- 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
-
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 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.
-
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.
-
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.
-
The final step is , two unconstrained additive branches. There is no orthogonality penalty and no diagnostic showing that carries market-wide commonality. Eq. 17's elementwise gate also collapses the rank- 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 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.
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.