acceptodds
Under review as a conference paper at ICLR 2027

CausalRank: Discovering Causal Factors for Cross-Sectional Stock Returns

Abstract

Empirical asset pricing in the finance literature requires not only identifying numerous predictive factors but also uncovering the reliable and explainable pricing mechanisms behind return predictability. To answer this question, we propose CausalRank, a foundation model designed for parent node ranking. For a given target variable, CausalRank aims to simultaneously identify its direct causal parents from a massive pool of candidate variables and rank them based on the strength of their causal effects. Specifically, we train CausalRank on synthetic and semi-synthetic data with known causal structures and intervention effects. This training enables the model to learn transferable causal inference capabilities across various causal structures and generating functions. CausalRank distinguishes direct parent variables from correlated noncausal ones by assessing each pairwise candidate-target relationship in the context of the other candidates. This enables it to predict both whether a direct causal relationship exists and its effect. Using the U.S. stock market data, we show that the investment strategy using the characteristics identified by CausalRank can achieve a higher risk premium with lower turnover. Meanwhile, it ranks these characteristics by causal effect strength, which is crucial for evaluating a factor's role in empirical asset pricing and understanding the complexity of financial markets.

Primary area:
causal reasoning
Keywords:
Empirical Asset Pricing, Causal Foundation Model, Causal Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_cvLzOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents CausalRank, a target-specific framework for jointly identifying direct causal parents and ranking their causal effect strengths, motivated by cross-sectional stock return analysis. It combines candidate–target representations with conditional information from other variables, using synthetic pretraining and financial semi-synthetic fine-tuning. Validation spans synthetic and semi-synthetic benchmarks, correlated-proxy tests, ablations, and a U.S. equity application. The results support the framework’s potential, although the justification of specific design choices, model selection procedures, and interpretation of some baseline comparisons require further clarification.

Critical strengths
  1. The paper combines direct parent identification and ranking by direct causal effect strength into a target-specific causal ranking task. The problem is well motivated and worth studying, particularly for identifying return drivers from a large pool of candidate stock characteristics.

  2. CausalRank integrates representation encoding, target-aware parent interaction, and causal ranking into a unified framework. Its components correspond to the stated motivations for parent identification, conditional evidence extraction, and effect ranking.

  3. The paper is clearly structured. The task definition, main method, and principal experimental results are easy to follow.

  4. Comparisons involving effect and ranking supervision, candidate competition, and semi-synthetic fine-tuning provide solid evidence for several key designs.

Critical weaknesses

1.Target parents and their effect ranking provide less structural context than a full causal graph, making it harder to inspect relationships among variables against domain knowledge.

2.The rationale and comparisons for specific design choices are insufficient. The paper explains some design motivations and provides relevant ablations supporting several components. However, it does not sufficiently justify why these particular choices are optimal. Specifically, the paper neither demonstrates that the four fixed transformations are the best among alternatives, nor clarifies how the combination of statistics and the network architecture were selected.

Questions for authors

1.The paper does not appear to clearly state whether CausalRank uses independent validation data or tasks during pretraining and fine-tuning, or how its parameters are selected. Could the authors clarify whether independent validation data are used and whether the test set enters parameter tuning?

2.DirectLiNGAM and NOTEARS learn causal structures from which the target’s parents can be extracted. Why do these methods, which directly address parent identification through structure learning, nevertheless underperform methods such as Lasso and LightGBM on Full-FCM?

  1. In Semi target-generating mechanism, each direct parent enters the target equation through a cubic nonlinear basis function with its first-order linear projection removed, reducing the usefulness of simple linear correlation. Does this disadvantage linear baselines and thereby make the comparison unfair to them?
4 comments

We thank Reviewer cvLz for the comments and constructive feedback. Below we address your questions in detail.


W1: "Limitations of Target-Specific Causal Modeling"

A:

We agree that a complete causal graph provides richer structural information about the relationships among variables. CausalRank primarily focuses on identifying the causal effects of financial characteristics on future stock returns. For this objective, identifying the direct parents of the target return provides the necessary causal information without the additional computational cost of recovering the complete causal structure among all financial characteristics.

Moreover, the target-conditioned architecture of CausalRank can, in principle, be extended to other target variables. By treating a characteristic of interest as YY, we can further explore the local causal relationships of other variables. However, such an extension would require additional pretraining and fine-tuning, which would introduce extra computational costs.

We will clarify the differences and connections between our approach and complete causal graph recovery in the revised manuscript.


W2: "Rationale and Effectiveness of the Model Design"

A:

We thank the reviewer for the suggestions regarding our model design. We further explain the rationale behind the statistical features and network architecture, and provide additional ablation experiments to validate their effectiveness.

Prior work has shown that nonlinear variable transformations and basis function expansions can capture statistical information that conventional linear correlations may miss [1]. Research on financial return prediction also highlights the importance of modeling nonlinear relationships between financial characteristics and returns [2]. Motivated by these findings, we select four simple transformations with different functional properties to preserve linear relationships, capture saturating nonlinearities, emphasize changes in variable magnitude, and represent non-monotonic responses, respectively. These transformations construct statistical features through simple element-wise operations and correlation calculations, providing complementary forms of marginal dependence information at low computational cost in high-dimensional financial panels without additional learnable parameters. Our experiments empirically support their effectiveness, although we cannot exhaustively establish their optimality among all possible alternatives.

We selected our combination of statistics based on the requirements. Regarding statistical information, Target-Factor Statistics provides explicit marginal relationships between XX and YY, while Conditional Incremental Evidence supplements these with conditional relationship information obtained from joint Ridge regression coefficients and partial correlations after controlling for other candidates.

Regarding the network architecture, the Panel Representation Encoder extracts cross-sectional and temporal information from financial panels. Target-Aware Parent Interaction combines target relationships with competition among candidate variables to distinguish direct parents from correlated non-parents. The Causal Ranking Decoder then learns direct-parent existence and intervention-defined causal effect strengths.

To further evaluate the effectiveness of combining these statistical features, we conducted four ablation experiments on Full-FCM. All other training settings, random seeds, and evaluation protocols are identical to those of the full model.

Target-Factor StatisticsConditional Incremental EvidenceAUPRCNDCG@20Parent Spearman
✗✗0.261 ± 0.0130.202± 0.0260.028 ± 0.028
✓✗0.585 ± 0.0300.651 ± 0.0380.541 ± 0.034
✗✓0.850 ± 0.0140.790 ± 0.0100.410 ± 0.025
✓✓0.984 ± 0.0010.965 ± 0.0010.809 ± 0.023

The new ablation results show that removing either type of statistical information leads to a performance decline, while the full model achieves the best results. Furthermore, the existing ablation experiments in our paper provide additional support for key design choices in the network modules and training objectives. Removing Candidate Competition reduces AUPRC to 0.7948, while introducing causal-effect and ranking supervision improves NDCG@20 from 0.9021 to 0.9647 and Parent Spearman from 0.6334 to 0.8093. Together, these results support the rationale behind our model design.

We will include the above design rationale and experimental results in the revised manuscript.

[1] Breiman, Leo, and Jerome H. Friedman. "Estimating optimal transformations for multiple regression and correlation." Journal of the American Statistical Association 80.391 (1985): 580-598. [2] Gu, Shihao, Bryan Kelly, and Dacheng Xiu. "Empirical asset pricing via machine learning." The Review of Financial Studies 33.5 (2020): 2223-2273.

Q1: "Data Splits and Parameter Selection"

A:

We thank the reviewer for pointing out this omission. We apologize for not adequately describing the data splits and model selection procedures. We clarify these details below.

Both pretraining and fine-tuning use independent training, validation, and test sets. The validation sets are used only for checkpoint selection, while the test sets in both Full and Semi experiments remain strictly held out and are never used for training, hyperparameter tuning, or model selection.

The model architecture and training hyperparameters were empirically determined before the experiments, based on panel dimensions, model capacity, and training stability. The main settings include a representation dimension of 64, 8 attention heads, and a feedforward dimension of 128. The learning rates for pretraining and fine-tuning were set to 2×10−42\times10^{-4} and 3×10−53\times10^{-5}, respectively, and gradually decreased during training. All these hyperparameters were predetermined and were not adjusted based on test results.

We will provide the complete data splits and model parameters in the revised manuscript.

Q2: "Performance of Causal Discovery Baselines"

A:

We thank the reviewer for raising this question. DirectLiNGAM and NOTEARS are representative methods for causal structure learning. We believe their relatively weak performance on Full-FCM mainly reflects differences between their modeling assumptions and the data-generating mechanisms. DirectLiNGAM assumes linear non-Gaussian structural equations with independent noise, while the original NOTEARS learns causal graphs using linear structural equations. Full-FCM includes nonlinear main effects, multiplicative interactions, and heteroscedastic mechanisms, which go beyond some of the assumptions of these standard models. When the structural equation assumptions do not fully match the underlying generating mechanisms, causal structure recovery may be affected, leading to lower accuracy in identifying the target's direct parents. In contrast, although Lasso also uses a linear model, it can directly exploit predictive signals for feature selection.

Moreover, the optimization objectives of these methods differ from ours. DirectLiNGAM and NOTEARS primarily focus on recovering complete causal graphs, whereas our task additionally requires identifying the direct parents of a specified target and ranking them by their intervention-defined causal effect strengths. We use the absolute weights of incoming edges to the target as ranking scores for both methods, but these weights may not accurately reflect the intervention-defined effect strengths under nonlinear and interaction mechanisms.

Therefore, although DirectLiNGAM and NOTEARS are well-established causal discovery methods, their relatively weak performance on Full-FCM may mainly stem from the mismatch between their standard modeling assumptions and the diverse causal generating mechanisms considered in our study.

Q3: "Fairness to Linear Baselines"

A:

The Semi dataset is designed to evaluate CausalRank's ability to identify nonlinear direct causal relationships and rank their effect strengths under real financial covariate distributions. The purpose of removing the first-order linear projection is to evaluate whether models can identify causal relationships beyond first-order linear information, rather than to disadvantage any particular baseline. Full-FCM and Semi serve different but complementary evaluation purposes. Full-FCM covers diverse causal generating mechanisms, including settings where first-order linear signals remain informative. Semi preserves the distributional properties and correlation structure of real financial covariates while evaluating performance under a nonlinear target-generating mechanism. The baselines is evaluated on both datasets, with identical input data and evaluation protocols for all methods within each dataset. Together, these experiments provide a broader assessment across different causal mechanisms and data distributions, rather than relying on a single experimental setting.

We will clarify the design objectives and complementary roles of these two datasets in the revised manuscript.


We appreciate the reviewer's time and thoughtful feedback. We hope our responses address your concerns and strengthen your positive assessment of our work. We would be happy to answer any further questions.

Thank you for the rebuttal. The authors’ responses have resolved my questions. However, I still see some minor limitations in the method’s design, as noted in my review. My overall assessment therefore remains unchanged.

Thank you for your follow-up and consideration. We fully acknowledge that some minor limitations remain in the current design, and we will continue to address them in our future work. We sincerely appreciate your positive assessment and valuable suggestion. Thank you again for your time and careful consideration.

Reviewer_TJ2HOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the Evidence-Guided Action World Model (EGA-WM) for time-series forecasting under external evidence (e.g., promotion schedules, weather/demand forecasts). The authors formulate time-series forecasting with external evidence as an action-conditioned world model problem. EGA-WM constructs a latent action space via a learnable primitive bank, infers state-dependent actions based on current system states and external evidence, and incorporates these actions into latent state transitions with persistent temporal dynamics. Experiments are conducted on electricity market (EPF) and retail datasets (Rossmann, M5).

Critical strengths
  1. Formulating the dynamic response to external evidence as state-dependent latent actions within a world model paradigm is an interesting and intuitive idea.

  2. The framework is evaluated across diverse datasets featuring different temporal resolutions, horizons, and types of external evidence (forecast-based vs. event-based).

Critical weaknesses
  1. Figure 1 and Figure 2 exhibit clear signs of being generated by generic generative AI image tools. The text labels within these diagrams are severely distorted, unreadable, or blurry (e.g., garbled axis and text labels in Figure 1; unreadable primitive/decoder annotations in Figure 2). Incorporating raw AI-generated diagrams with corrupt text strongly undermines the academic rigor and professional quality of the manuscript.
  2. The model's final prediction formula is defined as Y^t=bt+[D(za)−D(z0)]\hat{Y}_t = b_t + [D(z^a) - D(z^0)], where btb_t is directly provided by a strong, independently pretrained forecaster (Sonnet). Thus, EGA-WM operates essentially as a residual refinement module built on top of a powerful base forecaster.
  3. In Table 1, the authors compare EGA-WM (which leverages a pretrained SOTA Sonnet backbone plus residual tuning) directly against other standalone baseline models (e.g., TimeXer, DAG, TiDE) trained from scratch. This introduces a major confounding variable: it is impossible to determine whether the performance gains stem from the proposed "action world model" mechanism or simply from the unfair advantage of inheriting the pretrained Sonnet base forecaster.
  4. The manuscript aggressively packages residual modification as an "Action-Conditioned World Model". However, the framework lacks actual environment interaction, ground-truth action annotations, or causal supervision. The "primitives" are learned solely via end-to-end MSE loss on the final forecast errors. Without empirical proof that these learned actions represent true causal mechanisms rather than merely overfitting residual noise, the conceptual positioning remains an over-claim.
Questions for authors

See Weaknesses.

2 comments

Thank you for taking the time to review our work. Although we have not yet had the opportunity to hear your specific thoughts on our paper, we hope that the revisions made in response to the other reviewers may also help address some questions you might have. We would therefore like to briefly summarize the main additions and improvements in this round of revision and share the updated PDF for your reference. All revisions are highlighted in blue in the updated manuscript.

1. Methodological Design and Theoretical Discussion. We further clarified the distinction between full causal graph discovery and local causal discovery, explaining the focus of our work on identifying the direct causal parents of a specific target variable. We also added further rationale for the statistical features used in our model and conducted additional ablation experiments to demonstrate the effectiveness of Target-Factor Statistics and Conditional Incremental Evidence.

2. Synthetic Experiments and Model Evaluation. We extended the Semi experiments by adding tests under other nonlinear mechanisms, including Sine and Tanh, to further evaluate the model's capabilities across different nonlinear data-generating mechanisms. We also clarified the training, validation, and test splits used in the Semi experiments.

3. Real Financial Experiments and Economic Value Analysis. We adopted an experimental setting with no covariate overlap, using data from 1957 to 1986 for fine-tuning and validation, and replaced the real financial experimental results in the paper with those obtained using the new model. In addition, we clarified the predictive performance reflected by ROOS2R^2_{\mathrm{OOS}} and its economic interpretation, reported the standard error and bootstrap confidence interval of the Sharpe ratio, and added portfolio performance results net of different transaction cost levels. We also revised the abstract accordingly.

We hope these additions and improvements provide a more complete basis for reconsidering the overall value of our work. If you find that the revised manuscript addresses any concerns you may have, we would greatly appreciate your favorable consideration in the final evaluation.

Thank you again for your time and consideration.

AuthorsOpenReview ↗

Gentle Follow-up on Our Response

Dear Reviewer,

Thank you very much for taking the time to review our submission. We sincerely appreciate your time and effort.

We are writing to gently follow up on our earlier message, as the comments appeared not to correspond to our paper. We understand that such mix-ups can happen, and we would be very grateful if you could kindly take another look at our submission when your schedule permits.

We have also shared our revised manuscript and a brief summary of the improvements made during the rebuttal period. We hope these revisions may help address any concerns you might have about our work. We would greatly appreciate any further comments you could share and, if you feel it is appropriate, a reconsideration of your rating based on the revised manuscript. We would be happy to provide any additional clarification or results if needed.

Thank you again for your time, understanding, and consideration. We wish you all the best in your research.

Best regards, The Authors

Reviewer_EjQKOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper trains a model, CausalRank, on synthetic panels with known causal structure so that, given a panel of stock characteristics X and next-month returns Y, it outputs a score per characteristic which is the product of a parent existence probability g and an effect strength a. Supervision comes from parent labels and an intervention defined effect size (Section 3.3), through a BCE loss, an MSE loss on effects and a pairwise ranking loss (Appendix A.4). The main design idea is a leave-one-out attention over the other candidates, so each candidate is judged against what the rest already explain (Section 4.1.2). It is evaluated on fully synthetic and semi-synthetic test tasks against Lasso, LightGBM, DirectLiNGAM, NOTEARS and TCD-DL (Table 1), on correlation traps with constructed proxies (Section 5.2), and on US stocks 1987 to 2016, where the Top-20 characteristics from each ranker go into an Elastic Net and are judged by R2, Sharpe, return and turnover (Table 3).

Critical strengths

The formulation is the right one for asset pricing. You do not need the full graph among 94 characteristics, you need the direct parents of next-month returns and their order, and defining the ranking target through an intervention contrast between the 20th and 80th quantiles (Section 3.3) is a clean way to get that out of a synthetic generator. The s = g times a decomposition and the log additive ranking loss (Appendix A.4.3) are simple and they work. The Existence-Only ablation shows the effect supervision actually buys ranking quality, Parent Spearman goes from 0.63 to 0.81 on the fully synthetic set (Table 1). So this is good.

Critical weaknesses

The real data result does not support the abstract. CausalRank Top-20 has the lowest out-of-sample R2 of all seven methods, 0.19% against 0.23% for Random Top-20 and 0.26% for All Factors (Table 3), and the text describes this as retaining predictive information. The Sharpe of 0.77 against 0.74 for All Factors comes with no standard error and no test, and over 30 years a rough standard error on a Sharpe of 0.77 is about 0.2, so 0.77, 0.74 and 0.59 are basically not separated. The one large effect is turnover, 58.72% against 102% to 167%. But the characteristics selected every year are mvel1, rd_sale, divi, divo and securedind (Table 7), which barely move month to month, while momentum and volatility (mom12m, mom1m, retvol, idiovol, maxret) sit in the lower half of Figure 12. So low turnover and low R2 are both what you get from picking slow characteristics, not evidence that the picked ones are causal. Please report Sharpe standard errors or a bootstrap and net of cost returns, and say plainly that R2 is the lowest. Otherwise I think “higher risk premium” should not be in the abstract.

Questions for authors
  • Were the semi-synthetic fine-tuning tasks generated with the same cubic target mechanism as the semi-synthetic test tasks, and which calendar period of the real panels did they use? If the panels overlap 1987 to 2016 then the real data experiment is not out-of-sample on the covariate side, and if the mechanism matches then the Semi column of Table 1 is in-distribution for CausalRank only. This is the question which matters most for my rating.

  • Does the training generator ever leave a true parent of Y out of the candidate set, or put a latent common cause behind Y and some of the candidates? Appendix B.2 only says nonparents may be dependent on parents. In real data most of what moves next-month returns is not in the 94 characteristics, so I would like to know what g_j in Figure 11 actually means when the model has never seen a hidden parent.

4 comments

We thank Reviewer EjQK for the comments and constructive feedback. Below we address your questions in detail:


W1: "Insufficient Support for the Abstract from the Real Data Results"

A:

We thank the reviewer for the thorough analysis of our real financial experiments and the constructive suggestions.

1.Predictive Accuracy and Risk-Adjusted Returns

We agree with the reviewer's assessment of both metrics. Although CausalRank Top-20 has a positive ROOS2R^2_{OOS}, indicating some predictive information, it is still lower than that of competing methods. We will clarify this in the revision. CausalRank aims to identify direct causal parents rather than minimize mean squared predictive error. In other words, CausalRank seeks to identify characteristics that directly affect the first moment of future returns, namely expected returns, which is central to asset pricing, whereas ROOS2R^2_{OOS} evaluates predictive accuracy based on squared errors. These objectives do not necessarily favor exactly the same characteristics. By estimating causal effects from expanding historical data, CausalRank may favor characteristics with more persistent effects on expected returns across periods.

Following the reviewer's suggestion, we added a paired time-series bootstrap test for the Sharpe ratios. CausalRank's Sharpe ratio is 0.772 (SE 0.181, 95% CI [0.400, 1.111]). The difference in Sharpe ratio (0.028, 95% CI [-0.419, 0.427]) between CausalRank and All Factors is not statistically significant. However, the strategy using only 20 characteristics achieves a gross annualized return comparable to that using all of them while significantly reducing turnover (58.72% vs. 153.80%). Moreover, the remaining 74 characteristics yield only a 0.5677 Sharpe ratio (SE = 0.1980, 95% CI [0.1887, 0.9627]), 9.41% annualized return, and 150.66% turnover, which further supports the economic value of CausalRank's feature selection. We also show the results under one-way transaction costs:

Net Return / SharpeCausalRankAll Factors
5 bps11.73% / 0.72810.25% / 0.630
10 bps11.03% / 0.6848.42% / 0.517
20 bps9.63% / 0.5964.75% / 0.291

CausalRank has higher net return and Sharpe ratio point estimates, with a larger advantage as costs rise.

2. Turnover and Causality of Selected Characteristics

We agree that low turnover alone cannot establish causality. While slow-moving characteristics may partly explain low turnover, they do not necessarily lead to better portfolio returns. Conversely, even fast-moving characteristics may result in low turnover if the resulting portfolio positions remain stable. CausalRank selects characteristics based on their estimated causal effects on expected returns, regardless of how quickly or slowly they change. Thus, we view low turnover as an outcome of causal-effect-guided selection.

Moreover, we used the mean absolute change in rank and its autocorrelation to measure stability for all 94 characteristics. We find that the top-20 selection by CausalRank sits in the middle, around 60% in the full set, in terms of stability. We also construct a portfolio using 20 characteristics with similar temporal stability from the remaining 74. With comparable ROOS2R^2_{OOS} and turnover, its Sharpe is only 0.442 vs. CausalRank's 0.772. We note that CausalRank does not optimize portfolio returns or factor stability; instead, it selects characteristics from inferred direct causal relationships and effect strengths. Yet its selections achieve both lower turnover and favorable economic performance. We therefore consider it reasonable to conjecture that their causal properties help explain these benefits.

Furthermore, although ground-truth causal labels are unavailable in real data, CausalRank's selections do have economic foundations. For instance, R&D intensity (rd_sale) relates to intangible capital accumulation and long-term operating performance [1], while firm size (mvel1) is a classic characteristic associated with cross-sectional returns [2]. This supports their interpretation as potential long-term causal drivers. Meanwhile, the causal structure of financial markets may change across market conditions. In future work, we hope to incorporate both short-term and long-term causal parents to improve causal interpretability and predictive accuracy.

In summary, we thank the reviewer for these suggestions. We will explicitly report the above statistical tests and additional experimental results in the revised manuscript, and revise the phrase "higher risk premium" in the abstract to more accurately describe the economic value of CausalRank.

[1] Eberhart, Allan C., William F. Maxwell, and Akhtar R. Siddique. "An examination of long‐term abnormal stock returns and operating performance following R&D increases." The Journal of Finance 59.2 (2004): 623-650.

[2] Fama, Eugene F., and Kenneth R. French. "The cross‐section of expected stock returns." The Journal of Finance 47.2 (1992): 427-465.

Q1: "Data Splits and Covariate Overlap"

A:

We apologize for not including data splits in the manuscript.

1. Semi Benchmark and Mechanism Generalization

The training, validation, and test panels of Semi fine-tuning cover 1957.01–2002.06, 2002.07–2012.03, and 2012.04 onward, with strict temporal separation. Both fine-tuning and testing use the cubic family, but test panels, parent sets, and structural coefficients are generated independently.

The reviewer's observation that this mechanism is in-distribution reflects different learning paradigms. CausalRank learns causal-mechanism priors through cross-task pretraining and can adapt to other mechanisms with suitable structural priors by constructing new datasets. Other baselines instead fit each test task independently, without a pretraining mechanism distribution.

To further examine whether CausalRank's performance depends on the test mechanism, we freeze the original checkpoint and test Orthogonalized Sine and Tanh:

MechanismMethodAUPRCNDCG@20Parent ρ
SineCausalRank0.790.770.49
LightGBM0.710.710.40
Nonlinear Correlation0.660.650.41
TanhCausalRank0.800.780.45
LightGBM0.750.740.40
Nonlinear Correlation0.620.610.27

CausalRank remains best under both, so its performance is not limited to the mechanism of Semi fine-tuning. However, we acknowledge that CausalRank does not yet generalize to unseen causal mechanisms.

2. Covariate Overlap in the Real Financial Experiment

Semi fine-tuning uses real financial covariates to adapt to their joint distribution P(X)P(X) and distinguish parents from correlated nonparents. The same XX can have different target causal structures, e.g., Y(1)=f1(XP1)+ε1Y^{(1)}=f_1(X_{P_1})+\varepsilon_1 and Y(2)=f2(XP2)+ε2Y^{(2)}=f_2(X_{P_2})+\varepsilon_2, with P1≠P2P_1\neq P_2. Semi generates targets, parent sets, and coefficients independently, so sharing covariates does not expose the model to causal supervision from realized stock returns. Based on this causal perspective, we did not strictly separate covariates. However, the reviewer's concern is valid, as exposure to future covariates may affect the financial backtest. We apologize that our initial separation was not sufficiently thorough and conduct the following experiment to address it.

We therefore fine-tune a new model from the same Full-pretrained model on 1957–1983, using 1984–1986 for validation. We evaluate both models over the same 1987–2016 period using identical financial evaluation protocols.

MetricPre-1987 modelOriginal modelΔ (95% CI)
ROOS2R^2_{\mathrm{OOS}} (%)0.2340.195+0.040 [−0.026, 0.105]
Sharpe0.7750.772+0.004 [−0.285, 0.289]
Return (%)13.0312.43+0.60 [−3.76, 5.18]
Turnover (%)54.4858.72−4.24 [−11.66, 3.78]

The new model has higher point estimates for all predictive and return metrics. The correlation of selected characteristics between models is 0.932, and the rank correlations of mean gjg_j, aja_j, and sjs_j are 0.977, 0.971, and 0.972. Thus, CausalRank still maintains stable rankings and economic performance with or without covariate overlap. The evidence also does not suggest that overlap benefited the original model.

In the revised paper, we will replace all real data experiment results with those obtained from a model trained without covariate overlap, and retest the Semi experiments under multiple mechanisms.

Q2: "Unobserved Direct Parents and Latent Confounding"

A:

Thank you for raising this question. We considered an observable common cause to help distinguish direct parents from correlated nonparents. However, the current synthetic data do not explicitly simulate unobservable direct parents or unobserved common causes. Our training setup thus assumes all causally relevant factors needed to identify YY's direct parents are included among the candidates. We adopt this setting as a starting point, where a controlled setting provides exact structural supervision for learning direct-parent identification and causal-effect ranking in financial characteristics.

We agree that real markets may contain return drivers outside the 94 characteristics. When the assumptions fail, gjg_j in Figure 11 is an empirical direct-parent score based on observables and may reflect effects associated with hidden nodes. Future work will incorporate unobserved variables and study performance and applicability under more complex causal structures. We will clarify these assumptions and the interpretation of gjg_j in the revision.


We again thank the reviewer for the careful review and helpful suggestions. We hope these responses address your concerns and support reconsideration of our work for a more favorable score. If you still have any questions or need additional materials, please feel free to let us know. We will do our best to answer.

Thank you for your time. We have uploaded a revised PDF incorporating the requested changes. All revisions are highlighted in blue in the updated manuscript.

We extended the Semi experiments to additional nonlinear mechanisms and replaced the real financial results with those obtained from a model fine-tuned and validated exclusively on pre-1987 data, eliminating covariate overlap. We also clarified the interpretation of ROOS2R^2_{\mathrm{OOS}}, reported the Sharpe ratio's standard error and bootstrap confidence interval, added net-of-cost portfolio results, and revised the abstract. Furthermore, we clarified the causal sufficiency and direct-parent coverage assumptions, as well as the interpretation of gjg_j in real financial data.

Please note that the numerical results discussed in our previous response to the Weaknesses section were based on the original model, whereas the revised manuscript reports results from the new model. The differences are minor and do not affect our main conclusions. We hope these revisions address your concerns and would appreciate your reconsideration and, if warranted, a more favorable rating.

AuthorsOpenReview ↗

Gentle Follow-up on Our Response to Your Review

Dear Reviewer,

Thank you very much for taking the time to review our paper and for your thoughtful and constructive comments. We highly appreciate your suggestions on clarifying the presentation and interpretation of our financial results. We agree that the data splits used for fine-tuning on the Semi dataset warranted more rigorous validation, and we have made every effort to address your concerns through additional analyses and revisions.

We are writing a gentle follow-up and would be very grateful if you could take a moment to review our responses when your schedule permits. Please feel free to let us know if anything remains unclear. We would be happy to provide further clarification or additional results.

Thank you again for your valuable feedback, time, and consideration. We wish you all the best in your research.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.