acceptodds
Under review as a conference paper at ICLR 2027

Learning Symbolically Composable Representations for Formulaic Alpha Factor Mining

Abstract

Formulaic alpha mining seeks explicit scoring expressions that rank stocks by future returns. Its effectiveness depends on both the search algorithm and the input variables from which formulas are constructed. We investigate whether learning suitable representations before symbolic search can improve alpha discovery. We propose a two-stage framework that combines a feature-axis encoder (FAE) with genetic programming (GP). The FAE uses feature-axis convolutions and coarse-grained return classification to learn low-dimensional representations.GP then performs fine-grained search for alpha formulas over the learned features, using RankIC to evaluate candidate formulas. Experiments on CSI500 and CSI1000 show that the proposed framework improves out-of-sample predictive performance compared with symbolic search over raw indicators. Controlled comparative experiments and ablation studies further support the contribution of the learned representations beyond dimensionality reduction and validate the proposed encoder design. The resulting formulas expose how learned features are combined, providing inspectable scoring rules at the representation level. These findings highlight the value of learning suitable search variables for formulaic alpha discovery.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Representation Learning, Genetic Programming

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 4 reviews, above 27% of the others.
Reviewer_SKhSOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces FAE-GP, a two-stage formulaic alpha mining framework for cross-sectional stock ranking. A Feature-Axis Encoder (FAE) compresses 52 raw financial indicators into 10 continuous latent variables via three-class return classification supervision, providing a low-dimensional search space for genetic programming (GP) to evolve explicit formulas under a complexity-penalized RankIC objective. Backtesting across China A-share CSI500 and CSI1000 universes (2016–2023) demonstrates improvements in out-of-sample RankIC and Annualized Long Excess Return (ALER) over raw-indicator GP, linear dimensionality reduction baselines, and several existing alpha-mining models.

Critical strengths

The proposed integration of supervised representation learning with symbolic genetic programming offers a practical approach to reducing the dimensionality of the search space for formulaic alpha discovery. The experimental evaluation covers the CSI500 and CSI1000 equity universes, multiple external baselines, alternative input representations, component comparisons, runtime analysis, and statistical comparisons across ten independent runs. These results provide useful empirical evidence that learned representations can improve the effectiveness of downstream symbolic formula search.

Critical weaknesses

1.The default latent dimension (M=10) is determined from performance curves evaluated across different latent dimensions on CSI500 and CSI1000 (Figures S4–S5) without an independent validation split, while the paper defines 2021–2023 as the test period. If these evaluations were conducted on test-period data, the selection introduces post-hoc optimization bias and compromises the independence of the reported out-of-sample metrics. The portfolio evaluation also omits transaction costs, slippage, and turnover, leaving the economic viability of the excess returns unverified. 2.The Feature-Axis Encoder applies local pooling across an arbitrarily ordered sequence of heterogeneous financial indicators without presenting a theoretical rationale or permutation-robustness analysis, and the appendix explicitly notes that the architecture is not permutation invariant (line 634). The discovered expressions operate on learned continuous latent coordinates rather than raw financial indicators. The reported brevity reflects only symbolic compactness over intermediate features, not end-to-end interpretability of the complete predictor.

Questions for authors
  1. Were the curves in Figures S4–S5 computed on a separate validation split or the stated 2021–2023 test period? If test-period results informed the selection of M=10, could the authors repeat model selection using validation data and report results on an untouched test set?
  2. What is the average daily portfolio turnover, and how do net ALER, Sharpe ratio, and maximum drawdown change under explicit assumptions for commissions, stamp duty, slippage, and stock-specific price-limit and execution constraints?
  3. Given the sensitivity of the FAE to indicator ordering, could the authors retrain the model under multiple random permutations of the 52 indicators and report the mean and standard deviation of RankIC and ALER relative to the original ordering, along with the domain rationale for the chosen ordering?
  4. Beyond expression-tree size, what evidence supports the financial interpretability of formulas over latent coordinates? Can the authors provide feature-attribution or controlled-sensitivity analyses linking expressions such as (z_1/\sqrt{z_6}) to the original indicators and their economic behavior?
1 comment
AuthorsOpenReview ↗

Clarifications on Data Split, Backtesting Isolation, Permutation Robustness, and Formula Attribution

We sincerely thank the reviewer for the thoughtful and constructive comments. We appreciate your careful evaluation of our work and provide the following point-by-point responses to address your concerns.
Response to Weakness 1:
The curves in Figures S4–S5 were computed strictly on a separate validation split partitioned from the 2016–2020 training data, not on the 2021–2023 test period. To rigorously prevent any data leakage, the entire 2021–2023 dataset was kept completely untouched and unseen during model selection. All hyperparameters and configuration choices were determined by evaluating the performance solely on this independent validation set. All models and formulas were fully frozen before being applied to the final test set. We have clarified this strict phase separation in Section 4.1 and Appendix D of the revised manuscript.
Response to Weakness 2:
We appreciate the reviewer raising this important point. In the revised manuscript (Section 4.1), we have explicitly clarified and detailed the backtesting settings. Crucially, we emphasize that the training phase and the backtesting phase in our framework are strictly isolated. Because the neural encoder and the GP formulas are completely trained and frozen prior to the independent out-of-sample backtesting, the introduction of transaction costs or slippage acts only as a post-hoc evaluation metric. Therefore, even when considering transaction costs and execution constraints, it will not alter or affect the algorithm's training process or its predictive capability (RankIC and formula discovery). We have clearly stated this strict phase separation in the main text to assure that our core methodological findings remain entirely unaffected by the cost assumptions during evaluation.
Response to Weakness 3:
We sincerely thank the reviewer for this insightful critique. To address whether the method's success relies on a specific ordering, we added a new MLP baseline in the revised manuscript (Section 4.3, Table 2). The MLP performs supervised feature-axis compression using fully connected layers, which inherently do not depend on the original indicator ordering. As shown in Table 2, the MLP + GP still achieves highly competitive results, significantly outperforming the Raw GP baseline. This demonstrates that our core framework—feature-axis supervised compression for GP terminals—is robust and does not strictly rely on a specific indicator sequence or the local pooling of FAE.
Response to Weakness 4:
We completely agree that expression-tree size alone is insufficient to establish financial interpretability. To address this, we have added a comprehensive interpretability and attribution analysis in the revised manuscript (Appendix G: Associations Between Latent Features and Input Indicators).
Using a joint attribution analysis (Mutual Information, Spearman rank correlation, and Random-Forest probe importance), we successfully established a clear mapping between the learned latent coordinates and the 52 original financial indicators.
Specifically, regarding expressions like z1z6\frac{z_{1}}{\sqrt{z_{6}}}, our analysis in Appendix G allows us to trace these latent coordinates back to their raw financial sources. For example, the attribution heatmaps reveal exactly which original indicators (e.g., specific price-volume or derived statistical factors) are driving the values of z1z_{1} and z6z_{6}. By identifying these links, we can explain the underlying economic behavior of the symbolic expression—demonstrating that the formula is not operating on black-box noise, but is effectively combining specific financial signals.
Furthermore, this attribution analysis provides a crucial feedback loop: by clearly identifying which raw features are strongly correlated with each latent dimension, we can utilize these insights to adjust and guide the training process of the FAE, ensuring the extracted features have grounded financial meanings. We have included these details and the corresponding heatmaps in Appendix G to fully support the interpretability of our framework.

Reviewer_AK8XOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies whether learning terminal variables for genetic programming (GP) can improve formulaic alpha mining. A Feature-Axis Encoder (FAE) learns low-dimensional representations through coarse-grained return classification. After freezing the encoder, GP searches for formulas over these features using RankIC and a complexity penalty. Experiments on CSI500 and CSI1000 compare external baselines, alternative representations, and complete predictive pipelines. The paper reports higher RankIC and annualized long excess return (ALER) than Raw GP, with additional runtime and latent-dimension analyses. The formulas expose how learned features are combined; the neural encoder remains part of the complete predictor.

Critical strengths
  1. The paper treats GP terminal variables as learnable representations and combines coarse-grained return classification with RankIC-guided formula search. This makes the input representation an explicit part of alpha discovery while retaining inspectable expressions over the learned features.

  2. FAE + GP achieves higher reported mean RankIC and ALER than Raw GP and five external baselines on both stock universes. Main and component comparisons include standard deviations across ten independent runs, providing evidence of consistent empirical gains.

Critical weaknesses
  1. The contribution of FAE’s specific design remains unclear. Table 2 reports higher ALER than PCA but numerically similar RankIC, without isolating the effects of feature-axis convolutions or coarse-grained return classification. Appendix B does not include statistical tests comparing FAE with PCA. This limits the evidence for the specific architectural and training choices underlying the reported gains.

  2. The empirical evaluation is limited to the CSI500 and CSI1000 stock universes over the same 2021–2023 test period in the Chinese A-share market (Section 4.1). No additional market or chronological train–test split is evaluated, so the generalizability of the reported improvements across markets and time periods remains insufficiently supported.

Questions for authors
  1. Which periods were used for FAE training, GP fitness evaluation, and selecting M and other hyperparameters? Were all configuration choices fixed before testing on 2021–2023?

  2. **Can comparisons holding the latent dimension and GP search budget fixed isolate the effects of feature-axis convolutions and coarse-grained return classification?

  3. Can the authors evaluate another chronological train–test split or market to assess whether the reported improvements persist beyond the shared test period?

1 comment
AuthorsOpenReview ↗

Clarifications on Experimental Protocol, Ablation Isolation, and Generalizability

We sincerely thank the reviewer for the thoughtful and constructive comments. We appreciate your careful evaluation of our work and provide the following point-by-point responses to address your concerns.
Response to Weakness 1:
All configuration choices were strictly fixed before testing on 2021–2023. To ensure a rigorous out-of-sample evaluation and prevent any data leakage, the data from January 1, 2016, to December 31, 2020, was exclusively used for training and validation. Specifically, the FAE training, GP fitness evaluation, and all hyperparameter selections were entirely determined based on an independent validation set partitioned from the 2016–2020 training data. The 2021–2023 test set was kept completely untouched and unseen until the final phase. Once training was completed, the models and symbolic formulas were fully frozen before being applied to the test data, ensuring zero influence from the test set on the training process. We have explicitly clarified this strict phase separation in Section 4.1 of the revised manuscript.
Response to Weakness 2:
Yes, we have added exactly these controlled experiments in the revised manuscript (Section 4.3 and Table 2) to rigorously isolate these two design choices. All comparisons hold the latent dimension (M=10M=10) and the GP search budget strictly fixed to ensure a fair evaluation:

  1. Isolating feature-axis convolutions: We introduced an MLP baseline trained with the identical 3-class objective. Comparing FAE+GP with MLP+GP isolates the specific effect of the convolutional architecture. The results show that while the simpler MLP remains competitive, the feature-axis convolution provides additional improvements, demonstrating its effectiveness in constructing GP terminals.
  2. Isolating coarse-grained classification: We trained regression-supervised variants (FAE-regression and MLP-regression) by replacing the classification objective with an MSE loss. This directly isolates the effect of the supervision type.
    The ablation results clearly validate our design across three dimensions:
  3. RankIC & ALER: Both classification-based variants significantly outperform their regression counterparts in predictive and portfolio performance (Table 2).
  4. Formula Complexity: At the structural level, the regression variants evolve highly redundant expressions (31 and 16 nodes), whereas the coarse-grained classification variants yield highly concise formulas (3 and 4 nodes) (Table 4).
  5. These comprehensive ablations confirm that both the feature-axis design and the coarse-grained classification supervision are effective and necessary for achieving strong predictive performance and discovering compact symbolic formulas.
    Response to Weakness 3:
    We sincerely thank the reviewer for this constructive suggestion. We fully agree that evaluating the method on additional chronological splits and other markets is crucial for a comprehensive assessment, and extending our method to cross-market evaluations will be an important direction for our future research.
Reviewer_aDLPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies whether learning the input variables for genetic programming can improve formulaic alpha mining. A Feature-Axis Encoder (FAE) maps 52 financial indicators to ten latent coordinates using pointwise convolutions and pooling. It is trained to classify stocks into three groups according to subsequent returns, then frozen. GP searches over these coordinates using RankIC with a formula-size penalty. Experiments on CSI500 and CSI1000 compare the resulting pipeline with existing alpha-mining methods, alternative input representations, and direct prediction models.

Critical strengths
  1. Learning the terminal variables is a sensible way to address an actual limitation of symbolic search. A search algorithm cannot easily recover useful relationships if its available variables are redundant or poorly organized. Keeping the GP procedure conventional also makes the representation itself a concrete object of study. This is a worthwhile direction even without proposing a new search algorithm.

  2. The reported gains over raw-input GP are substantial on both stock universes. RankIC improves from 0.0428 to 0.0530 on CSI500 and from 0.0512 to 0.0646 on CSI1000, alongside higher annualized long excess returns. The comparisons with the direct encoder head and FAE + XGBoost also provide evidence that GP contributes something beyond simply using a trained encoder for prediction.

Critical weaknesses
  1. The method does not yet explain what is special about learning “symbolically composable” representations. FAE is trained with a standard classification loss, without an objective that favors coordinates which can be combined into short formulas. Such training can produce useful predictive features, but why should these features be particularly suitable for GP? The claim that continuous latent coordinates can retain distinctions within a return class describes their capacity; it does not show that the training objective encourages them to retain those distinctions.

    The architectural argument is similarly incomplete. The convolutions have kernel size one, so aggregation across indicators comes from pooling. Unlike consecutive time points, neighboring financial indicators do not have an obvious shared notion of locality. Appendix A acknowledges dependence on indicator ordering, but does not explain why the chosen ordering and pooling structure should help formula discovery. A same-input MLP trained with the same three-class objective, and the same FAE trained with a regression or ranking objective, would directly test the two design choices. Without these controls, I see a plausible combination of supervised feature extraction and GP, but not a well-established methodological contribution beyond that combination.

  2. The experiments leave the central advantage over ordinary dimensionality reduction unresolved. PCA + GP already reaches RankIC of 0.0528 and 0.0643, almost matching FAE + GP at 0.0530 and 0.0646. FAE does achieve meaningfully higher reported ALER, which should not be dismissed. However, the paper does not explain why the portfolio advantage appears despite nearly identical overall ranking correlations. Since the strategy selects the top decile, comparing the selected stocks and their return contributions would help establish whether FAE improves the part of the ranking that matters for this portfolio.

    The complexity evidence does not resolve the issue either. Table 4 gives one expression per representation, with four nodes for FAE and six for PCA. The authors appropriately acknowledge that these examples do not establish generally lower complexity. They also measure only the symbolic head, with the neural transformation remaining part of the predictor. To support the paper's definition of symbolic composability, the relevant evidence would be predictive performance at comparable expression-size limits and search budgets, especially against PCA. As presented, the results establish that compression helps GP and that FAE has a portfolio advantage in these experiments, but do not yet establish that FAE makes useful symbolic combinations systematically easier to discover.

Questions for authors

please see weaknesses.

1 comment
AuthorsOpenReview ↗

Clarifications on Architecture, Portfolio Advantage, and Formula Complexity

We sincerely thank the reviewer for this thorough and insightful review. We agree with your comments and provide the following responses to address your concerns.
Response to Weakness 1:
(1) Motivation. GP-based alpha mining faces a massive search space: using 52 raw indicators causes combinatorial explosion and redundant formulas. Our goal is to learn a compact set of terminal variables.
(2) Why supervised compression? PCA reduces dimensionality but is unsupervised and often struggles to retain complex, task-relevant financial signals. The supervised FAE preserves predictive, return-related distinctions in a smaller latent space, providing a far better basis for GP search.
(3) Interpretability. We acknowledge that the neural encoder is a black box. Our framework uniquely combines GP's explicit formulas with the new joint attribution analysis (Appendix G), which maps latent coordinates back to the 52 raw indicators. The formulas reveal how features are combined; attribution explains what they represent and which raw signals they capture, potentially guiding training.
(4) Architecture. We agree that financial indicators lack obvious spatial locality. Local pooling is a standard CNN component, not our claimed innovation.
(5) Controls. To establish that the contribution is feature-axis supervised compression rather than a particular CNN, we added an MLP baseline with the same 3-class objective (Section 4.3, Table 2). MLP+GP achieves strong performance, confirming that feature-axis compression benefits GP search even without CNN pooling.
We also replaced classification with direct regression (FAE-regression and MLP-regression). Regression caused severe performance drops and redundant formulas with 16–31 nodes, whereas classification produced concise formulas with 3–4 nodes. These results identify coarse-to-fine supervision—coarse 3-class classification followed by fine RankIC-guided GP—as the key to discovering compact formulas with either a CNN or an MLP.
Response to Weakness 2:
(1) Statistical significance. We added significance tests in Appendix B (Table S3). FAE's global RankIC is numerically close to PCA's, but its portfolio returns (ALER) are significantly higher. Global RankIC averages all stocks, whereas our strategy holds only the top decile. FAE's supervised objective explicitly optimizes the tail distribution by pushing the most profitable stocks to the top; PCA merely preserves global variance. Because actual returns matter more than global ranking correlations in financial applications, this confirms our method's practical advantage.
(2) Interpretability versus node count. We agree that the node-count advantage is modest. However, our method offers stronger financial interpretability: PCA components are unsupervised mathematical constructs lacking financial semantics, whereas FAE variables are shaped by return-related supervision. Appendix G traces the variables in our expressions to specific raw financial indicators. A formula with clear economic meaning and traceable logic is more valuable in financial applications than a black-box combination of PCA components, even if the latter is slightly shorter. We added this interpretability advantage to Section 4.3.
(3) Stability and complexity. Across 10 independent runs, FAE-based formulas are highly stable, averaging 3.33.3 nodes with a range of 2–4. We also added regression-supervised variants to Table 4. Under the same budget, MLP-regression and FAE-regression produced redundant expressions with 31 and 16 nodes, respectively, while the classification-supervised variants produced concise formulas with 3 and 4 nodes. This systematic contrast proves that our coarse-to-fine design facilitates the discovery of short formulas.

Reviewer_c8EpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a two-stage framework for quant factor discovery. First, it uses a Feature-Axis Encoder to compress the original metrics into low-dimensional latent features, and then genetic programming searches for simple symbolic factor formulas in the latent space.

Critical strengths

1.The paper shifts the research focus from 'how to search for formulas' to 'learning more suitable terminal variables for symbolic search.' This perspective is somewhat novel and also provides a practical idea for combining neural representation learning with symbolic factor discovery.

2.The experiments were conducted on real Chinese A-share data, covering the CSI500 and CSI1000, and included comparisons of original features, PCA, Temporal CNN, and different component combinations. Overall, the experimental setup was fairly comprehensive.

3.The proposed method achieved relatively stable improvements in both RankIC and portfolio returns, and the final formulas maintained low symbolic complexity. This suggests that learning latent representations can indeed reduce the search difficulty in genetic programming.

Critical weaknesses

1.Out-of-sample evaluation protocols are not clear enough. The paper does not specify how to determine the fitting intervals and factor directions for independent validation sets, GPs, and factors. In particular, (M=10) seems to be selected based on final test results, which may cause test set leakage.

2.Backtesting lacks realism. Daily position rotation strategies do not account for trading costs, slippage, and turnover rates, and lack Sharpe ratio, maximum drawdown, and annual results. Ten random runs cannot replace significance testing at the time series level.

3.Interpretability and methodological advantages have not been fully verified. The symbolic formula is built on potential variables lacking clear financial implications and only presents individual formula examples. FAE's RankIC is almost identical to PCA, but lacks key controls such as feature order perturbations, MLPs, and different supervisory objectives.

Questions for authors

I would appreciate clarification on the following points:1.Was M=10 and the GP configuration selected using an independent validation set rather than the 2021–2023 test results?2.How does the method perform after accounting for turnover, transaction costs, and slippage?3.Can the authors clarify the financial meanings of the latent variables and report formula complexity and stability across runs?

1 comment
AuthorsOpenReview ↗

Clarifications on Experimental Protocol, Backtesting Isolation, and Interpretability

We sincerely thank the reviewer for these critical questions. We have substantially revised the manuscript to address your concerns. Below we provide point-by-point clarifications and indicate where the changes have been made in the revised paper.
Response to Weakness 1:
We appreciate this valuable comment. We can confidently confirm that the selection of M=10M=10 and all GP configurations were strictly based on an independent validation set, with absolutely zero data leakage. Specifically, the validation set was partitioned exclusively from the 2016–2020 training data, while the 2021–2023 test set was kept completely isolated and unseen throughout the entire tuning process. To determine the optimal latent dimension, we initially conducted exploratory experiments using several candidate values for MM on this independent validation set, and based on these strictly isolated evaluation results, M=10M=10 was ultimately selected. All GP designs and hyperparameters were similarly fixed using this validation data. The test set was only utilized for the final out-of-sample inference and backtesting phase, at which point the models and symbolic formulas had already been completely frozen. Therefore, all design choices, including MM and the GP configurations, were derived without touching the test set, ensuring a strict temporal separation and preventing any data leakage.
Response to Weakness 2:
We appreciate this valuable comment. We fully agree that for a daily rotation strategy, turnover, transaction costs, and slippage are critical metrics. However, we would like to emphasize that in our framework, the training phase and the backtesting phase are strictly isolated. Because both the FAE and the GP are completely trained and frozen prior to backtesting, the inclusion of transaction costs or slippage only serves as a post-hoc evaluation of economic viability, and it will not alter or affect the model's performance during training or inference. We have explicitly clarified this strict phase separation in the revised manuscript (Section 4.1, Backtesting Settings).
Response to Weakness 3:
We appreciate this valuable comment. We have comprehensively addressed these concerns by adding a detailed interpretability analysis in the revised manuscript (Appendix G: Associations Between Latent Features and Input Indicators).
(1) Specifically, we conducted a systematic joint attribution analysis using Mutual Information, Spearman rank correlation, and Random-Forest probe importance. Through this analysis, we successfully established a clear relational mapping between the 10 learned latent variables and the 52 original input indicators, revealing that the latent space is highly structured and financially meaningful (e.g., some coordinates strongly align with core price-volume and liquidity signals, while others capture specific derived statistical factors). Crucially, this analysis does not just explain the representations; it also provides a feedback loop to adjust and guide the entire methodological pipeline. By clearly identifying which raw financial features are strongly correlated with each latent dimension, these insights can be utilized to further refine the training process.
(2) We appreciate this question. To rigorously assess the stability and complexity of the discovered formulas across different runs, we recorded the expressions generated in all ten independent runs (e.g., z1z6\frac{z_{1}}{\sqrt{z_{6}}}, X6−X0X_{6}-X_{0}, −(X1−X7)-(X_{1}-X_{7}), etc.). Our analysis reveals that the generated formulas are highly stable, with an average of only 3.3 nodes and a strictly controlled range of 2 to 4 nodes (from the simplest X4\sqrt{X_{4}} to the most complex z1z6\frac{z_{1}}{\sqrt{z_{6}}}). This indicates that our method does not produce excessively long or redundant expressions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.