acceptodds
Under review as a conference paper at ICLR 2027

Dynamic Cross-Market Systematic Attention for Stock Prediction

Abstract

Stock price prediction is challenging due to the extremely low signal-to-noise ratio in financial markets, where stock returns reflect both market-wide systematic effects and stock-specific dynamics. Recent machine learning approaches have increasingly explored cross-stock relationships for prediction, but often overlook the underlying market-wide systematic effects and their potential interactions across different markets. To address these limitations, we propose a model-agnostic dynamic cross-market systematic attention method that improves existing stock prediction models with dynamic latent systematic factors and cross-market factor interactions. The method learns market-specific latent factors that adapt to changing market conditions and models their interactions across markets through factor-level attention, providing systematic information that can be incorporated into diverse stock prediction architectures. Experiments across the China, United States, and Hong Kong stock markets demonstrate consistent improvements when the method is integrated with different prediction models. Further analysis reveals time-varying cross-market interactions, providing insights into how the method captures changing patterns of systematic dependence across markets.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Finance, Stock Forecast, Deep Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_n2EGOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

DCSA is a plug-in module for existing stock return predictors. It learns market-specific latent factors, adapts them to a market state built from index price and volume, lets the target market's factors attend to other markets' factors, and then has stocks attend to those factors. It is attached to three host models and evaluated on Chinese, US and Hong Kong universes. Adding DCSA improves every reported metric in every host-market block.

Critical strengths
  • The improvement percentages I spot-checked match the ratios of the printed means, the ablation table agrees with the main table for the same model and dataset, and I did not find copied or contradictory numbers.
  • A foreign market's features are admitted only after that market closes, so information arrives with at least a one-session lag, and non-trading dates are dropped rather than filled.
  • Evaluating three markets and three structurally different host families, with licensed CRSP and CSMAR data that retain delisted names, is a substantial empirical effort.
Critical weaknesses
  1. The mechanism recomposes known parts. The systematic and idiosyncratic split is the CAPM to APT to Fama-French line, which the paper cites. Latent factors in stock prediction are standard, and MCI-GRU [5] and FactorGCL [6] already do it. State-dependent factor loadings are the conditional CAPM tradition [4], also cited. The one composite is cross-market factor-to-factor attention, which is multi-head attention with scalar gates over JJ tokens. The sigmoid in Eq. 15 also has a direct published precedent with the same motivation [8], and it is uncited. Most troubling, MASTER [7] is one of the three hosts and already builds a market-status gate from index price and volume, so the first stage of DCSA already lives inside a host it claims to improve.

  2. Appendix C describes both stages only in DCSA terms. Stage one is 40 epochs on the home market with foreign sources off. Stage two enables the foreign sources and trains 10 more epochs on all three markets jointly, with host parameters left unfrozen. A bare host has none of that machinery, so the paper never says what it does in stage two. Either it gets 40 epochs against 50, or it gets 50 without any three-market data. Either way the augmented model also has its host parameters fine-tuned on foreign panels and sees 2,100 panels instead of 800 or 500. The Section 4.5 claim that each row pair held the training protocol fixed is therefore unsupported. The effect sizes are +0.013 to +0.033 IC, which is the order of what ten extra epochs on 2.6 to 4.2 times more data would produce.

  3. Significance comes from a t-test over five seeds, but seeds only change the initialisation. The test set and the portfolio path are identical across seeds, so the seed spread measures optimisation noise rather than sampling error of realised performance. From the reported IC and ICIR, the day-level standard error of mean IC is four to ten times the seed-based one. For CN800 DTML it is 0.0066 against 0.00089. Only 35 of the 54 comparisons survive Bonferroni, and the full test family across the appendix is closer to 400. Two printed stars also contradict the numbers. CN800 MASTER AR has p = 0.069 and is starred, while US500 MASTER ICIR has p = 0.017 and is not.

  4. The final step is O=L+G′O = L + G', two unconstrained additive branches. There is no orthogonality penalty and no diagnostic showing that G′G' carries market-wide commonality. Eq. 17's elementwise gate also collapses the rank-JJ representation back to a per-stock vector before it is added, which discards the low-rank structure that would justify calling the output systematic. A control that replaces G′G' with an equally sized unconstrained branch is the natural check, and it is missing.

Minor issues

  • Three sentences claim more than the tables show. "Eight of the nine pairs falling between 21% and 44%" holds for six of nine, since US500 MASTER is 20.83% and HK800 FinD3 is 44.44%. "Best or joint-best on every reported metric" fails on HK800 DTML RankICIR, where SSA is 0.87 against DCSA 0.85. "US was the stronger foreign source in five of six cases" describes comparisons the appendix does not contain.
  • Appendix C changes DTML's objective from binary classification to MSE, so those rows are not the published DTML and the loss change is a confound of its own.

[1] Capital Asset Prices: A Theory of Market Equilibrium under Conditions of Risk. Journal of Finance 1964.

[2] The Arbitrage Theory of Capital Asset Pricing. Journal of Economic Theory 1976.

[3] Common Risk Factors in the Returns on Stocks and Bonds. Journal of Financial Economics 1993.

[4] Learning about Beta: Time-Varying Factor Loadings, Expected Returns, and the Conditional CAPM. Journal of Empirical Finance 2009.

[5] MCI-GRU: Stock Prediction Model Based on Multi-Head Cross-Attention and Improved GRU. Neurocomputing 2025.

[6] FactorGCL: A Hypergraph-Based Factor Model with Temporal Residual Contrastive Learning for Stock Returns Prediction. AAAI 2025.

[7] MASTER: Market-Guided Stock Transformer for Stock Price Forecasting. AAAI 2024.

[8] Theory, Analysis, and Best Practices for Sigmoid Self-Attention. ICLR

Questions for authors

See weaknesses for details, please.

Reviewer_1rpkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Existing studies often overlook the underlying market-wide systematic effects and their potential interactions across different markets. To address these limitations, this paper proposes a model-agnostic dynamic cross-market systematic attention method that enhances the performance of existing stock prediction models by incorporating dynamic latent systematic factors and cross-market factor interactions. Experiments conducted on the stock markets of China, the United States, and Hong Kong validate the effectiveness of the proposed method.

Critical strengths
  1. DCSA integrates with host models through an independent systematic information modeling branch and a gated residual connection, without requiring modifications to the core prediction architectures of the original models.

  2. This paper conducts experiments on three datasets, CN800, US500, and HK800, covering the stock markets of mainland China, the United States, and Hong Kong, which exhibit different market characteristics. Six evaluation metrics, including IC, ICIR, RankIC, RankICIR, AR, and SR, are employed to validate the effectiveness of the proposed method.

  3. In Appendix E.2, this paper analyzes the information contributions across different markets and their temporal variations through single-source-market experiments and gate-weight visualizations. This further analysis of cross-market information transmission enhances the interpretability of the experimental results.

Critical weaknesses
  1. Although the paper interprets the learned latent representations as market-wide systematic factors, the factors in Section 3.2.1 are optimized entirely through the prediction objective, and their explicit correspondence with traditional financial risk factors has not been established. This makes it difficult to verify whether they genuinely capture financially meaningful systematic risks.

  2. Although Appendix A accounts for differences in trading time zones across markets and avoids future information leakage by delaying the use of information from other markets, Section 3.2.2 primarily employs currently available cross-market factors for attention-based interactions. This design does not explicitly model potential lagged effects between different markets.

  3. The paper does not further distinguish prediction performance across bull markets, bear markets, high-volatility periods, and major financial events. Given that the core motivation of DCSA is to capture systematic effects that evolve with changing market conditions, rolling-window experiments and period-specific evaluations during episodes of extreme market volatility are necessary to verify whether the dynamic factors and cross-market attention mechanism remain effective when market structures undergo substantial changes.

Questions for authors

1.Although the paper interprets the learned latent representations as market-wide systematic factors, the factors in Section 3.2.1 are optimized entirely through the prediction objective, and their explicit correspondence with traditional financial risk factors has not been established. This makes it difficult to verify whether they genuinely capture financially meaningful systematic risks.

2.Although Appendix A accounts for differences in trading time zones across markets and avoids future information leakage by delaying the use of information from other markets, Section 3.2.2 primarily employs currently available cross-market factors for attention-based interactions. This design does not explicitly model potential lagged effects between different markets.

3.The paper does not further distinguish prediction performance across bull markets, bear markets, high-volatility periods, and major financial events. Given that the core motivation of DCSA is to capture systematic effects that evolve with changing market conditions, rolling-window experiments and period-specific evaluations during episodes of extreme market volatility are necessary to verify whether the dynamic factors and cross-market attention mechanism remain effective when market structures undergo substantial changes.

4.Although DCSA has been evaluated on three architectures, DTML, MASTER, and FinD3, Appendix C indicates that the full model adopts a two-stage training strategy, consisting of 40 epochs of single-market training followed by 10 epochs of cross-market joint fine-tuning. It is recommended to conduct additional controlled experiments under identical training epochs, parameter counts, and computational budgets.

1 comment

Thank you for the authors' response, which has addressed some of my concerns. Regarding the lead-lag relationships, I believe that implicit modeling through the DCSA module alone is insufficient to capture such effects.

Reviewer_LW5wOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DCSA augments stock prediction models with market-conditioned latent tokens, cross-market attention, and gated residual fusion. Experiments cover three host architectures in China, the U.S., and Hong Kong, with factor ablations, source-market analyses, and simulated portfolios.

Critical strengths
  1. The module is clearly presented and can be integrated into different hosts. A small set of market-level tokens provides a practical interface for sharing cross-market information without attending over all source stocks.

  2. Mean IC improves in all nine host-market configurations. Recent 2025 factor baselines, five-run standard deviations, and source-market ablations support the evaluation. The chronological split and session-close availability rule are useful, and the authors appropriately interpret transmission gates as routing behavior.

Critical weaknesses
  1. DCSA adds foreign-market inputs, capacity, and joint training, so host comparisons do not isolate its mechanism. Include index-feature concatenation/gating, standard cross-market attention, and a parameter-matched auxiliary branch under identical inputs and training budgets. Clarify the claimed input equivalence between Host and DSA, particularly for FinD3.

  2. Equation (19) does not identify systematic and idiosyncratic components; the host may also capture shared effects. Since DCSA combines conditional tokens, attention, and gates, its distinct benefit needs conditional-token controls and comparison with a dynamic-factor baseline such as the already-cited FactorVAE [1]. Validate the economic interpretation or state it more cautiously.

  3. Appendix D reports a Hong Kong cost of 0.01085%, below the 0.1% stamp duty per side on ordinary taxable stock trades [2]. Clarify whether this is a reporting error or an implementation omission; recompute AR/SR if needed and report turnover, cost sensitivity, and spread/slippage assumptions. This concerns economic performance and does not itself invalidate IC gains.

  4. Appendix A addresses foreign-session availability, but execution timestamps/prices and holding intervals remain unclear, especially given the different return targets across markets. Explain when forecasts become tradable and whether execution assumes an already-observed closing price. Specify point-in-time turnover-universe selection, reconstitution, and suspension/delisting treatment; including delisted stocks in the data does not establish unbiased selection.

  5. Five-seed statistics are useful, but the significance-test unit is unclear: seed averages mainly reflect initialization uncertainty, while daily observations require temporal-dependence correction. Specify paired versus unpaired testing and multiple-comparison handling. Yearwise/rolling results, paired block-bootstrap intervals, and benchmark-relative portfolio returns would strengthen the evidence.

[1] Duan, Y., Wang, L., Zhang, Q., and Li, J. FactorVAE: A Probabilistic Dynamic Factor Model Based on Variational Autoencoder for Predicting Cross-Sectional Stock Returns. AAAI, 36(4), 4468-4476, 2022. doi:10.1609/aaai.v36i4.20369.

[2] The Stock Exchange of Hong Kong Limited. Reduction of Stamp Duty on Stock Transactions Effective 17 November 2023. Circular CT/158/23, 15 November 2023. https://www.hkex.com.hk/-/media/HKEX-Market/Services/Circulars-and-Notices/Participant-and-Members-Circulars/SEHK/2023/ce_SEHK_CT_158_2023.pdf

Questions for authors
  1. Which index-level features are available to Host, SSA, DSA, and DCSA? Can simpler fusion methods match the gains with identical inputs, parameter counts, and training budgets?

  2. For each market, provide a timestamped example of the latest usable observations, execution time/price, and portfolio return interval. How are non-trading dates aligned in the ten-step lookback?

  3. Does the Hong Kong backtest include stamp duty beyond the stated 0.01085% charge? If correction is needed, how do AR/SR change under realistic costs?

  4. What point-in-time rules determine turnover-universe membership? Do the gains persist separately in 2024 and 2025 and under rolling evaluation?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.