Pith. sign in

REVIEW 3 major objections 6 minor 103 references

This review argues that all automated alpha-discovery methods — GP, RL, GFlowNets, MCTS, LLM agents — are one noisy evolutionary loop, and that its weakest link is the fitness signal.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

A review that reframes automated trading-signal (alpha) discovery as noisy, dynamic, multiobjective evolutionary optimization, and proposes six-component and eight-dimension frameworks for comparing and evaluating methods.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Worth reading for its evaluation matrix and coherent EC framing, but the main 'fitness is the bottleneck' diagnosis is partly built into Table IV's coding rule and needs a transparent rubric before it carries weight. the 3 major comments →

arxiv 2608.01789 v1 pith:ASQ62A7G submitted 2026-08-03 cs.NE

Towards Autonomous Formulaic Alpha Discovery: An Evolutionary Computation Perspective

Xinwei Yu , Yiyang Fu , Mingcheng Fan , Enqi Li , Yilin Gao , Shugong Xu This is my paper
classification cs.NE MSC 68W5090C59
keywords automated alpha discoveryformulaic alpha factorsevolutionary computationgenetic programmingnoisy fitness optimizationnonstationary marketsLLM agentsquantitative investment
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that automated formulaic alpha discovery — the search for interpretable trading-signal formulas from large symbolic factor spaces — is not a sequence of isolated methodological waves but a single noisy, dynamic, multiobjective symbolic evolutionary optimization problem. It organizes every method family (genetic programming, reinforcement learning, GFlowNets, Monte Carlo tree search, LLM-guided generation, agentic workflows) into one closed loop with six components: representation, variation, fitness evaluation, selection, memory, and adaptation. The payoff of the reframing is a diagnosis: representation and variation have advanced much faster than fitness reliability, validated memory, and regime-aware adaptation, so the noisy empirical fitness signal is the main bottleneck. If the diagnosis is right, more autonomy without stronger validation does not improve discovery — it just fits historical noise faster, which is why the paper's proposed remedy is an eight-dimension, reliability-centered evaluation protocol rather than a more powerful generator.

Core claim

This paper's central claim is that automated formulaic alpha discovery is not a sequence of isolated methodological waves — genetic programming, reinforcement learning, GFlowNets, MCTS, LLM prompting, and agentic workflows — but one noisy, dynamic, multiobjective symbolic evolutionary optimization problem. All method families are instantiations of a single closed loop, C_{t+1} = Sel(C_t ∪ V(C_t, M_t), F̂_t), whose state has six components: representation, variation, fitness evaluation, selection, memory, and adaptation. Reading the literature through this loop yields the paper's diagnosis: representation and variation have advanced faster than fitness reliability, validated memory, and regim

What carries the argument

The load-bearing object is the evolutionary loop equation (Eq. 10), C_{t+1} = Sel(C_t ∪ V(C_t, M_t), F̂_t), together with the six-component state S_t = (R_t, V_t, F_t, Sel_t, M_t, A_t) (Eq. 16). It recasts every alpha-discovery method — GP trees, RL policies, GFlowNet samplers, MCTS trees, LLM proposals, agent workflows — as a different implementation of the same cycle: representation defines the symbolic search space, variation generates candidates, fitness evaluation estimates noisy empirical quality, selection updates the retained factor pool, memory stores validated experience, and adaptation responds to market regime change. A second carrier is the qualitative component-coverage coding

Load-bearing premise

The diagnosis that representation and variation outpace fitness reliability, memory, and adaptation is read directly off the paper's own qualitative strong/moderate/weak ratings of each method family in Table IV; if an independent rater judged, say, one of the flagship GP or RL systems' validation design as strong, the bottleneck conclusion weakens.

What would settle it

Re-code the component coverage of the cited method families with independent raters using a fixed rubric — does the method apply purged cross-validation, multiple-testing correction, walk-forward validation, and memory validation before reuse? If most families score strong on fitness reliability or adaptation, the asymmetry diagnosis and its conclusion that fitness evaluation is the main bottleneck fail. A complementary experiment: hold the generator fixed and swap the fitness estimator between raw IC and a deflated, purged version; if the selected alphas barely change, fitness reliability is

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the framing is right, method families should be compared by which loop components they automate, not by their algorithmic labels: GP, RL, GFlowNets, MCTS, and LLM agents are not rival paradigms.
  • Fitness evaluation is the binding constraint: under a noisy reward signal, stronger generators, larger memories, and more complex agent loops amplify historical noise rather than improve discovery.
  • Evaluation of alpha-discovery systems must report process-level evidence — search budget, multiple-testing correction, leakage controls, memory validation, regime splits — not just the best reported IC or long-short return.
  • Memory must be validated before reuse: an archive that stores overfitted factors or lucky trajectories biases all later search toward the same failures.
  • Nonstationarity requires active adaptation — regime-conditioned fitness, decay-aware memory, adaptive resource allocation — because passive rolling windows cannot track a moving fitness landscape.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The framework suggests a direct controlled test of the bottleneck claim: run the same generator twice, differing only in the fitness estimator (raw IC versus purged, multiple-testing-corrected fitness); if the selected alphas barely change, fitness reliability is not the binding constraint for that method.
  • The component-coverage ratings in Table IV could be turned into a quantitative instrument: independent raters applying a fixed rubric across the cited families would show whether the asymmetry diagnosis is stable or an artifact of coding choices.
  • The same unbalanced loop — generation outpacing validation — plausibly applies to LLM-based scientific discovery beyond finance, where autonomous systems propose hypotheses faster than they can be tested; the paper gestures at this but does not develop it.
  • A practical consequence the authors leave implicit: reported IC differences between published alpha-mining papers are largely uninterpretable until the evaluation protocol, not the generator, is held fixed, since noisy fitness and undisclosed search budgets can dominate method quality.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. This review-style paper proposes that automated formulaic alpha discovery should be understood as a noisy, dynamic, multiobjective symbolic evolutionary optimization problem, rather than as a sequence of isolated methodological waves. It introduces a six-component analytical framework—representation, variation, fitness evaluation, selection, memory, adaptation—formalized as the closed loop in Eq. (10) and the system state in Eq. (16), and it surveys method families ranging from human-guided formula libraries to GP/EA, RL, GFlowNets, MCTS, LLM-guided generation, and agentic workflows. The paper's central diagnostic claim is that representation and variation have advanced faster than fitness reliability, validated memory, and active adaptation, with fitness evaluation identified as the main bottleneck (Sections III.D and IV.L). It also proposes an eight-dimensional, autonomy-oriented evaluation protocol (Table V) and a set of research priorities for reliable, reproducible alpha discovery. The quantitative content (Eqs. 4–9, 11–15) restates standard definitions correctly, and the paper makes no strong empirical claims of its own beyond the qualitative comparative synthesis.

Significance. If the central asymmetry diagnosis is accepted, the paper provides a useful unifying vocabulary for a fragmented literature and makes a constructive practical point: increasing search sophistication without reliable fitness and validated memory may amplify backtest overfitting. The formalization of the discovery loop in Eqs. (10) and (16) is a coherent and pedagogically valuable framing. The evaluation matrix in Table V is a genuinely useful contribution, as are the repeated calls for reproducible search-process reporting, multiple-testing correction, and memory validation. The paper is honest about the limits of current evidence and does not overclaim numerical superiority of any method. However, the central claim about component-level imbalance rests on a small number of qualitative table entries rather than on independent measurement; the paper is best read as a perspective and research agenda, not as an empirically established diagnosis. Its value will depend on whether the field adopts the proposed evaluation dimensions.

major comments (3)
  1. [Table IV / Section IV.L] The paper's main conclusion—that representation and variation have advanced faster than fitness reliability, validated memory, and active adaptation—is read directly from Table IV. The table's note defines F as 'the explicit treatment of fitness reliability and validation design, rather than the mere use of a reward, IC, RankIC, or backtesting score.' Under this definition, nearly every family that relies on ordinary backtest metrics is coded ◦ in F, so the weak-F result is substantially built into the coding rule rather than independently derived from the surveyed literature. No coding rubric, inter-rater protocol, or raw per-method ratings are provided. This is load-bearing: if an independent rater counted AlphaQCM's distributional uncertainty modeling or RiskMiner's risk-seeking reward as explicit treatment of fitness risk, the F column would change and the 'main bottleneck' claim wou
  2. [Table IV and Sections IV.D–IV.K] Each Table IV row assigns a single symbol per method family, which obscures within-family variance. For example, the RL row groups AlphaGen, Alpha2, and AlphaQCM, yet AlphaQCM models return uncertainty with distributional RL; the MCTS row groups RiskMiner and alphaCFG, yet RiskMiner explicitly searches trajectories with high reward variance. The table marks all these F entries as weak (◦). Because the paper's framework is component-level, a single symbol per family cannot support statements about what 'RL-based methods' or 'MCTS-based methods' do in general. A range of ratings, per-method coding, or a supplementary table with the underlying evidence would make the taxonomy falsifiable and would substantially strengthen the reproducibility of the central diagnosis.
  3. [Section III.D / Section V.A] The claim that 'fitness evaluation is the main bottleneck' is asserted in Section III.D and reiterated in V.A, but it is not derived from the formalization. Equations (10)–(16) are definitions of a loop and a state; they do not imply an ordering of bottleneck severity among the six components. The sentence 'This explains why fitness evaluation is the main bottleneck' (Section III.D) overstates the logical connection. As a review, the manuscript may legitimately advance this as a synthesis, but it should be framed as an interpretive judgment, with the basis (largely Table IV) made explicit and its subjectivity acknowledged. Otherwise the reader cannot distinguish a conclusion from a framing commitment.
minor comments (6)
  1. [Section II heading] Typo: 'Formulaicalpha' should be 'Formulaic Alpha' in the section heading.
  2. [Abstract / Section I] The phrase 'for the first time' appears in the abstract and contributions. Since no prior EC-unification review is cited or discussed, this claim is hard to verify and unnecessary; consider softening to 'this article presents a unified'.
  3. [Section V.D] AlphaLogics [95] is mentioned only as a citation without the method name in the text; the sentence reads as though the mechanism, not a named system, is being introduced. Add the system name for consistency with other families.
  4. [Table IV / Section III] The table note uses '⊙' and '◦' which may render inconsistently in some viewers; consider replacing with letters (S/M/W) or adding a legend in the table caption.
  5. [Section II.E / Section III.A] The term 'selection' is used both for the tree-search policy in MCTS and for the factor-pool update in the six-component loop. A brief terminological distinction would prevent confusion in the taxonomy sections.
  6. [Section IV.A] The inclusion criteria are described, but the actual number of screened/included studies is not reported. Since this is positioned as a systematic taxonomy, a PRISMA-style count or a simple summary of the screening outcome would improve transparency.

Circularity Check

1 steps flagged

The unifying EC framing is self-contained; the main bottleneck diagnosis is partly read off a coding rule that defines fitness evaluation as explicit reliability/validation treatment, making the asymmetry conclusion partially tautological.

specific steps
  1. self definitional [Section III.D; Section IV.L / Table IV note; Section VI]
    "†The F column evaluates the explicit treatment of fitness reliability and validation design, rather than the mere use of a reward, IC, RankIC, or backtesting score. ... This explains why fitness evaluation is the main bottleneck of automated formulaic alpha discovery. ... The main pattern in Table IV is structural asymmetry. ... representation and variation have advanced faster than fitness reliability, validated memory, and active adaptation."

    The central asymmetry diagnosis is read directly off Table IV. Because F is defined as explicit treatment of fitness reliability and validation design, any family that merely reports IC/RankIC/backtesting scores is coded weak by construction. The paper then uses these weak F entries to conclude both that fitness reliability is underdeveloped and that fitness evaluation is the main bottleneck. Thus the headline diagnosis is partly a restatement of the coding rule rather than an independently derived empirical finding, especially since no inter-rater rubric or raw ratings are provided.

full rationale

The paper is a position/review article, not an empirical system, so there is no fitted-to-prediction loop and no self-citation chain; the authors do not lean on their own prior results. The EC reframing (Eqs. 10 and 16) is a definitional modeling choice, not a derived prediction, and the evaluation roadmap in Table V is a normative checklist. The only notable circularity-adjacent concern is Table IV: the F-column coding rule makes the 'weak fitness reliability' finding a near-tautology, and the same issue applies in milder form to M and A. Because this coding is load-bearing for the paper's main bottleneck diagnosis, a modest circularity score is warranted, but the core unified-framework contribution retains independent conceptual content and the score should stay well below the levels reserved for fitted predictions or self-citation chains.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

The paper introduces no fitted constants: its equations (4)-(9) restate standard financial metrics, and Eq (13)'s thresholds tau and delta are uninstantiated placeholders. The load-bearing assumptions are domain premises: noisy fitness (Eq 11), nonstationary landscape (Eq 12), and the multiobjective characterization (Eq 14). These are reasonable and standard in the cited literature, but asserted rather than demonstrated. The one ad hoc construction is the six-component equivalence claim (Section III.A, Eq 16), which is the paper's own lens: all subsequent conclusions, including the asymmetry diagnosis, are stated in that vocabulary. No invented entities.

free parameters (2)
  • quality threshold tau (Eq 13)
    Symbolic threshold in the quality-diversity archive definition. Never instantiated or fitted; no derivation depends on its value.
  • diversity threshold delta (Eq 13)
    Symbolic diversity threshold in Eq (13). Also uninstantiated and does not affect the central claim.
axioms (5)
  • domain assumption Empirical fitness is a noisy estimate of latent utility: F_hat_t(alpha) = F*_t(alpha) + eps_t(alpha) (Eq 11).
    Section II.D.1. The bottleneck diagnosis assumes selection pressure is driven by noisy fitness. Reasonable for empirical finance but asserted, not derived.
  • domain assumption The fitness landscape is time-varying, so the optimal formula depends on the current regime: alpha*_t = argmax F_t(alpha) (Eq 12).
    Section II.D.2. Nonstationarity premise inherited from dynamic EC literature (refs [26], [27]); it is the basis for the adaptation component.
  • ad hoc to paper All surveyed method families (GP, RL, GFlowNets, MCTS, LLM agents) are implementations of a common six-component evolutionary loop (Eq 16).
    Section III.A. This equivalence is the paper's own construction; the taxonomy and asymmetry conclusions inherit it. A reader who rejects the equivalence loses the unifying claim.
  • domain assumption Useful alpha discovery is multiobjective: predictive quality, stability, diversity, tradability, simplicity, reproducibility (Eq 14).
    Section II.D.4. Standard practitioner characterization, asserted without empirical derivation.
  • standard math Standard statistical definitions: Pearson/Spearman correlation, ICIR, turnover, net return (Eqs 4-9).
    Section II.C. Correct restatements of conventional definitions.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Autonomous Formulaic Alpha Discovery: An Evolutionary Computation Perspective." pith.science (2026). https://pith.science/paper/ASQ62A7G

@misc{pith2026260801789,
  author       = {Pith},
  title        = {Pith review of: Towards Autonomous Formulaic Alpha Discovery: An Evolutionary Computation Perspective},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ASQ62A7G}},
  note         = {Machine review of arXiv:2608.01789}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Automated formulaic alpha discovery aims to generate predictive and interpretable trading signals from large symbolic factor spaces. Its effectiveness is constrained by noisy fitness estimates, market nonstationarity, costly backtesting, semantic redundancy, and conflicting practical objectives. Existing studies employ diverse techniques, including genetic programming (GP), evolutionary algorithms (EAs), reinforcement learning (RL), generative flow networks (GFlowNets), Monte Carlo tree search (MCTS), large language models (LLMs), and agentic workflows, but generally examine them as separate algorithmic families. This article introduces, for the first time, a unified evolutionary computation (EC) perspective on automated formulaic alpha discovery, formulating it as a noisy, dynamic, and multiobjective symbolic evolutionary optimization problem. A six-component analytical framework is developed to characterize existing methods through representation, variation, fitness evaluation, selection, memory, and adaptation. Furthermore, an eight-dimensional, autonomy-oriented evaluation framework is proposed, covering search efficiency, fitness reliability, residual alpha quality, economic diversity, tradability, evolutionary autonomy, robustness to nonstationarity, and reproducibility. Together, these frameworks provide a systematic foundation for unifying heterogeneous approaches, diagnosing component-level limitations, and guiding the development of reliable, adaptive, interpretable, and reproducible autonomous alpha discovery systems.

Figures

Figures reproduced from arXiv: 2608.01789 by Enqi Li, Mingcheng Fan, Shugong Xu, Xinwei Yu, Yilin Gao, Yiyang Fu.

Figure 1
Figure 1. Figure 1: A general framework for automated formulaic alpha discovery from [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Six-component framework for autonomous alpha evolution. The framework organizes automated formulaic alpha discovery as a closed-loop evolutionary [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

103 extracted references · 57 canonical work pages · 1 internal anchor

  1. [1]

    101 formulaic alphas,

    Z. Kakushadze, “101 formulaic alphas,”Wilmott, vol. 2016, no. 84, pp. 72–81, 2016, doi: 10.1002/wilm.10525

    work page doi:10.1002/wilm.10525 2016

  2. [2]

    R. C. Grinold and R. N. Kahn,Active Portfolio Management: A Quantitative Approach for Producing Superior Returns and Controlling Risk, 2nd ed. New York, NY , USA: McGraw-Hill, 2000

    2000

  3. [3]

    Simple technical trading rules and the stochastic properties of stock returns,

    W. Brock, J. Lakonishok, and B. LeBaron, “Simple technical trading rules and the stochastic properties of stock returns,”J. Finance, vol. 47, no. 5, pp. 1731–1764, 1992

    1992

  4. [4]

    Foundations of technical analysis: Computational algorithms, statistical inference, and empirical implementation,

    A. W. Lo, H. Mamaysky, and J. Wang, “Foundations of technical analysis: Computational algorithms, statistical inference, and empirical implementation,”J. Finance, vol. 55, no. 4, pp. 1705–1765, 2000

    2000

  5. [5]

    Common risk factors in the returns on stocks and bonds,

    E. F. Fama and K. R. French, “Common risk factors in the returns on stocks and bonds,”J. Financ. Econ., vol. 33, no. 1, pp. 3–56, 1993

    1993

  6. [6]

    Empirical asset pricing via machine learning,

    S. Gu, B. Kelly, and D. Xiu, “Empirical asset pricing via machine learning,”Rev. Financ. Stud., vol. 33, no. 5, pp. 2223–2273, 2020, doi: 10.1093/rfs/hhaa009

    doi:10.1093/rfs/hhaa009 2020

  7. [7]

    Returns to buying winners and selling losers: Implications for stock market efficiency,

    N. Jegadeesh and S. Titman, “Returns to buying winners and selling losers: Implications for stock market efficiency,”J. Finance, vol. 48, no. 1, pp. 65–91, 1993

    1993

  8. [8]

    On persistence in mutual fund performance,

    M. M. Carhart, “On persistence in mutual fund performance,”J. Finance, vol. 52, no. 1, pp. 57–82, 1997

    1997

  9. [9]

    Value and momentum everywhere,

    C. S. Asness, T. J. Moskowitz, and L. H. Pedersen, “Value and momentum everywhere,”J. Finance, vol. 68, no. 3, pp. 929–985, 2013

    2013

  10. [10]

    AutoAlpha: An efficient hierar- chical evolutionary algorithm for mining alpha factors in quantitative investment,

    T. Zhang, Y . Li, Y . Jin, and J. Li, “AutoAlpha: An efficient hierar- chical evolutionary algorithm for mining alpha factors in quantitative investment,”arXiv preprint arXiv:2002.08245, 2020

    Pith/arXiv arXiv 2002

  11. [11]

    AlphaEvolve: A learning framework to discover novel alphas in quantitative investment,

    C. Cui, W. Wang, M. Zhang, G. Chen, Z. Luo, and B. C. Ooi, “AlphaEvolve: A learning framework to discover novel alphas in quantitative investment,” inProc. 2021 Int. Conf. Management of Data (SIGMOD), 2021, pp. 2208–2216, doi: 10.1145/3448016.3457324

    work page arXiv 2021

  12. [12]

    Generating synergistic formulaic alpha collections via reinforcement learning,

    S. Yu, H. Xue, X. Ao, F. Pan, J. He, D. Tu, and Q. He, “Generating synergistic formulaic alpha collections via reinforcement learning,” inProc. 29th ACM SIGKDD Conf. Knowledge Discovery and Data Mining (KDD), 2023, pp. 5476–5486, doi: 10.1145/3580305.3599831

    arXiv 2023

  13. [13]

    RiskMiner: Discovering formulaic alphas via risk seeking Monte Carlo tree search,

    T. Ren, R. Zhou, J. Jiang, J. Liang, Q. Wang, and Y . Peng, “RiskMiner: Discovering formulaic alphas via risk seeking Monte Carlo tree search,” inProc. 5th ACM Int. Conf. AI in Finance (ICAIF), 2024, pp. 752–760, doi: 10.1145/3677052.3698613

    arXiv 2024

  14. [14]

    AlphaSAGE: Structure-aware alpha mining via GFlowNets for robust exploration,

    B. Chen, H. Ding, N. Shen, J. Huang, T. Guo, L. Liu, and M. Zhang, “AlphaSAGE: Structure-aware alpha mining via GFlowNets for robust exploration,”arXiv preprint arXiv:2509.25055, 2025

    Pith/arXiv arXiv 2025

  15. [15]

    Navigating the Alpha Jungle: An LLM- powered MCTS framework for formulaic alpha factor mining,

    Y . Shi, Y . Duan, and J. Li, “Navigating the Alpha Jungle: An LLM- powered MCTS framework for formulaic alpha factor mining,”Proc. AAAI Conf. Artif. Intell., vol. 40, no. 2, pp. 997–1005, 2026, doi: 10.1609/aaai.v40i2.37069

    work page doi:10.1609/aaai.v40i2.37069 2026

  16. [16]

    AlphaAgent: LLM-driven alpha mining with regularized exploration to counteract alpha decay,

    Z. Tang, Z. Chen, J. Yang, J. Mai, Y . Zheng, K. Wang, J. Chen, and L. Lin, “AlphaAgent: LLM-driven alpha mining with regularized exploration to counteract alpha decay,” inProc. 31st ACM SIGKDD Conf. Knowledge Discovery and Data Mining (KDD), 2025, pp. 2813– 2822, doi: 10.1145/3711896.3736838

    arXiv 2025

  17. [17]

    FactorMiner: A self-evolving agent with skills and experience memory for financial alpha discovery,

    Y . Wang, J. Xu, H. Zhang, S.-L. Huang, D. D. Sun, and X.-P. Zhang, “FactorMiner: A self-evolving agent with skills and experience memory for financial alpha discovery,”arXiv preprint arXiv:2602.14670, 2026

    arXiv 2026

  18. [18]

    Distilling free-form natural laws from experimental data,

    M. Schmidt and H. Lipson, “Distilling free-form natural laws from experimental data,”Science, vol. 324, no. 5923, pp. 81–85, 2009

    2009

  19. [19]

    Order of nonlin- earity as a complexity measure for models generated by symbolic re- gression via Pareto genetic programming,

    E. J. Vladislavleva, G. F. Smits, and D. den Hertog, “Order of nonlin- earity as a complexity measure for models generated by symbolic re- gression via Pareto genetic programming,”IEEE Trans. Evol. Comput., vol. 13, no. 2, pp. 333–349, 2009, doi: 10.1109/TEVC.2008.926486

    work page arXiv 2009

  20. [20]

    AI Feynman: A physics-inspired method for symbolic regression,

    S.-M. Udrescu and M. Tegmark, “AI Feynman: A physics-inspired method for symbolic regression,”Sci. Adv., vol. 6, no. 16, Art. no. eaay2631, 2020, doi: 10.1126/sciadv.aay2631

    doi:10.1126/sciadv.aay2631 2020

  21. [21]

    Contemporary symbolic regression methods and their relative performance,

    W. La Cava et al., “Contemporary symbolic regression methods and their relative performance,” inNeurIPS Datasets and Benchmarks, 2021

    2021

  22. [22]

    J. H. Holland,Adaptation in Natural and Artificial Systems. Ann Arbor, MI: University of Michigan Press, 1975

    1975

  23. [23]

    D. E. Goldberg,Genetic Algorithms in Search, Optimization and Machine Learning. Reading, MA: Addison-Wesley, 1989. 17

    1989

  24. [24]

    Predicting excess stock returns out of sample: Can anything beat the historical average?

    J. Y . Campbell and S. B. Thompson, “Predicting excess stock returns out of sample: Can anything beat the historical average?”Rev. Financ. Stud., vol. 21, no. 4, pp. 1509–1531, 2008

    2008

  25. [25]

    A comprehensive look at the empirical performance of equity premium prediction,

    A. Goyal and I. Welch, “A comprehensive look at the empirical performance of equity premium prediction,”Rev. Financ. Stud., vol. 21, no. 4, pp. 1455–1508, 2008

    2008

  26. [26]

    Tracking extrema in dynamic environments,

    P. J. Angeline, “Tracking extrema in dynamic environments,” in Evolutionary Programming VI, ser. Lecture Notes in Computer Sci- ence, vol. 1213. Berlin, Germany: Springer, 1997, pp. 335–345, doi: 10.1007/BFb0014823

    work page doi:10.1007/bfb0014823 1997

  27. [27]

    R. W. Morrison,Designing Evolutionary Algorithms for Dynamic Environments. Berlin: Springer, 2004

    2004

  28. [28]

    Does academic research destroy stock return predictability?

    R. D. McLean and J. Pontiff, “Does academic research destroy stock return predictability?”J. Finance, vol. 71, no. 1, pp. 5–32, 2016

    2016

  29. [29]

    C. A. Coello Coello, G. B. Lamont, and D. A. Van Veldhuizen, Evolutionary Algorithms for Solving Multi-Objective Problems, 2nd ed. New York, NY , USA: Springer, 2007

    2007

  30. [30]

    Multiobjective evolutionary algorithms: A comparative case study and the strength Pareto approach,

    E. Zitzler and L. Thiele, “Multiobjective evolutionary algorithms: A comparative case study and the strength Pareto approach,”IEEE Trans. Evol. Comput., vol. 3, no. 4, pp. 257–271, 1999

    1999

  31. [31]

    Approximating the nondominated front using the Pareto archived evolution strategy,

    J. D. Knowles and D. W. Corne, “Approximating the nondominated front using the Pareto archived evolution strategy,”Evol. Comput., vol. 8, no. 2, pp. 149–172, 2000

    2000

  32. [32]

    Surrogate-assisted evolutionary computation: Recent advances and future challenges,

    Y . Jin, “Surrogate-assisted evolutionary computation: Recent advances and future challenges,”Swarm Evol. Comput., vol. 1, no. 2, pp. 61–70, 2011

    2011

  33. [33]

    A comparative analysis of selection schemes used in genetic algorithms,

    D. E. Goldberg and K. Deb, “A comparative analysis of selection schemes used in genetic algorithms,” inFoundations of Genetic Algo- rithms, G. J. E. Rawlins, Ed. San Mateo, CA, USA: Morgan Kaufmann, 1991, pp. 69–93

    1991

  34. [34]

    Adaptive operator selection and management in evolution- ary algorithms,

    J. Maturana, Á. Fialho, F. Saubion, M. Schoenauer, F. Lardeux, and M. Sebag, “Adaptive operator selection and management in evolution- ary algorithms,” inAutonomous Search, New York: Springer, 2012, pp. 161–189

    2012

  35. [35]

    The PRISMA 2020 statement: An updated guideline for reporting systematic reviews,

    M. J. Page et al., “The PRISMA 2020 statement: An updated guideline for reporting systematic reviews,”BMJ, vol. 372, p. n71, 2021

    2020

  36. [36]

    Kitchenham and S

    B. Kitchenham and S. Charters,Guidelines for Performing Systematic Literature Reviews in Software Engineering, EBSE Technical Report EBSE-2007-01, 2007

    2007

  37. [37]

    Improving reproducibility in machine learning re- search: A report from the NeurIPS 2019 reproducibility program,

    J. Pineau et al., “Improving reproducibility in machine learning re- search: A report from the NeurIPS 2019 reproducibility program,”J. Mach. Learn. Res., vol. 22, no. 164, pp. 1–20, 2021

    work page 2019

  38. [38]

    Best practices for scientific computing,

    G. Wilson et al., “Best practices for scientific computing,”PLoS Biol., vol. 12, no. 1, p. e1001745, 2014

    work page 2014

  39. [39]

    Reproducible research in computational science,

    R. D. Peng, “Reproducible research in computational science,”Science, vol. 334, no. 6060, pp. 1226–1227, 2011

    work page 2011

  40. [40]

    Qlib: An AI-oriented quantitative investment platform,

    X. Yang, W. Liu, D. Zhou, J. Bian, and T.-Y . Liu, “Qlib: An AI-oriented quantitative investment platform,”arXiv preprint arXiv:2009.11189, 2020

    Pith/arXiv arXiv 2009

  41. [41]

    A multi-factor stock selection system based on short-term price-volume characteristics

    C. Li and F. Liu, “A multi-factor stock selection system based on short-term price-volume characteristics” (in Chinese), Guotai Junan Securities, Quantitative Special Report No. 93, Jun. 2017

    work page 2017

  42. [42]

    J. R. Koza,Genetic Programming: On the Programming of Computers by Means of Natural Selection. Cambridge, MA: MIT Press, 1992

    work page 1992

  43. [43]

    A. E. Eiben and J. E. Smith,Introduction to Evolutionary Computing, 2nd ed. Berlin: Springer, 2015

    work page 2015

  44. [44]

    Alpha2: Dis- covering logical formulaic alphas using deep reinforcement learning,

    F. Xu, Y . Yin, X. Zhang, T. Liu, S. Jiang, and Z. Zhang, “Alpha2: Dis- covering logical formulaic alphas using deep reinforcement learning,” arXiv preprint arXiv:2406.16505, 2024

    Pith/arXiv arXiv 2024

  45. [45]

    AlphaQCM: Alpha discovery in finance with distributional reinforcement learning,

    Z. Zhu and K. Zhu, “AlphaQCM: Alpha discovery in finance with distributional reinforcement learning,” inProc. 42nd Int. Conf. Machine Learning (ICML), ser. Proceedings of Machine Learning Research, vol. 267, 2025, pp. 80463–80479

    work page 2025

  46. [46]

    AlphaForge: A framework to mine and dynamically combine formulaic alpha factors,

    H. Shi, W. Song, X. Zhang, J. Shi, C. Luo, X. Ao, H. Arian, and L. A. Seco, “AlphaForge: A framework to mine and dynamically combine formulaic alpha factors,”Proc. AAAI Conf. Artif. Intell., vol. 39, no. 12, pp. 12524–12532, 2025, doi: 10.1609/aaai.v39i12.33365

    doi:10.1609/aaai.v39i12.33365 2025

  47. [47]

    AlphaPROBE: Alpha mining via principled retrieval and on-graph biased evolution,

    T. Guo, H. Shen, J. Luo, B. Chen, H. Ding, J. Huang, L. Liu, Y . Ma, and M. Zhang, “AlphaPROBE: Alpha mining via principled retrieval and on-graph biased evolution,”arXiv preprint arXiv:2602.11917, 2026

    arXiv 2026

  48. [48]

    Quality diversity: A new frontier for evolutionary computation,

    J. K. Pugh, L. B. Soros, and K. O. Stanley, “Quality diversity: A new frontier for evolutionary computation,”Front. Robot. AI, vol. 3, p. 40, 2016

    work page 2016

  49. [49]

    Alpha dis- covery via grammar-guided learning and search,

    H. Yang, D. Hao, Z. Wang, Q. Shi, and X. Li, “Alpha dis- covery via grammar-guided learning and search,”arXiv preprint arXiv:2601.22119, 2026

    arXiv 2026

  50. [50]

    Alpha- GPT: Human-AI interactive alpha mining for quantitative investment,

    S. Wang, H. Yuan, L. Zhou, L. Ni, H.-Y . Shum, and J. Guo, “Alpha- GPT: Human-AI interactive alpha mining for quantitative investment,” inProc. 2025 Conf. Empirical Methods in Natural Language Process- ing: System Demonstrations (EMNLP Demo), 2025, pp. 196–206, doi: 10.18653/v1/2025.emnlp-demos.14

    work page doi:10.18653/v1/2025.emnlp-demos.14 2025

  51. [51]

    Can large language models mine interpretable financial factors more effectively? A neural-symbolic factor mining agent model,

    Z. Li, R. Song, C. Sun, W. Xu, Z. Yu, and J.-R. Wen, “Can large language models mine interpretable financial factors more effectively? A neural-symbolic factor mining agent model,” inFindings of the Association for Computational Linguistics: ACL 2024, 2024, pp. 3891– 3902, doi: 10.18653/v1/2024.findings-acl.233

    doi:10.18653/v1/2024.findings-acl.233 2024

  52. [52]

    R&D- Agent-Quant: A multi-agent framework for data-centric factors and model joint optimization,

    Y . Li, X. Yang, X. Yang, M. Xu, X. Wang, W. Liu, and J. Bian, “R&D- Agent-Quant: A multi-agent framework for data-centric factors and model joint optimization,”arXiv preprint arXiv:2505.15155, 2025

    arXiv 2025

  53. [53]

    From Feedback Loops to Policy Updates: Reinforcement Fine-Tuning for LLM-Based Alpha Factor Discovery

    L. Zhang, T. Jia, Y . Zhai, Z. Xie, C. Duan, M. He, P. S. Yu, and Y . Li, “From feedback loops to policy updates: Reinforcement fine-tuning for LLM-based alpha factor discovery,”arXiv preprint arXiv:2605.15412, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  54. [54]

    and the cross-section of expected returns,

    C. R. Harvey, Y . Liu, and H. Zhu, “. . . and the cross-section of expected returns,”Rev. Financ. Stud., vol. 29, no. 1, pp. 5–68, 2016

    work page 2016

  55. [55]

    Human-level control through deep reinforcement learning,

    V . Mnih et al., “Human-level control through deep reinforcement learning,”Nature, vol. 518, pp. 529–533, 2015

    work page 2015

  56. [56]

    A distributional per- spective on reinforcement learning,

    M. G. Bellemare, W. Dabney, and R. Munos, “A distributional per- spective on reinforcement learning,” inProc. 34th Int. Conf. Machine Learning (ICML), ser. Proceedings of Machine Learning Research, vol. 70, 2017, pp. 449–458

    work page 2017

  57. [57]

    Implicit quantile networks for distributional reinforcement learning,

    W. Dabney, G. Ostrovski, D. Silver, and R. Munos, “Implicit quantile networks for distributional reinforcement learning,” inProc. ICML, pp. 1096–1105, 2018

    work page 2018

  58. [58]

    Abandoning objectives: Evolution through the search for novelty alone,

    J. Lehman and K. O. Stanley, “Abandoning objectives: Evolution through the search for novelty alone,”Evol. Comput., vol. 19, no. 2, pp. 189–223, 2011

    work page 2011

  59. [59]

    Uncertain quality-diversity: Evaluation methodology and new methods for quality-diversity in uncertain do- mains,

    M. Flageat and A. Cully, “Uncertain quality-diversity: Evaluation methodology and new methods for quality-diversity in uncertain do- mains,”IEEE Trans. Evol. Comput., vol. 28, no. 4, pp. 891–902, 2024, doi: 10.1109/TEVC.2023.3273560

    work page arXiv 2024

  60. [60]

    Mastering the game of Go with deep neural networks and tree search,

    D. Silver et al., “Mastering the game of Go with deep neural networks and tree search,”Nature, vol. 529, pp. 484–489, 2016

    work page 2016

  61. [61]

    Efficient selectivity and backup operators in Monte-Carlo tree search,

    R. Coulom, “Efficient selectivity and backup operators in Monte-Carlo tree search,” inComputers and Games, ser. Lecture Notes in Computer Science, vol. 4630. Berlin, Germany: Springer, 2007, pp. 72–83, doi: 10.1007/978-3-540-75538-8_7

    doi:10.1007/978-3-540-75538-8_7 2007

  62. [62]

    Language models are few-shot learners,

    T. B. Brown et al., “Language models are few-shot learners,” inAdv. Neural Inf. Process. Syst. (NeurIPS), 2020

    work page 2020

  63. [63]

    Chain-of-thought prompting elicits reasoning in large language models,

    J. Wei et al., “Chain-of-thought prompting elicits reasoning in large language models,” inAdv. Neural Inf. Process. Syst. (NeurIPS), 2022

    work page 2022

  64. [64]

    ReAct: Synergizing reasoning and acting in language models,

    S. Yao et al., “ReAct: Synergizing reasoning and acting in language models,” inICLR, 2023

    work page 2023

  65. [65]

    Toolformer: Language models can teach themselves to use tools,

    T. Schick et al., “Toolformer: Language models can teach themselves to use tools,” inAdv. Neural Inf. Process. Syst. (NeurIPS), 2023

    work page 2023

  66. [66]

    Tree of thoughts: Deliberate problem solving with large language models,

    S. Yao et al., “Tree of thoughts: Deliberate problem solving with large language models,” inAdv. Neural Inf. Process. Syst. (NeurIPS), 2023

    work page 2023

  67. [67]

    AutoGen: Enabling next-gen LLM applications via multi- agent conversation,

    Q. Wu et al., “AutoGen: Enabling next-gen LLM applications via multi- agent conversation,”arXiv preprint arXiv:2308.08155, 2023

    Pith/arXiv arXiv 2023

  68. [68]

    Emergent autonomous scientific research capabilities of large language models,

    D. A. Boiko, R. MacKnight, and G. Gomes, “Emergent autonomous scientific research capabilities of large language models,”arXiv preprint arXiv:2304.05332, 2023

    Pith/arXiv arXiv 2023

  69. [69]

    Large language models for scientific discovery in molecular property prediction,

    Y . Zheng et al., “Large language models for scientific discovery in molecular property prediction,”Nat. Mach. Intell., vol. 7, no. 3, pp. 437–447, 2025, doi: 10.1038/s42256-025-00994-z

    doi:10.1038/s42256-025-00994-z 2025

  70. [70]

    Efficient global optimiza- tion of expensive black-box functions,

    D. R. Jones, M. Schonlau, and W. J. Welch, “Efficient global optimiza- tion of expensive black-box functions,”J. Global Optim., vol. 13, no. 4, pp. 455–492, 1998

    1998

  71. [71]

    Forrester, A

    A. Forrester, A. Sóbester, and A. Keane,Engineering Design via Surrogate Modelling. Chichester: Wiley, 2008

    work page 2008

  72. [72]

    Lopez de Prado,Advances in Financial Machine Learning

    M. Lopez de Prado,Advances in Financial Machine Learning. Hobo- ken, NJ: Wiley, 2018

    work page 2018

  73. [73]

    A reality check for data snooping,

    H. White, “A reality check for data snooping,”Econometrica, vol. 68, no. 5, pp. 1097–1126, 2000

    2000

  74. [74]

    The deflated Sharpe ratio: Correcting for selection bias, backtest overfitting and non-normality,

    D. H. Bailey and M. Lopez de Prado, “The deflated Sharpe ratio: Correcting for selection bias, backtest overfitting and non-normality,” J. Portfolio Management, vol. 40, no. 5, pp. 94–107, 2014

    work page 2014

  75. [75]

    The probability of backtest overfitting,

    D. H. Bailey et al., “The probability of backtest overfitting,”J. Computational Finance, vol. 20, no. 4, pp. 39–69, 2017

    work page 2017

  76. [76]

    A five-factor asset pricing model,

    E. F. Fama and K. R. French, “A five-factor asset pricing model,”J. Financ. Econ., vol. 116, no. 1, pp. 1–22, 2015

    work page 2015

  77. [77]

    R. S. Tsay,Analysis of Financial Time Series, 3rd ed. Hoboken, NJ: Wiley, 2010. 18

    work page 2010

  78. [78]

    Optimal execution of portfolio transac- tions,

    R. Almgren and N. Chriss, “Optimal execution of portfolio transac- tions,”J. Risk, vol. 3, no. 2, pp. 5–39, 2001

    work page 2001

  79. [79]

    Trading costs of asset pricing anomalies,

    A. Frazzini, R. Israel, and T. J. Moskowitz, “Trading costs of asset pricing anomalies,”J. Financ. Econ., vol. 132, no. 2, pp. 365–382, 2019

    work page 2019

  80. [80]

    Evolutionary principles in self-referential learning,

    J. Schmidhuber, “Evolutionary principles in self-referential learning,” Diploma thesis, TU Munich, 1987

    work page 1987

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.