifies that the test prior uses only outcomes fully realized before January 1, 2024, remains fixed throughout testing, and supplies identical updated prior outputs to the matched JEPA baselines. [Reviewers 1h3k, m3na]

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.