acceptodds
Under review as a conference paper at ICLR 2027

Decorrelated Is Not Diversified: Style-Robust Selection and Search of Formulaic Alpha Factors

Abstract

Formulaic alpha factors are interpretable predictive signals mined from market data, and recent work explores the formula space with reinforcement learning and generative models rather than genetic programming. The mined pools, however, are still screened by the information coefficient (IC) and de-duplicated by pairwise correlation of factor values, and the resulting portfolios lose predictive power when market style rotates. We trace this to a blind spot in the selection criterion: IC measures fit to the prevailing regime, and mutual-IC screening removes redundancy only in factor values, so a value-decorrelated set can still concentrate on a single capitalization exposure and fail in unison when that exposure rotates. We propose the Exposure–Cycle Residual (ECR), a behavioral descriptor that characterizes a factor by the direction of its net exposure across capitalization layers and by the phase and period of its capitalization tilt, and defines a candidate's novelty as its residual energy relative to the factors already admitted. Screening on ECR produces factor sets whose measured IC converts into realized return more stably across style regimes than IC or mutual-IC screening. We further embed the residual into search itself as a reward-shaping term over an online pool, steering generation toward exposures and cycles not yet covered. Finally, we present a controlled comparison of formulaic miners at high feature dimensionality, with hundreds of tick-derived fields, a shared single-formula evaluator, and both linear and nonlinear combiners, re-running random search, genetic programming, and recent reinforcement-learning and generative miners under identical screening. Across screens, miners and combiners, ECR-based selection and ECR-guided search consistently improve out-of-sample predictive power and portfolio return, with the largest gains during style-rotation periods.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
formulaic alpha mining, style rotation, regime robustness, behavioral novelty, quality-diversity search, reinforcement learning, nonlinear factor combination

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_Wzi7OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies redundancy in formulaic alpha mining beyond pairwise correlation of factor values. It argues that factors can be decorrelated in value space while remaining concentrated in a common capitalization-related exposure and therefore fail jointly under style rotation. The paper introduces the Exposure–Cycle Residual (ECR), a descriptor intended to characterize where along the capitalization spectrum a factor is effective and the temporal phase/persistence of that exposure. ECR is used both as a selection criterion and as a reward-shaping term during formula search. The paper further compares several formula-mining procedures under a shared evaluation pipeline and reports improved out-of-sample IC-related metrics and label Sharpe for ECR-guided search. I find the general distinction between value decorrelation and common exposure potentially interesting. However, in its current form, the experimental evidence does not support the strength or generality of the paper’s conclusions, particularly the claims about style robustness and about where diversity should enter the mining pipeline.

Critical strengths
  1. The paper identifies a meaningful limitation of factor-value correlation as a redundancy criterion. Two factors can be weakly correlated in their realized values while depending on similar market conditions. Asking whether formula mining should explicitly encourage diversity in conditional exposures is a useful research question, and one that is not fully addressed by standard mutual-IC screening.
  2. The attempt to evaluate multiple mining families under a shared downstream pipeline is valuable in principle. Formula-mining papers are often difficult to compare because they use different expression spaces, evaluators, screens, and combiners. A controlled common evaluation framework could therefore be a useful contribution if the search budgets and experimental conditions were made genuinely comparable.
Critical weaknesses

1. The claims about style robustness are not fully supported by the experiments.

ECR mainly captures how a factor performs across capitalization buckets over time, rather than its exposure to broader investment styles. Since its signature also uses realized training-label performance, it partly reflects where a factor has fitted the data. The paper does not explicitly evaluate sector, momentum, volatility, or other style rotations. I therefore think the claims about general style robustness should be narrowed. The revised OpenReview abstract also appears stronger than the attached manuscript, particularly regarding style regimes and ECR-based selection.

2. The large improvements need stronger statistical evidence and clearer attribution.

All experiments use only seed 0, despite involving highly stochastic search procedures. Given the substantial reported Sharpe improvements, repeated runs and uncertainty estimates are important.

Moreover, the strongest batched-search result changes several components simultaneously, including the reward, archive admission, and policy update. It is therefore difficult to isolate the contribution of ECR itself. I would particularly like to see multi-seed results and an ablation that varies only the ECR component.

3. Some comparisons do not adequately isolate the proposed mechanism.

The selection baseline uses a two-dimensional projection rather than the full ECR signature and ignores IC when selecting factors. Its poor performance does not necessarily show that diversity should be encouraged during search rather than selection.

Similarly, Figure 1 compares candidate pools with different sizes and filtering procedures. The control pool is truncated by training IC, whereas the ECR pool is not. A comparison using consistent filtering and selection criteria would be more convincing.

4. The experimental comparisons and financial interpretation need clarification.

Search budgets vary substantially across miner families, and some baselines have modified objectives. This makes it difficult to draw conclusions about the relative effectiveness of different mining algorithms.

I am also concerned about the use of terms such as "portfolio return" and "Sharpe." The main metric is based on standardized labels rather than actual tradable portfolio returns. Although this distinction is acknowledged in the manuscript, it should be communicated more consistently, especially in the abstract.

5. Reproducibility remains limited.

The supplementary code omits several essential components, including operator definitions and the execution environment. I could not run the provided mock baseline because a required module (model_core_v3.vm) was missing.

In addition, the default configuration includes mechanisms that are not clearly described in the manuscript. Without the exact experimental configurations and run logs, it is difficult to verify which components produced the reported results. Providing a minimal runnable example and the configurations used for the main experiments would substantially improve reproducibility.

Questions for authors
  1. Robustness and causal attribution. Can the authors provide matched multi-seed experiments in which ECR and non-ECR methods use identical proposal budgets, identical post-filter candidate-pool sizes, and identical screening rules? In particular, for the batched method, can the authors isolate the ECR residual itself from archive admission, shape rewards, and batch penalties? Confidence intervals or paired significance tests would materially affect my assessment.
  2. Selection versus search. The conclusion that diversity should be introduced during search rather than selection appears to be based on farthest-point sampling in a lossy two-dimensional projection of ECR, while the search reward uses the full signature. Can the authors repeat the selection experiments using the same full ECR representation, for example with a joint IC/diversity criterion, MMR, DPP-style selection, or full-signature max-min selection? Without this comparison, I do not think the current experiments justify the stated conclusion.
  3. Meaning of “style robustness.” Because the ECR descriptor is specifically capitalization-based and partially constructed from realized label efficacy, what evidence shows that the method is robust to style rotation in a broader sense rather than simply diversifying capitalization-conditioned modes of training-period predictability? Results using additional ex-ante style dimensions, or a clearer narrowing of the claim, would materially change my view.
Reviewer_BnYsOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper argues that two formulaic factors can have nearly uncorrelated values and still earn their IC on the same part of the capitalization spectrum and on the same rhythm. It defines an Exposure–Cycle Residual and multiplies a candidate’s fitness by the share of that behavior not yet stored in an online archive. Eight miners are rerun in one formula space on Chinese A-shares. Under a mutual-IC screen, adding the residual raises out-of-sample label Sharpe while IC moves little. The authors conclude that this diversity belongs in the search reward, and that the final screen should keep the IC order.

Critical strengths

1.The placement question is the right one. The paper asks whether behavioral diversity should enter the reward or the final selection, and it reports that IC-blind farthest-point selection lowers the label Sharpe of the pools already mined with the residual.

2.One comparison is isolated. For tournament GP on RPN, the residual is the only change, and the mutual-IC cap is the same with and without it. The paper also states that the reported Sharpe is a moment of a standardized label, not a tradable return.

Critical weaknesses

1.Every number is a single run with seed 0. The claimed gains are about four label-Sharpe points for batched search and for tournament GP. Search variance is large enough that one seed cannot support that difference, and the paper reports no second seed and no uncertainty.

2.The headline comparison does not isolate the residual, and the metric is not a portfolio result. Batched search with the residual also changes archive admission, the policy update, a shape bonus, and a batch penalty. Its pool is not truncated, while the control is cut to the top 1,000 formulas by training IC, and the two sets fill only at mutual-IC caps of 0.30 and 0.90. The label itself is a within-capitalization-bucket standardized residual. A label Sharpe near 9 is a property of that series, not evidence of a tradable book.

3.The baselines are too weak to rank the objective. Uniform RPN sampling, which reads no score, reaches a higher label Sharpe than both IC-trained controls and the reimplemented AlphaSAGE. Until those controls are competitive, the ranking in the last experiment cannot be read as evidence that behavioral diversity is what separates the miners.

Questions for authors

1.Do the tournament-GP gains, where only the reward changes and the cap is held fixed, remain across several seeds?

2.On the same selected factors, what is the net long–short return after costs, rather than the Sharpe of the standardized label?

3.Why does uniform sampling beat the trained controls? Were those miners given a comparable proposal budget and a tuned objective?

1 comment

Thank you for recognizing our placement question and the causal isolation in Tournament GP on RPN. Below we address your three questions on multi-seed variance, reward isolation and label design, and the Uniform RPN sample ordering via new experiments in Section 5.3, Figure 5, Appendix F, and Appendix G (Tables 5-6).

1. Multi-seed stability of isolated Tournament GP on RPN (Q1 & W1 -> Section 5.3, Figure 5, Appendix G Tables 5-6) Inspecting Seed-0 GP (4.29) linked W1 and W3: unconstrained chains of >= 3 same-family operators inflated in-sample F alongside behavioral crowding. Bounding same-family chains to depth <= 2 in both arms yields a strong 5-seed GP baseline (5.63 ± 1.79 OOS Sharpe across Seeds 0-4, per-seed values in Table 5):

  • Mean, variance, and floor (Table 5, Figure 5 left): Adding Equation 5 alone lifts 5-seed mutual-IC Sharpe to 6.03 ± 1.18 (+0.40 paired gain), cuts cross-seed SD by 34%, and raises the worst-seed floor from 2.69 to 4.33. As you suspected, +4 jumps occur specifically when an unregularized seed collapses onto one regime (Seed 0: 2.69 -> 6.62), while on naturally diverse seeds ECR preserves peak Sharpe. Search-time ECR + mutual-IC (6.03 ± 1.18) also beats unregularized search + farthest-point selection (5.20 ± 1.55).
  • Left-tail risk across 3,170 pooled OOS seed-days (Table 6, Figure 5 right): Pooling all 5 x 634 = 3,170 OOS days shows ECR raises the daily top-100 mean, lowers daily SD, and compresses 1% and 5% left-tail regime drawdowns while matching 95%/99% right-tail payoffs.

2. Isolating ECR, unbucketed return labels, and fee-simulation scope (Q2 & W2 -> Sections 4.2, 5.1, 5.3, Table 2, Appendix G)

  • Isolating the reward: We agree Batched search + ECR alters admission, batch penalties, and truncation (Changed: R, A, U). Sections 4.2 & 5.1 and Table 2 now anchor our isolated causal claim on Tournament GP on RPN (Changed: R only) and frame Batched search + ECR strictly as a system-level GPU loop.
  • Unbucketed return labels (Section 5.3, Page 10; Appendix G): Retraining and evaluating Tournament GP (depth <= 2, 2,000 proposals) on an unbucketed full-market cross-sectional z-score of raw close-to-close returns (with the 5-layer ECR descriptor fixed) still raises OOS Sharpe from 9.08 to 9.86, confirming our gains do not depend on bucketed labels.
  • Why we avoid unbuffered fee deductions: Broad single-stock shorting is unavailable in A-shares, and F(f), the mutual-IC screen, and the fixed published AlphaForge combiner are all 1-day turnover-blind (~90% daily turnover). Charging stamp duty to unbuffered 1-day rankings punishes baselines for trading churn and market beta rather than redundancy, whereas adding a turnover buffer breaks the fixed-combiner control.

3. Why Uniform RPN sample (6.79) beat several Seed-0 baselines (Q3 & W3 -> Section 5.2 RQ4, Appendix F) All single-formula miners share the exact 2,000-proposal budget, objective F(f), filters, and causal combiner (Appendix F, Page 16). Three effects explain Seed 0:

  1. 388-token vocabulary (344 fields + 44 ops): Sequential policies (AlphaSAGE, AlphaGen), designed for ~6 OHLCV fields, under-converge in 2,000 episodes (196 and 460 eligible formulas, Appendix F) when kept faithful without custom field pruning (Section 1).
  2. Filtered, combined random sampling avoids regime collapse: Only 296 Uniform RPN formulas pass |IC| >= 0.01 into the expanding |RankICIR| > 0.15 combiner, and score-blind sampling naturally spreads across behaviors (0.30 cap).
  3. Fast-converging miners over-concentrate without regularization: Adding Equation 5 alone lifts unconstrained Seed-0 GP from 4.29 to 8.64 (above 6.79), and bounding depth <= 2 strengthens GP across 5 seeds (5.63 -> 6.03) and on full-market z-score returns (9.08 -> 9.86).
Reviewer_gTMZOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies behavioral redundancy in formulaic alpha mining. The authors argue that mutual-IC screening, which removes factors with correlated outputs, can miss factors that rely on similar market exposures or exhibit similar temporal patterns. They propose Exposure–Cycle Residual (ECR), a behavioral descriptor based on how a factor’s ranking efficacy varies across capitalization layers and over time. ECR compares a candidate’s descriptor with an online archive and scales its search reward according to the residual behavior not already covered. The authors evaluate ECR in a shared formula space using eight mining procedures on Chinese A-share data. They report that ECR improves several out-of-sample metrics under mutual-IC screening, particularly for batched search and tournament GP on RPN. They also compare search-time behavioral diversity with a farthest-point selection rule and report that behavioral selection can help pools mined without ECR but can hurt pools mined with ECR.

Critical strengths
  1. The paper studies a meaningful limitation of standard factor screening. Mutual-IC measures similarity between factor outputs, but it does not directly characterize whether factors depend on similar capitalization exposures or temporal behavior. ECR offers a concrete and interpretable way to investigate this issue.

  2. The experimental framing separates search and selection. The paper compares ECR as a search reward with behavioral farthest-point selection, and includes matched ECR/control pairs for multiple search procedures. This is a useful design for studying where diversity should enter a factor-mining pipeline.

Critical weaknesses
  1. The empirical evidence is insufficient to establish that the reported gains are reliable or general. All experiments use a single random seed and one chronological train/test split, with no variability estimates across runs or test periods. This is particularly concerning given the large reported differences. The paper does include ECR/control comparisons, but the batched-search pair changes multiple components in addition to the reward, including archive admission and batch-level terms, so it cannot isolate the contribution of the ECR reward. The tournament-GP and REINFORCE pairs more closely isolate the reward, but they are still based on one run and one test period; moreover, the REINFORCE improvement is small. Overall, the results are promising but do not yet support a robust or general claim that ECR improves search.

  2. The paper does not sufficiently establish ECR’s methodological contribution relative to closely related diversity-based alpha-mining methods. ECR’s capitalization-exposure and temporal-cycle descriptor is a potentially useful contribution. However, prior work such as AlphaSAGE already combines novelty incentives with structure-aware representations and rewards in formulaic alpha mining. The paper should more precisely explain how ECR differs from value-correlation novelty, structure-aware behavioral distance, and archive-based quality-diversity methods, and what capability its exposure-and-cycle descriptor adds. The current comparison to AlphaSAGE is not enough to establish that ECR’s particular descriptor or reward is responsible for the reported gains, especially given the different implementations and objectives.

  3. The contribution of the proposed reward is not sufficiently isolated against alternative reward formulations. ECR scales predictive fitness by an archive-based behavioral novelty term. The paper compares ECR with the original fitness in some miners, but does not sufficiently compare this reward formulation with closely related novelty or multi-objective reward designs under a controlled setting. As a result, it remains unclear whether the gains come from the capitalization-exposure/cycle descriptor, the archive-based novelty signal, or the particular multiplicative reward formulation. This question is especially relevant for batched search, where the ECR version also changes archive admission and batch-level terms.

  4. The behavioral descriptor and its conclusions are not sufficiently validated against the chosen label construction. The paper standardizes next-day returns within five capitalization buckets, and the proposed ECR descriptor measures factor effectiveness across those same capitalization layers. This may be a reasonable design, but it couples the label construction to the behavior being measured. The paper does not show whether the main findings persist under reasonable alternatives, such as different bucket counts or a label without within-bucket standardization. A sensitivity analysis is needed to establish that the observed gains are not specific to this construction.

Questions for authors
  1. Can the authors report the main results across multiple random seeds and additional chronological test periods, with uncertainty estimates for the ECR/control differences? For the batched-search comparison, can they disentangle the effects of the reward from archive admission and batch-level terms?

  2. How does ECR differ technically from prior diversity-based formulaic alpha-mining approaches, particularly AlphaSAGE’s novelty and structure-aware rewards, as well as value-correlation novelty and quality-diversity archives? Could the authors provide a more controlled comparison or targeted ablation that isolates the contribution of the capitalization-exposure and cycle descriptor?

  3. Could the authors compare the proposed multiplicative reward with controlled alternatives, such as an additive novelty bonus or a value-correlation novelty reward, while holding the miner, proposal budget, and evaluation pipeline fixed? This would help separate the contribution of the ECR descriptor from that of the reward formulation itself.

  4. Could the authors report a sensitivity analysis using alternative capitalization bucket counts or a label without within-bucket standardization, and explain whether the main ECR conclusions persist?

  5. Since the ECR-shaped reward multiplies the base fitness, can the authors clarify whether the base fitness can be negative in their experiments and how the scaling affects the reward and policy update in that case?

  6. Could the authors provide a portfolio-level evaluation of the selected factors, with a clearly specified construction protocol and transaction costs and turnover taken into account? Reporting net return, Sharpe ratio, and maximum drawdown would help assess practical investment implications. If this is outside the scope of the paper, the conclusions should remain focused on predictive performance.

1 comment

Thank you for your thorough review and clear framing of search-time vs. post-search diversity. Below we address all six questions with pointers to the new experiments and revisions.

1. Multi-seed variability, OOS regimes, and isolating ECR (Q1 & W1 -> Sections 4.2, 5.1, 5.3, Table 2, Figure 5, Appendix G Tables 5-6)

  • Isolating ECR: We agree Batched search + ECR modifies archive admission, batch penalties, and updates (Changed: R, A, U). Sections 4.2 & 5.1 and Table 2 now frame it as a system-level GPU loop and anchor our isolated causal test on Tournament GP on RPN (Changed: R only, Algorithm 1). (AlphaGen-style REINFORCE yields only 460 eligible formulas over 388 tokens in 2,000 episodes, leaving little exploitation pressure to concentrate.)
  • Five-seed GP ablation (Table 5, Figure 5 left): Bounding same-family operator chains to depth <= 2 in both arms strengthens the GP control (5.63 ± 1.79 OOS Sharpe across Seeds 0-4). Adding Equation 5 raises mutual-IC Sharpe to 6.03 ± 1.18, cuts cross-seed SD by 34%, lifts the worst-seed floor (2.69 -> 4.33), and beats unregularized search + behavioral farthest-point selection (5.20 ± 1.55).
  • Regimes & 3,170 pooled OOS seed-days (Figure 4, Table 6, Figure 5 right): Short test sub-windows starve the expanding |RankICIR| filter in the fixed combiner, whereas our continuous 634-day (~2.6-year) OOS span already spans sharp A-share size reversals (Figure 4). Across all 3,170 pooled seed-days (Table 6), ECR compresses 1% and 5% left-tail regime losses while preserving 95%/99% right-tail payoffs.

2. Distinction from AlphaSAGE, output-correlation novelty, and QD (Q2 & W2 -> Section 2, Page 3) Revised Section 2 contrasts ECR along three axes: (a) representation—output correlation (mutIC/PCA in AutoAlpha, AlphaGen, AlphaSAGE) and GNN syntax terms miss shared size exposure across disjoint small-cap subsets, whereas ECR tracks efficacy across size layers and regime phases; (b) archive query—an orthonormal linear-subspace residual (res_amp) plus phase/frequency coverage (res_cyc), rather than pairwise mutIC or fixed grids; and (c) selection—unlike QD niche outputs, post-search niche sampling lowers Sharpe (Tables 1 & 5) by dropping higher-IC signals.

3. Multiplicative ECR vs. correlation and additive bonuses (Q3 & W3 -> Section 4.1, Table 1, Appendix E-F) Under the identical engine and pipeline, AlphaSAGE's additive 1 - max |mutIC| + structure bonus (5.69 Sharpe) trails GP + ECR because output-correlation novelty duplicates the post-search mutual-IC screen. Algebraically (Equation 5), over 344 fields noise formulas easily hit novelty ≈ 1 with IC ≈ 0; additive bonuses reward zero-signal drift once strong behaviors fill, whereas multiplicative scaling keeps F = 0 at zero reward.

4. Unbucketed return labels (Q4 & W4 -> Section 5.3, Page 10; Appendix G) Retraining and evaluating Tournament GP (depth <= 2) on an unbucketed full-market cross-sectional z-score of raw close-to-close returns (with the 5-layer ECR descriptor unchanged) still lifts OOS Sharpe from 9.08 to 9.86, confirming ECR does not rely on bucket-normalized labels.

5. Negative base fitness F(f) < 0 (Q5 -> Section 4.2, Page 6) By sign invariance F(-f) = F(f) (Section 3.3), F < 0 occurs in only 7.5% (150/2,000) of Seed-0 GP proposals (Section 4.2). Scaling by [0.3, 1] shrinks covered losses toward zero while keeping all F < 0 candidates below positive ones in tournament selection; Batched search applies degenerate-panel penalties outside fill.

6. Signal quality vs. costed trading (Q6 -> Sections 3.1, 5.1, 5.3, Appendix D) Because F(f), the mutual-IC screen, and the fixed published AlphaForge combiner are 1-day turnover-blind (~90% daily turnover) and broad A-share shorting is unavailable, unbuffered stamp-duty deductions mainly measure trading churn and unhedged beta, while adding a turnover buffer breaks the fixed-combiner control. We clarify this scope and validate upstream signal quality via unbucketed z-score returns (9.08 -> 9.86) and 3,170-day tail compression (Table 6).

Reviewer_DJG4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces the Exposure–Cycle Residual (ECR), which describes a factor by where on the capitalization spectrum its ranking has been effective and by the period of that tilt. During search, ECR scales each candidate’s fitness by the share of its behavior that an online archive does not yet cover, so rediscovering a known behavior earns only a fraction of the fitness. Experimental results show that adding ECR to the reward raises the out‑of‑sample label Sharpe of the combined factor while its IC barely changes.

Critical strengths

This paper contributes in the following aspects: 1) It proposes ECR, a descriptor of where and on what rhythm a factor’s ranking works along the capitalization spectrum, and uses its residual energy against an online archive as a multiplicative term in the search reward. 2) Under mutual‑IC screening, the term improves all three miners it is added to. 3) It uses two coordinates of the same descriptor as a diagnostic plane rather than as a selection rule, and shows that IC‑driven search concentrates.

Critical weaknesses
  1. The problem formulation in Section 2 needs improvement. For example, the notations f⋅,tf_{\cdot,t} and g⋅,tg_{\cdot,t} in Equations (1) and (2) lack formal definitions.
  2. The experimental setup should cover broader datasets commonly adopted in related research. Current experiments are conducted only on Chinese A‑share stocks using 344 tick‑derived daily fields, and this dataset does not appear to be open‑sourced. This limits the persuasiveness of the proposed method.
Questions for authors
  1. why used 8 and 4 as hyper-parameter in this equation on Page 4 Line 182?
  2. See weakness.
1 comment

Thank you for your clear summary of our contributions and for pointing out where our notation, hyperparameter choices, and data access needed fuller explanation. We address each point below.

1. Formal definitions of cross-sectional slices in Equations 1 and 2 (Section 3.2, Page 4) Fixed. Immediately before Equations 1 and 2 in Section 3.2, we now explicitly define the daily stock-level factor and label cross-sections on date t, specify that the daily cross-sectional correlation is computed via Pearson correlation over stocks where both entries are finite (skipping dates with fewer than two valid pairs), and define the valid-date set for mutual-IC.

2. Rationale for weights 8 and 4 in Equation 4 (Section 3.3, Page 4) We added the explanation for both the 2 : 1 ratio and the overall scale to Section 3.3:

  • 2 : 1 ratio (|IC| vs. r_book): In cross-sectional equity models, extreme-quantile book returns have expectation proportional to IC, but their sample mean is substantially noisier because it averages only over tail names rather than all stocks (Grinold & Kahn, 2000). Giving full-cross-section |IC| twice the weight stabilizes the training signal while keeping r_book aligned with top-end payoff.
  • Overall scale (x4, giving 8 and 4 instead of 2 and 1): Scaling by 4 places typical training scores in an O(1) range (~0.3-1.0) convenient for policy-gradient updates without changing the relative trade-off; for rank-based Tournament GP on RPN, any positive scalar leaves tournament comparisons strictly invariant. All control and + ECR pairs optimize this identical F(f).

3. Dataset scope, label-decoupling check, and supplementary panel sample (Section 5.3, Appendix C Tables 3-4, Reproducibility Statement)

  • Why 344 A-share microstructure fields & label check: Behavioral crowding—where syntactically distinct, low-mutual-IC formulas quietly harvest the same size tilt—emerges at realistic feature widths (hundreds of order-book, trade-tape, and chip-state reductions) rather than 6-field OHLCV toys. Moreover, retraining and evaluating on an unbucketed full-market cross-sectional z-score of raw close-to-close returns (Section 5.3, Page 10) still raises OOS Sharpe from 9.08 to 9.86, confirming our findings do not hinge on the 5-bucket label design.
  • Auditable January 2021 panel slice in Supplementary Material: While commercial licensing prohibits redistributing the full multi-year tick feed, our supplementary archive includes anonymized code plus a 20-trading-day January 2021 daily panel slice (panel.npz, fields.csv, summary.json across 4,158 anonymized stocks and all 344 fields, documented in Appendix C and the Reproducibility Statement). It provides the daily features alongside raw returns, within-bucket excess returns, final labels, public AKShare log-capitalization (stock_value_em), and size bins—allowing direct inspection of all 344 fields, step-by-step verification of label construction, and interface testing of our code.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.