acceptodds
Under review as a conference paper at ICLR 2027

AlphaGEAR: A Self-Evolving Agent for Alpha Mining via Library-Guided Exploration and Auditable Refinement

Abstract

LLM agents can easily search alpha factors and explore program spaces at a scale difficult to achieve manually. The resulting candidates must then be selected and combined into a final predictor. Yet generating individually promising or diverse factors does not ensure that they contribute useful information to the final predictor, while repeated evaluation can amplify selection and feedback risks. We introduce AlphaGEAR, a self-evolving agent for constructing refined factor libraries. AlphaGEAR aligns search with deployment by guiding factor evolution with library-relative residual information and retaining complementary factors through stream-aware admission and marginal-contribution refinement for an equal-weight predictor. A discrete, query-limited certification interface makes additional held-out feedback bounded and auditable. Under a common equal-weight evaluation on CSI 300, AlphaGEAR achieves an average development RankIC of 0.0517 across three random seeds, compared with 0.0043 for the strongest competing agent, while retaining only 4-8 members against 23-44 in the competing libraries; the advantage persists after style orthogonalization. Frozen libraries pooled across search seeds attain forward RankICs of 0.0473, 0.0503, and 0.0589 on CSI 300, CSI 500, and CSI 1000, respectively, retaining 81%, 95%, and 103% of their corresponding development RankICs; every completed certified individual export is also positive out of sample. These results show that aligning LLM-driven exploration with library-level construction yields compact, auditable factor libraries that retain predictive value across windows and market universes.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Quantitative Stock Factor Mining, LLM Agents, Self-Evolving Agent, Adaptive Search, Reliable ML Evaluation, Financial Time Series

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_dXXLOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

AlphaGEAR is an LLM-based framework for constructing factor libraries for an equal-weight rank predictor. It combines library-relative residual feedback, a state-dependent program-evolution controller, statistical admission, development-window refinement, and a query-limited certification interface.

Critical strengths

The paper makes a useful distinction between standalone factor performance, incremental evidence, and improvement to the deployed combination. The frozen forward evaluation and fixed-candidate construction experiments are valuable. Appendix C.2 also provides paired temporal bootstrap analyses with multiplicity correction.

Critical weaknesses

1. The residual statistic does not establish incremental predictive information.

In Eq. (1), linear residualization ensures orthogonality to the library ranks, but the subsequent Spearman calculation reranks the residuals and can destroy this property. For example, let the library and return ranks be x = y = (1, ..., 24), and candidate ranks be z = (1, ..., 8, 21, ..., 24, 9, ..., 20). The residual–return Spearman correlation is approximately 0.44696, despite the library already perfectly predicting returns. The candidate–library correlation is 0.66609, below the 0.7 redundancy threshold. Thus, a positive residual statistic need not indicate incremental predictive information, although downstream marginal filtering may reject such candidates.

2. The ablations do not establish the central mechanisms’ added value.

In Table 4, the full method achieves a forward RankIC of 0.0430, compared with 0.0436 without residual-guided search, 0.0471 without residual-based admission, and 0.0451 with library-family BH. Removing the marginal filter or LOO pruning yields 0.0429 and 0.0438, respectively. With only three seeds, these results do not establish that the components are harmful, but they provide insufficient evidence of improved forward prediction. Benefits in compactness or reliability require matched-size comparisons or direct validation.

Questions for authors

How should the positive residual Spearman statistic in the counterexample above be interpreted as evidence beyond the existing library? Does an alternative partial-rank statistic change the experimental findings?

How does Full compare with direct greedy marginal-RankIC selection and backward elimination under identical initialization, orientation, and candidate budgets?

What measurable benefit does certification provide relative to no certification or equally query-limited continuous feedback? Can its false-promotion behavior be evaluated in a controlled setting?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 3 (1/3) — Residual statistic and ablations

R3-W1

The counterexample identifies a re-ranking artifact, and the corrected partial-rank statistic returns zero under our degeneracy convention.

We thank the reviewer for the counterexample. Re-ranking the projected candidate residual can reintroduce association with the library ranks. The submitted statistic gives 0.44696 in this example, which should not be interpreted as evidence beyond the library. We correct Eq. (1) by projecting both candidate and label ranks on the library ranks and correlating the two residuals directly, without re-ranking. The corrected statistic measures linear partial association in rank space. In the counterexample, the label residual is zero, so the statistic returns zero under our degeneracy convention [Sec. 3.2, Appendix C.4].

Adding the candidate also lowers the equal-weight RankIC from 1.0000 to 0.8982, so it fails the separate marginal-contribution check during refinement [Sec. 3.3, Appendix C.4]. R3-Q1 reports the empirical impact of the correction, and R3-Q3 reports the controlled false-promotion experiments.

R3-W2

On the same candidate streams, Full retains fewer factors than the variants without LOO pruning, without residual-based admission, or with library-family BH. Table 4 leaves the added value of residual-guided search and residual-based admission for forward prediction unresolved.

Removing either gives a higher observed forward mean (0.0436 and 0.0471, respectively, versus 0.0430 for the full construction policy). Removing the marginal filter changes the observed forward mean only slightly, from 0.0430 to 0.0429 (Table 4).

On the same recorded candidate streams, the full construction policy retains 6.3 members on average, compared with 7.7 without leave-one-out pruning, 8.0 without residual-based admission, and 8.3 with library-family BH. All three larger variants have higher observed forward means, by 0.0008 to 0.0041. The full construction policy produces smaller libraries and higher observed forward RankIC than simple screening on all four fixed candidate collections, with gains for two sources surviving Holm correction (Tables 3 and 10).

Greedy selection outperforms the full construction policy on our candidate collections (+0.0066 forward RankIC, Holm-adjusted p = 0.029), while the policy outperforms backward elimination on the R&D-Agent(Q) collections (+0.0076, Holm-adjusted p = 0.021). Appendix C.5 reports all matched-size comparisons. In the fixed-size replacement diagnostic, the historical RankIC improvement rate for replacements was below 50% in six of nine Full runs, short of the prespecified minimum of eight runs (Appendix C.8).

AuthorsOpenReview ↗

Response to Reviewer 3 (2/3) — Partial-rank reruns and matched-size selection

R3-Q1

The alternative partial-rank statistic preserves positive mean RankIC across all evaluation windows in complete closed-loop reruns.

Appendix C.4 reports three complementary checks.

  • Fixed-library diagnostics. Holding each candidate's reference library and decision threshold fixed, 278 of 303 admission verdicts agree. The remaining 25 comprise 22 pass-to-reject and 3 reject-to-pass changes.
  • Candidate-stream replay. Replaying the recorded candidate streams with unchanged thresholds, while allowing the reference library to evolve, preserves two of the three libraries exactly. The third library's development RankIC changes from 0.0547 to 0.0523. The union grows from 10 to 13 members, and its development RankIC changes from 0.0584 to 0.0565.
  • Complete closed-loop reruns. We reran the search with the corrected partial-rank statistic in online fitness, admission, refinement, and certification, using three independent CSI 300 runs and 100 iterations per run, against three matched runs with the original statistic. Development RankIC is 0.0507 ± 0.0012 versus 0.0494 ± 0.0029, and historical RankIC is 0.0303 ± 0.0031 versus 0.0309 ± 0.0074. In the post-submission supplementary evaluation, all six frozen libraries have positive forward RankIC (Newey–West t ≥ 5.0 for each run). Forward RankIC is 0.0469 ± 0.0009 with the corrected statistic and 0.0419 ± 0.0020 with the original statistic. Values are means ± sample standard deviations across three runs. The forward difference of +0.0050 is not statistically significant (95% interval [−0.0013, +0.0115]).

The main results retain the submitted residual-Spearman implementation. Table 13 reports the separate closed-loop evaluation of the corrected partial-rank implementation.

R3-Q2

With the same candidate pools and matched library sizes, direct greedy marginal-RankIC selection achieves higher forward RankIC than Full on AlphaGEAR’s collections, while Full scores higher than backward elimination on R&D-Agent(Q)’s collections.

We compared the methods on the same 12 frozen candidate collections, including the shared initial programs and the available candidates rejected by the full construction policy. All three methods retain native signal orientations. Both direct selectors optimize development RankIC of the same equal-weight combination and match Full’s final library size in each run. The full construction policy and greedy forward selection start from an empty library. Backward elimination starts from the entire pool and removes members.

On the forward window, greedy selection beats the full construction policy on our candidate collections by +0.0066 (pointwise 95% interval [0.0022, 0.0114], Holm-adjusted p = 0.029). Backward elimination is also higher on these collections by +0.0050, with an adjusted p = 0.103. The full construction policy beats backward elimination on the R&D-Agent(Q) collections by +0.0076 (pointwise 95% interval [0.0025, 0.0125], adjusted p = 0.021). On the historical window, the policy beats both selectors on the AlphaAgent collections by +0.0026 and +0.0030 (adjusted p = 0.028 and 0.005). After Holm correction, two of the eight forward contrasts and two of the eight historical contrasts reached statistical significance (Appendix C.5). The bootstrap intervals quantify temporal uncertainty conditional on the observed runs.

The number of development-window objective evaluations can differ across selectors. The direct selectors optimize the combination score without the policy’s statistical screen or redundancy cap, so the contrasts measure the effect of changing the construction policy as a whole. The full construction policy’s admission rule makes the statistical evidence and redundancy criteria for accepting a candidate explicit.

AuthorsOpenReview ↗

Response to Reviewer 3 (3/3) — Certification and false promotion

R3-Q3

Under matched search budgets, the forward RankIC benefit of certification remains uncertain, and controlled simulations show that the corrected statistic reduces false promotion.

Certification under matched search budgets. We compared certification with no certification using the original residual statistic (Appendix C.6). In the post-submission supplementary evaluation of frozen libraries, mean forward RankIC was 0.0419 with certification and 0.0451 without it. The enabled-minus-disabled difference was −0.0032 (95% interval [−0.0097, +0.0030]).

The interface caps queries, restricts each response to a finite alphabet, and logs the feedback returned to search (Appendix D). Whether discrete feedback reduces adaptive overfitting remains unresolved pending a comparison with equally query-limited continuous feedback.

Per-query false promotion under the two statistics. The synthetic experiment evaluates the submitted residual-Spearman statistic and the corrected partial-rank statistic on panels with known truth (360 days × 300 names, 2,000 replicates per setting), using the certification service’s margin and threshold for a PROMOTE verdict. The partial-rank statistic reduces false-promotion rates in all 36 library-dependent settings without incremental information, covering linear, quadratic, and rearrangement-based candidate families. In the primary noiseless-candidate setting of a synthetic family based on the reviewer’s rearrangement, the rate falls from 99.8% to 10.4% with one library member and from 95.3% to 9.2% with five members. Here the labels contain independent noise, distinguishing this simulation from the exact counterexample discussed above.

For candidates with genuine incremental information, the two statistics yield similar observed detection rates in 20 of 24 settings. The corrected statistic yields rates 2.45–4.20 percentage points lower in the remaining four weak-signal settings correlated with the library. Pure-noise promotion rates range from 1.9% to 3.0%, near the 2.31% finite-sample reference rate.

Calibration varies across settings. Across the 36 library-dependent settings, corrected rates range from 0% to 10.4%. In 29 of the 36 settings, the entire 95% Wilson interval lies above the finite-sample reference rate of 2.31% (P(t₃₅₉ ≥ 2)). The statistic measures linear partial association in rank space, so nonlinear dependence on the library can remain after projection. These simulations characterize per-query behavior under the specified synthetic models (Appendix C.7). The comparison with equally query-limited continuous feedback and the effect on adaptive overfitting across a search remain open.

The certification verdicts reported in the submission were produced with the original statistic and are reported as such.

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

This paper proposes AlphaGEAR, a self-evolving agent for constructing refined financial factor libraries. It focuses on the gap between the individual factor's predictive value and its contribution to the final predictor, utilizing the combination of residual-guided search, stream-aware admission, and marginal-contribution refinement under a limited query budget. The constructed and frozen libraries on experiments of CSI 300, CSI 500 and CSI 1000 are used to report the positive forward RankIC. Under the common equal-weight setup, AlphaGEAR outperform other competing agents while using fewer factors to improve mean forward RankIC.

Critical strengths
  1. This paper discusses the mentioned problem clearly. A single well-performing factor or a factor providing other useful information does not mean that their direct combination would perform well also. It then devises a framework to collect the factor's generation, filtering and combining together to link the final combination, which overall approach is coherent.

  2. Under the fixed candidate sets, the paper compares simple screening with full construction policy across four different frameworks. Also, it shows that the AlphaGEAR's ability to select factors from existing candidates and construct an equal-weight predictor. Based on the multiple random seeds' experiment results, it includes forward evaluation of frozen factor libraries and ablation studies. It then helps reader to understand the method's performance and limitation depending on individual components on the predictive performance improvement.

Critical weaknesses
  1. The methodological contribution should be clearly stated compared to the existing works. The paper emphasizes that effects of single factor does not exactly equal to the contribution to the final combination. However, existing works like AlphaGen and AlphaForge already consider this aspects. Therefore, it would be better to explain clearly what empirical insights AlphaGEAR offer beyond these approaches. Also, providing concrete examples of discovered factors or evolution progress would better what this framework actually explores.

  2. The reason why the performance gains could not be explained well by the current comparison. The main experiments shows that full construction policy outperforms simple screening; however, a stronger baseline that directly selects factors based on the performance of the final combination is missing. Therefore, it still remains unclear how much the full policy improves over direct forward selection and backward elimination.

Questions for authors
  1. The current evaluation mainly focuses on RankIC, providing a limited view of practical financial performance. Related work, like AlphaAgent and R&D-Agent (Q), also reports return and risk metrics and accounts for transaction costs. Indeed, the RankIC shows the ability of AlphaGEAR, but other metric evaluation could help reader to understand AlphaGEAR better from other perspectives.

  2. Current experiment only use deepseek-v4.1-flash to evaluate the AlphaGEAR. LLMs with stronger reasoning or code generation ability could affect candidate quality, the relative performance of different framework. It therefore remains unclear if the findings generalize beyond the tested model configuration.

  3. Table 4 reports mean forward RankIC without residual-guided search is close to the full method. The paper explains that some constraints help produce smaller factor libraries. Additional evidence would help clarify if the full method achieves a better trade-off between library size and predictive performance than simpler variants.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 2 (1/2) — Method, factor examples and selection baselines

R2-W1

Under the same construction policy, libraries built from AlphaGEAR’s searched candidates improve on the shared initialization by +0.0067 historical and +0.0114 forward RankIC.

AlphaGen already rewards its generator based on the performance of a fitted combination. It refits the weights by gradient descent after adding a candidate and drops the factor with the smallest absolute weight when the pool exceeds capacity. Its results also show the limits of mutual-IC screening. AlphaForge trains its generator on individual-factor IC, then re-selects factors and fits a linear combination each trading day [Sec. 2].

In AlphaGEAR, online admission updates the library that guides later proposals. Final refinement checks both a candidate's library-adjusted rank evidence and its marginal effect on the predictor's development RankIC. Screening includes recorded rejected trials. The library-level combiner averages member ranks with equal weights, and the certification interface separately caps and logs sealed queries.

On the same candidate pools, Full improves forward RankIC over simple screening with fewer factors. With candidates and the combiner fixed, the observed gains are 0.0088–0.0149 with 78–92% fewer members across all four sources, including other agents' exports (Table 3). Gains for our collections and QuantaAlpha's survive Holm correction (p = 0.029 and 0.036, Table 10). This comparison evaluates the complete construction policies, including direction handling.

The pointwise 95% intervals for both gains exclude zero (Appendix C.9). Under identical construction, our observed collections also score higher on forward RankIC than each comparator's (+0.0042 to +0.0116, Holm-adjusted p ≤ 0.042, Table 10). At matched sizes, forward selection scores higher than Full on our collections (R2-W2), so the advantage over simple screening does not establish a general advantage over direct selection.

Search-generated programs account for 17 of the 19 member entries across the three CSI 300 libraries in Table 1. Appendix C.12 shows the first search-generated admission in each run. One early admission combines the cross-sectional ranks of 26 price and volume features using ridge regression, refitted every 10 days on a rolling 120-day window using only outcomes available when the signal is formed. The other examples are a return-distribution signal and a market-adjusted reversal and volatility signal. In separate matched search runs, online development RankIC rises from 0.0358 at initialization to 0.0395, 0.0486, and 0.0509 on the selection window.

R2-W2

At matched library sizes, greedy forward selection based on the final combination’s development RankIC achieves higher forward RankIC than Full on AlphaGEAR’s candidate pools (0.0496 vs. 0.0430, a gain of +0.0066, Holm-adjusted p = 0.029).

On the R&D-Agent(Q) collections, Full achieves higher forward RankIC than backward elimination (0.0314 vs. 0.0238, Holm-adjusted p = 0.021). The other six forward contrasts do not reach statistical significance after Holm correction. On the historical window, both significant contrasts favor Full on the AlphaAgent collections. AlphaGEAR’s candidate pools achieve the highest observed forward RankIC under Full and both direct selectors, while Full and greedy forward selection both exceed simple screening across all four sources. Tables 14–16 report the complete results.

For each run on the Table 3 collections, both direct selectors use the same candidate pool and equal-weight aggregation rule as Full. They optimize development RankIC and use Full’s final library size as their stopping point. Their forward scores come from post-submission supplementary evaluations of frozen libraries.

These results support evaluating candidates through the deployed combination and leave the relative performance of Full and direct selection dependent on the candidate source in the observed comparisons. Full provides a rule for determining library size through statistical screening, redundancy control, and refinement.

AuthorsOpenReview ↗

Response to Reviewer 2 (2/2) — Portfolio metrics, LLMs and library size

R2-Q1

After accounting for transaction costs, all four methods underperform the CSI 300 on the forward window, with AlphaGEAR achieving the highest observed net excess return.

We evaluated all four methods using the same top-50 equal-weight long-only portfolio rule, with at most five replacements per day and assumed transaction costs of 7.5 bp on purchases and 17.5 bp on sales (Appendix C.11).

In the post-submission forward evaluation of frozen libraries, AlphaGEAR’s mean annualized net excess return relative to the CSI 300 was −7.8%, its mean information ratio was −0.84, and its mean maximum drawdown of excess returns was −42.1%. Its mean annualized net excess return was 2.2–3.4 percentage points higher than the comparator means, with all three 95% confidence intervals from the post hoc paired comparisons including zero. Estimated annualized trading costs ranged from 6.1% to 6.4% across methods.

On the development window used for selection, our method’s mean annualized net excess return was +8.8%, compared with −0.1% to +0.5% for the three comparators. Table 22 reports net excess returns, information ratios, and maximum drawdowns of excess returns for all four methods.

R2-Q2

AlphaGEAR achieves positive mean RankIC on all three evaluation windows with each of the three tested LLMs, spanning two model families.

We reran the full CSI 300 pipeline with two additional models, changing only the model identifier and using three independent runs per model. In the post-submission supplementary evaluation of the additional models’ frozen libraries, mean forward RankIC was 0.0470 for Claude Opus 5.5 and 0.0460 for DeepSeek V4 Pro, compared with 0.0430 for the DeepSeek V4.1 Flash reference cohort also used in Table 3. Both additional models had higher observed development means and lower historical means. Table 21 reports the full results, including sample standard deviations.

Table 1 compares all four frameworks using the same LLM. Whether their relative rankings persist with other LLMs remains open.

R2-Q3

Without residual-guided search, forward RankIC is close to Full’s and the resulting factor library is larger on average (Table 4).

Removing residual-guided search yields 7.0 members on average and a forward RankIC of 0.0436, compared with 6.3 members and 0.0430 for Full (Table 4). The fixed-stream variants without LOO pruning, with library-family BH, and without residual-based admission retain 7.7, 8.3, and 8.0 members, respectively, with higher observed forward means of 0.0438, 0.0451, and 0.0471. These are means across three runs per configuration.

Appendix C.8 (Table 19) reports eleven threshold and component configurations replayed on the same three recorded CSI 300 candidate streams. Compared with BH q = 0.20, removing LOO pruning, removing residual-based admission, and library-family BH, Full uses 18–24% fewer members, while observed historical RankIC differences remain below 0.001. Each configuration is evaluated at its natural final library size.

These results describe the observed compactness–performance trade-off and leave a forward predictive advantage for Full over the simpler variants unresolved. The matched-size comparisons with direct selection are reported in R2-W2 and Appendix C.5.

Thank you for the detailed response and the considerable effort put into the additional experiments. I really appreciate the candid reporting of all results. After careful consideration, the benefit of the proposed construction over direct selection remains unclear to me, so I will keep my original rating.

AuthorsOpenReview ↗

Follow-up to Reviewer 2

We sincerely thank the reviewer for engaging so carefully with our response and for raising this important question. In each run, both greedy forward selection and backward elimination stopped at the final library size determined by Full. The greedy rule tells us which factor to add next and needs a separate stopping rule to determine when to stop. Both direct selectors therefore inherit Full's decision about how many factors to retain. The comparison evaluates the subsets each method selects from the same candidate pool at that size. How either direct selector would perform with its own stopping rule remains an open question.

AlphaGEAR's candidate pools achieved the highest observed mean forward RankIC among the four sources under every construction policy in Table 14. The highest value overall was 0.0496 for greedy forward selection on AlphaGEAR's pools, compared with 0.0430 for Full. Both Full and greedy forward selection exceeded score-and-correlation screening in observed mean forward RankIC across all four sources, which supports using the deployed equal-weight predictor to guide selection.

The table reports paired RankIC differences as Δ = Full − selector. Positive values favor Full. Bold indicates Holm-adjusted p < 0.05, with correction applied separately to the eight contrasts in each evaluation window (Appendix C.5, Tables 15–16).

Candidate sourceForward vs. greedyForward vs. backwardHistorical vs. greedyHistorical vs. backward
AlphaGEAR−0.0066−0.0050−0.0030−0.0026
QuantaAlpha−0.0023+0.0034+0.0015+0.0022
R&D-Agent(Q)+0.0018+0.0076+0.0009+0.0034
AlphaAgent+0.0014+0.0013+0.0026+0.0030

Of the 16 contrasts, 11 favor Full in sign, and three of the four significant ones favor Full; the exception is greedy selection on AlphaGEAR's own pools. Between Full and greedy selection alone, one significant contrast favors each method. Full's contribution is therefore not a higher RankIC than direct selection at a fixed size, but comparable accuracy from a procedure that determines library size and membership itself, through explicit and auditable statistical screening, redundancy control, and leave-one-out pruning.

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

AlphaGEAR uses LLM agents to search for financial factors and assemble them into an equal-weight predictor. Search uses residual predictive information relative to the current factor library. Admission and refinement select factors based on statistical evidence, redundancy, and their contribution to the combined predictor. A query-limited service provides discrete feedback from a separate historical window during search. Experiments report stronger forward prediction with smaller libraries, including improvements when the same construction policy is applied to baseline candidates.

Critical strengths

The paper is fairly well written and readable, except that it has some random math in the middle that I am not sure adds much to the paper itself

I like the focus on the final predictor. A factor can add residual information without improving an equal-weight combination. Checking its actual marginal contribution addresses that mismatch directly. The resulting libraries are also much smaller, which makes them easier to inspect.

Libraries are frozen before forward evaluation, and the paper reports repeated searches, several stock universes, and paired uncertainty estimates. It also distinguishes retrospective results from forward results. These checks give the empirical contribution more weight.

Critical weaknesses

One concern is the novelty, as the agentic evolutionary search is just an application of existing work to this problem of finance forecasting. For example, QuantaAlpha (https://arxiv.org/abs/2602.07085) already studies evolutionary alpha mining, and AlphaForge (https://arxiv.org/abs/2406.18394) combines factor generation with selection and weighting.

Another thing is that there aren't really any ablations on the restricted certification at the end. A simple greedy selector at matched library sizes would help establish whether the statistical rules are needed.

Certification is audited, but its benefit is not tested. Query counts show that the interface respects its budget, not that discrete feedback reduces overfitting or improves selection. Positive results for certified subsets do not isolate this effect either. I would compare no certification and discrete feedback under matched search budgets.

Questions for authors

Weaknesses section

2 comments
AuthorsOpenReview ↗

Response to Reviewer 1 (1/2) — Method and matched-size selection

R1-W1

AlphaGEAR extends evolutionary alpha mining by using the evolving factor library to guide program search and refining that library for a fixed equal-weight rank predictor, with a separate budget for held-out feedback.

Evolutionary alpha mining and combination-aware selection are established. QuantaAlpha evolves mining trajectories through mutation and crossover and scores trajectories through the factors they produce. Its case study builds the pool greedily in RankIC order with a 0.7 correlation cap, and its downstream predictor uses LightGBM. AlphaForge trains its generator on individual-factor IC, screens factors using IC, ICIR and correlation, and re-selects and linearly combines factors each trading day. Its Discussion proposes using combined IC as a mining objective.

In AlphaGEAR, each online admission updates the library that guides subsequent search. During final refinement, we replay recorded candidates, recompute library-adjusted evidence, and check how each candidate changes the development RankIC of the equal-weight rank predictor used at deployment. That aggregation fits no library-level weights. The screening family includes recorded rejected trials, and the certification interface caps and logs sealed queries [Secs. 1–2].

On fixed candidate pools, Full improves forward RankIC over simple screening with fewer factors. Table 3 holds the candidate collections and equal-weight predictor fixed. Simple orients factors by development RankIC, ranks them by absolute RankIC, and applies a 0.7 correlation cap. Full gains 0.0088–0.0149 in observed forward RankIC with 78–92% fewer members across all four sources, including exports from QuantaAlpha, R&D-Agent(Q), and AlphaAgent. Gains for our collections and QuantaAlpha's survive Holm correction (p = 0.029 and 0.036, respectively). AlphaAgent's adjusted p is 0.056 (Table 10). The forward difference is positive in all 14 additional stored pairs, including CSI 500 and CSI 1000 [Appendix C.2]. These comparisons test the complete construction policies, including their direction handling.

The libraries constructed from our searched candidates improve on the shared initialization. Under Full, these libraries gain +0.0067 historical RankIC (95% interval [0.0029, 0.0104]) and +0.0114 forward RankIC ([0.0050, 0.0179]) over the initialization-only reference [Appendix C.9]. With the same construction policy, our observed collections also score higher on forward RankIC than each comparator's (+0.0042 to +0.0116, Holm-adjusted p ≤ 0.042, Table 10).

Greedy selection scores higher than Full on our collections at matched sizes (R1-W2). Table 4 leaves the forward benefit of residual-guided search unresolved. The evidence supports the framework under the evaluated equal-weight setting and leaves the necessity of individual constraints open.

R1-W2

A simple greedy selector at matched library sizes achieves higher observed forward RankIC than simple screening across all four candidate sources without applying the statistical admission rules.

On our fixed candidate collections from Table 3, greedy selection also scores higher than the full construction policy, at 0.0496 versus 0.0430 (+0.0066, Holm-adjusted p = 0.029). Appendix C.5 reports all four comparisons.

The full construction policy provides a rule for deciding which candidates to admit and how many to retain. It screens each candidate against the current library, includes rejected trials in the screening family, and enforces a redundancy cap. Final refinement also checks whether a candidate improves the equal-weight predictor. The matched-size results do not establish that the statistical admission rules are needed for higher forward RankIC. The greedy baseline uses Full’s final library size as its stopping point and selects directly on the combination’s development RankIC. The matched-budget certification ablation is reported in our response to R1-W3 and Appendix C.6.

AuthorsOpenReview ↗

Response to Reviewer 1 (2/2) — Certification and presentation

R1-W3

We compared no certification and discrete certification feedback under matched search budgets.

On CSI 300, we used the original residual statistic, a 100-iteration budget, and three independent runs per arm. In the post-submission supplementary evaluation of frozen libraries, mean forward RankIC was 0.0451 without certification and 0.0419 with discrete certification feedback. The enabled-minus-disabled difference was −0.0032, with a 95% moving-block bootstrap interval of [−0.0097, +0.0030]. Appendix C.6 also reports the development and historical scores.

The forward RankIC benefit of enabling certification remains uncertain. Whether discretizing feedback reduces adaptive overfitting remains open pending a comparison with numerical feedback under the same query budget. The interface caps queries and logs the discrete response together with its candidate and reference library.

R1-Other

We streamlined Sections 3.3–3.4 around the algorithmic decisions and moved the BH threshold and feedback bound to the appendices. The equations defining the deployed predictor and marginal contribution remain in the main text, while the BH screening threshold and the feedback-information bound are presented in Appendices A and D. The main text explains their roles directly and points to the detailed definitions and derivation.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to Reviewers

We thank the reviewers for their careful reading and for raising specific questions about our method and evaluation.

AlphaGEAR couples library-conditioned program search, construction of a fixed equal-weight predictor, and explicitly budgeted held-out feedback. The new comparisons support the predictive value of the searched candidate collections and clarify the limits of the construction and certification claims.

  1. AlphaGEAR’s candidate collections achieve the highest observed forward RankIC among the four sources under Full, forward selection, and backward elimination. Under the same Full construction policy, the searched collections improve on the shared initialization by +0.0067 historical and +0.0114 forward RankIC, with both 95% intervals excluding zero. These comparisons assess the observed candidate collections [Appendices C.5, C.9].

  2. Full improves on simple screening, while its advantage over direct selection depends on the candidate source. With candidates and the combiner held fixed, Full improves observed forward RankIC by 0.0088–0.0149 with 78–92% fewer members across all four sources; gains for our collections and QuantaAlpha’s survive Holm correction. Including stored-run comparisons, the forward difference is positive in all 26 completed pairs [Appendix C.2]. At matched library sizes, forward selection exceeds Full on our collections by +0.0066 (Holm-adjusted p = 0.029). The new baselines do not establish a general predictive advantage for Full over direct combination-based selection [Appendix C.5].

  3. Correcting the re-ranking artifact preserves positive mean RankIC across all three windows in complete closed-loop reruns. We project both candidate and label ranks on library ranks and correlate the residuals directly. The statistic measures linear partial association in rank space. Tables 1–5 and Figure 3 retain the submitted residual-Spearman results; Table 13 separately reports the corrected implementation. The correction also reduces per-query false promotion in all 36 controlled library-dependent null settings, although calibration remains imperfect [Appendices C.4, C.7].

  4. Mean RankIC remains positive across three LLMs spanning two model families. We also provide concrete discovered-program examples, size–performance comparisons, and portfolio return, risk, and cost evaluations [Appendices C.8, C.10–C.12].

  5. The matched-budget comparison does not establish a forward RankIC benefit from certification. The enabled-minus-disabled difference is −0.0032 (95% interval [−0.0097, +0.0030]). Whether discrete feedback reduces adaptive overfitting remains untested against equally query-limited numerical feedback. The interface caps queries and records the discrete feedback returned to search [Appendix C.6; Sec. 5].

The bootstrap intervals quantify temporal uncertainty conditional on the observed runs; they do not quantify uncertainty from new search runs.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.