acceptodds
Under review as a conference paper at ICLR 2027

RRST: Recurrent Relational Stock Transformer for Cross-Sectional Stock Ranking

Abstract

Cross-sectional stock ranking orders hundreds of stocks by their next-day return, and the public benchmarks judge the order two ways: by the correlation of scores with returns over the whole universe, and by the return of the few stocks at the top. We present the Recurrent Relational Stock Transformer (\rrst), one recurrent cell that steps over the last days of a window and carries two states per stock: a stock state that carries each stock's representation from one day to the next through a per-channel retention spread over timescales, so that recent days dominate while longer memory is kept, and a relation state holding each stock's loadings on latent factors, trained by an auxiliary loss to explain the next day's cross-section and consumed by the scoring layer as risk exposures. A scoring layer assembles one score for both kinds of metric. On the two public benchmark families, \rrst is above all 24 methods reported on NASDAQ and NYSE in cumulative return and Sharpe ratio (SR and against the best published and ) and above StockMixer on all four of its metrics on the same markets, with every hyperparameter chosen by a validation-only rule; the ablation locates the consistent top-of-ranking gain in the relation state used as exposures. We also prove that the benchmarks' price normalisation, by each stock's maximum over the whole series, encodes the future, measure what that column is worth, and report every result with and without the level inputs: NASDAQ and NYSE stay above every published top-of-ranking result in both settings, and S&P500 gains in IC and ICIR with the level inputs.

Primary area:
learning on time series and dynamical systems
Keywords:
cross-sectional stock ranking, financial time series, recurrent transformer, relational learning, dynamic stock relations, attention, quantitative finance

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_gu6gOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

RRST is a recurrent stock-ranking model with two states: a stock state for historical information and a relation state for learned factor exposures. It combines cross-stock attention, recurrent updates, and an exposure-adjusted scoring layer.

Experiments on NASDAQ, NYSE, and S&P500 report strong ranking results and component ablations. The paper also identifies future-information leakage in public benchmark normalization and evaluates the model with and without explicit price-level features.

Critical strengths
  1. Valuable benchmark leakage analysis. The paper identifies future-information leakage from full-series price normalization and measures its impact. It also acknowledges that removing price-level features alone does not guarantee fully causal preprocessing.

  2. Coherent architecture and strong results. RRST combines recurrent stock representations, learned relational exposures, and a scoring layer designed for stock ranking. The NASDAQ and NYSE results remain strong without explicit price-level features. The theoretical analysis and ablations provide useful evidence about the model's components.

Critical weaknesses
  1. The main results lack a fully causal, matched comparison. Removing level features does not fully rule out leakage from other preprocessing operations. This matters on S&P500, where IC falls from 0.0473 to 0.0065. The baseline results are also taken from previous publications rather than rerun under identical causal preprocessing and training settings. This limits the evidence for RRST's reported advantage.

  2. The contributions of individual mechanisms remain unclear. Table 3 shows mixed results. Removing cross-stock attention or leaky recurrence improves NYSE cumulative returns while reducing Sharpe. Removing learned relation exposures reduces returns but raises NYSE Sharpe from 7.02 to 9.74. These results suggest that features, exposure adjustment, and scoring also play important roles. More closely matched architectural comparisons would help isolate the proposed mechanisms.

Questions for authors

Can RRST and the strongest baselines be evaluated using the same fully causal preprocessing and training protocol?

Can the recurrent and relational mechanisms be compared with simpler models using identical features, losses, and scoring?

Can the authors clarify the selection losses, differentiable Top-K objectives, valid-stock masking, and exposure normalization in Equation (5)?

2 comments

We thank the reviewer for recognizing our benchmark leakage analysis, architectural design, and empirical contributions.

1. Causal validity and benchmark fairness.

Tables 1–2 distinguish RRST (as given), which follows the original benchmark's future-dependent normalization for fair comparison, from RRST (no levels), our verified leakage-free evaluation. The latter uses no future information in normalization, price recovery, validity masks, or numerical thresholds. We will correct the outdated manuscript qualifications.

Without leakage, RRST achieves NASDAQ/NYSE Sharpe ratios of 4.65/5.65 and cumulative returns of 1.02/1.24, exceeding the best published values of 2.32/2.35 and 0.82/0.79.

We attempted to reproduce StockMixer using its official repository but could not recover the published results; similar problems are documented in GitHub Issues #12–14. We therefore use published results for comparisons, including graph-based methods.

2. Independent architectural contributions.

Tables 3 and 12 isolate modules while keeping features, losses, and scoring identical. Removing cross-stock attention reduces NASDAQ Sharpe from 4.48 to 3.04 (95% CI for the difference: [0.14, 2.68]). Replacing recurrence with a single-step model reduces NASDAQ IC by approximately 0.002.

Removing learned relational exposures while retaining historical PCA exposures reduces NASDAQ/NYSE cumulative returns from 0.90/0.92 to 0.67/0.67. Although NYSE Sharpe increases from 7.02 to 9.74, this reflects a trade-off between cumulative and risk-adjusted performance, not a uniform improvement. The relation loss also increases explained next-day cross-sectional variation from 4.0%/4.6% to 10.2%/12.8%.

3. Training objectives and implementation.

The first selection head minimizes negative Pearson IC against volatility-adjusted returns; the second uses binary cross-entropy for nonnegative returns. Differentiable top-K objectives use soft portfolio weights:

w=softmax⁡(maskedzscore⁡(s)/0.25)w = \operatorname{softmax}(\operatorname{masked_zscore}(s)/0.25)

They optimize negative batch Sharpe and weighted hit rate, with respective weights 0.1 and 0.2. Additional objectives include forecast IC, deep supervision, relation reconstruction, and residual prediction.

Training masks require valid input windows and target-day observations. Equation (5) uses masked, exposure-standardized ridge regression with coefficient 10−610^{-6}, γ=γR=0.5\gamma=\gamma_R=0.5; relation reconstruction separately uses 10−210^{-2}.

These clarifications and controlled ablations support RRST's relational modeling contribution, and we agree the value of additional matched external comparisons.

Thank you for the detailed response. The training objectives and implementation details are now clearer. I also appreciate the controlled ablations, which provide stronger evidence for the attention and relational components.

However, I will maintain my rating of 2 (Weak rejection) for two reasons. First, the main comparison is still not fully matched. The authors report strong results from the leakage-free RRST variant, with NASDAQ/NYSE Sharpe ratios of 4.65/5.65. However, the baselines still use published results rather than being evaluated under the same corrected preprocessing. This leaves the central performance comparison unresolved. Second, the ablations show useful but mixed contributions. Removing learned relational exposures reduces NASDAQ/NYSE cumulative returns from 0.90/0.92 to 0.67/0.67, but increases NYSE Sharpe from 7.02 to 9.74. The single-step ablation also changes NASDAQ IC by only about 0.002. These results support parts of the design, but not a consistent advantage across markets and metrics. The response addresses several of my concerns, but the remaining evaluation limitations still affect the paper's main claims. I therefore keep my original rating.

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

RRST maintains recurrent stock states and factor-loading states over historical sub-windows. It combines exposure-adjusted forecasts with two selection heads to produce stock scores.

Critical strengths

Training the relation state as predictive factor loadings gives it a clear role, and the exposure ablation tests its contribution beyond historical PCA loadings.

Critical weaknesses
  1. Table 12 gives mixed evidence for recurrence, and published baselines were not rerun with the same features and scoring layer.
  2. Appendix A.1 claims that scale-dependent preprocessing may still transmit future information after level removal, so the no-level results do not establish a fully past-only forecasting result.
Questions for authors
  1. Do the gains remain when every feature and preprocessing decision is constructed only from prices available at the forecast date, including the scale-dependent recovery checks and floors?
  2. Can a separately tuned StockMixer or simpler nonrecurrent model match RRST when given the same past-only features, risk adjustment, selection heads, and tuning budget?
1 comment

We thank the reviewer for recognizing the contribution of predictive relational representations and their role beyond historical PCA exposures.

1. Causal preprocessing and clarification of Tables 1–2.

We clarify that the two RRST settings serve different purposes:

  • RRST (as given) follows the original benchmark normalization, including future-dependent statistics, to ensure direct comparability with published methods.
  • RRST (no levels) is our verified leakage-free evaluation. Neither validation nor testing data are used for training normalization, and the complete preprocessing pipeline, including price recovery, numerical thresholds, and validity masks, introduces no future-information leakage. We will correct the outdated manuscript statements accordingly.

Under the leakage-free setting, NASDAQ/NYSE achieve cumulative returns of 1.02/1.24 and Sharpe ratios of 4.65/5.65, compared with the published best values of 0.82/0.79 and 2.32/2.35.

2. Matched StockMixer comparison.

We attempted to reproduce StockMixer using its official repository but could not recover the published performance. Similar concerns have been publicly reported in GitHub Issues #12–14, including reproduction difficulties and implementation inconsistencies.

Therefore, we compare against the published results using identical benchmark data and normalization conventions rather than potentially unreliable reproduced results. RRST outperforms StockMixer on all four metrics on NASDAQ and NYSE.

Furthermore, Tables 3 and 12 provide controlled architectural ablations with identical features, losses, and scoring. Removing cross-stock attention reduces NASDAQ Sharpe from 4.48 to 3.04, while removing learned relational exposures reduces cumulative returns from 0.90/0.92 to 0.67/0.67.

Reviewer_GNytOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The author addresses three limitations of current methods in stock selection: 1. Lack of explicit treatment modeling the evolution of stock relations 2. Retention of long distance stock state 3. Indiscrepency between benchmark evaluation protocol and optimization target of model. To address these problems, the authors propose RRST methods. Specifically RRST models the evolution of stock relations with explicit recurrent modeling of stock relations models long distance dependency of stock state through intermediate deep supervision and introduced multiple optimization objective that better align with evaluation protocol.

Critical strengths
  • The authors addresses three critical problems in stock selection namely the modeling of stock relationship evolution, retention of long distance dependency of stock states and misalignment between benchmark evaluation protocol and optimization target.
  • The authors propose the recurrence mechanism to explicitly model the evolution of stock relationships rather than relying on predefined relationships or attention mechanisms without explicit evolution modeling.
  • The authors propose retention-based stock state update and deep supervision at intermediate steps to improve the retention of long distance dependency.
  • The multiple prediction heads and associated optimization objectives explicitly account for the differences between overall cross-sectional ranking quality and top-of-ranking stock selection performance.
  • RSST demonstrates superior performance on NASDAQ, NYSE and S&P500 benchmarks showing the effectiveness of overall design.
  • Extensive ablation experiments validated the effectiveness of each design.
Critical weaknesses
  • Limited Justification for the Novelty of Retention Mechanism. The authors introduce a retention-based recurrent update and intermediate deep supervision to address the difficulty of retaining long-range historical information by recurrence modeling techniques. However, attention-based architectures can also model long-range temporal dependencies through direct interactions between distant time steps. It remains unclear why the proposed recurrent mechanism is preferable to temporal attention. A comparison with such alternatives would help establish the necessity and benefits of the proposed design.
  • The Overall Presentation Needs Improvement. The paper would benefit from clearer, more concise, and better-organized writing. This would make the paper easier to follow.
Questions for authors
  • Could the author better discuss the novelty of the multi-task optimization targets beyond plain combination of well established IC and top-of-ranking selection objectives?
  • Could the author better discuss the necessity of the retention mechanism compare with attention mechanism which is well-known for good at long-distance dependency modeling?
1 comment

We thank the reviewer for recognizing the motivation, methodological soundness, and empirical contributions of RRST.

1. Recurrence versus temporal attention.

We agree that temporal attention can capture long-range dependencies. However, RRST addresses a different modeling objective: explicitly maintaining and evolving both per-stock representations and cross-stock relational factor states across successive historical sub-windows.

Our contribution is not a new retention operator in isolation, but its integration with recurrent relational modeling. Theorem 1 establishes a decoder-independent gradient path, while Table 13 shows that RRST retains approximately 5–12% of its state-gradient norm across seven steps, compared with only 0.28% for the overwriting cell.

Furthermore, the recurrent relation state learns predictive cross-sectional structure. Its explained share of next-day standardized returns increases from 4.0%/4.6% to 10.2%/12.8% on NASDAQ/NYSE when trained with the relation objective.

These findings support our recurrent modeling design, although we acknowledge that a matched temporal-attention baseline would be needed to establish comparative architectural advantages.

2. Novelty of multi-objective optimization.

We do not claim that individual IC or selection losses are novel. Proposition 4 establishes that optimizing cross-sectional IC does not necessarily maximize equal-weight top-K returns.

Motivated by this distinction, RRST combines exposure-adjusted forecasts with two specialized selection heads through a tiered scoring mechanism. The resulting composite score supports both cross-sectional ranking and top-of-ranking stock selection within one framework.

Thus, our contribution lies in the task-specific integration of recurrent relational modeling, predictive factor learning, and metric-aligned scoring, rather than introducing entirely new individual losses.

3. Presentation.

We appreciate this suggestion and will improve the organization of the method section, particularly the connections between historical features, recurrent states, relational learning, and final scoring. We will also distinguish theoretical properties from experimentally demonstrated performance gains.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.