Pith. sign in

REVIEW 3 major objections 4 minor 55 references

AQuA: Recursively Self-Improving Quantitative Trading Research Agents

T0 review · 3 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Two autonomous loops recursively improve quantitative research without leaking data, reaching a held-out Sharpe of +2.50.

desk verdict The sealed-sandbox design and honest failure case are genuinely valuable, but the headline evidence for recursive self-improvement is the validation metric the loop optimizes, so the central claim is not established as stated. read the letter →

arxiv 2608.12841 v1 pith:TZGYOUZM submitted 2026-08-13 cs.CL cs.AI

Jiacheng Guo , Suozhi Huang , Yunlong Gao , Zihao Li , Jian Ge , Xu Kuang , Mengdi Wang This is my paper · ORCID
classification cs.CLcs.AI
keywords recursiveself-improvementLLMagentsquantitativetradingfactordiscoverymodeldevelopmentdataleakagepreventionsealedsandboxinformationcoefficient
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that an autonomous system can improve its own quantitative-investment research when the definition of success is fixed and the agent cannot reach the data or the evaluator. It describes two separate loops: one discovers symbolic trading factors, the other develops a trainable model, and each retains validated evidence that guides its next proposal. In that bounded sense the paper claims recursive self-improvement happens at the level of the research process, not by rewriting the language model or the test. The reported payoffs are a combined factor information coefficient of about 0.190 on crypto data and a per-stock information coefficient of +0.0843 on US equities, which converts to a held-out long/short Sharpe of up to +2.50 at a two-leg cost.

What carries the argument

The load-bearing object is the sealed sandbox: a fixed tuple of data splits, feature and label definitions, and evaluator, paired with a domain-specific language whose expressions cannot reach those sealed components. The paper defines causality as closed under composition in the factor DSL, meaning the agent cannot express a future-looking normalizer or label even by accident. The second half of the mechanism is the split metric: the loop optimizes a validation score but reports a test metric computed once on an untouched window, so the number the agent is judged by is never the number it selects on.

What would settle it

Run the Part II loop a second time with the same seeds but return the test-window score during search instead of validation only; if the reported 2021–2025 IC and Sharpe survive that change, the sealed test window was not doing the work, and if they collapse, the headline numbers depend on never seeing the test window. Separately, audit the operator's access logs to the test store during the search: any read before the configuration is frozen falsifies the test-isolation claim.

Watch

Extended reading notes

Core claim

In AQuA's design, recursive self-improvement is made safe by asymmetric freedom: the agent can explore freely inside a restricted specification language, but every admissible action is compiled and scored by a sealed, human-authored sandbox whose data splits, features, labels, and evaluator the agent cannot modify. Because every time-series operator reads only a trailing window and every cross-sectional operator only the current timestamp, any expression the agent assembles is causal by construction. Selection leakage is handled separately: during search the agent receives only a fixed validation-score signal, while the designated test window is scored once after the configuration is frozen and never returned to the loop. The paper reports that this design lets the factor system reach a combined validation information coefficient of about 0.190 and lets the model system beat a GRU baseline (per-stock IC +0.0843 vs +0.0613) and produce a sector-neutral, volatility-targeted long/short book with a held-out Sharpe of +2.50, positive in every year from 2021 to 2025.

Load-bearing premise

The fixed validation slice that the harness returns to the agent during search is assumed to remain an honest measure of out-of-sample skill after many rounds of selection on that same slice, and the paper concedes that keeping the final test window untouched rests on operator discipline rather than a hard technical barrier.

Editorial extensions

If this is right

  • If the reported results hold, an autonomous research loop can accumulate validated evidence across iterations and improve factor and model quality without human review of each candidate.
  • The test-window numbers (per-stock IC +0.0843, Sharpe +2.50, positive in every year from 2021 to 2025) would be out of sample with respect to the whole search, not just the final model fit.
  • The sealed-sandbox recipe would make leakage-inducing actions structurally unavailable in any autonomous research agent, rather than relying on a reviewer to catch subtle errors.
  • Causality by construction in the factor DSL means every expression the agent can emit is a legitimate candidate, so the loop can run without per-candidate human review within its fixed research contract.
  • The fully causal walk-forward evaluation retains a Sharpe of about +2.0, implying the headline result is not the artifact of one favorable configuration or parameter choice.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond finance, the two-channel split (generation leakage versus selection leakage) is a general recipe: any LLM-driven discovery loop should make the evaluator unreachable and report a metric the loop never optimizes.
  • A testable extension is to track the gap between the validation IC the loop optimizes and the untouched test IC over successive iterations; a widening gap would reveal residual selection leakage even under a nominally sealed test window.
  • The paper itself flags coupling the two systems as the next step, but also notes that coupling introduces a new leakage channel; a concrete corollary is that the factor library must be frozen before model search begins if the model loop is not to steer factor selection.
  • The demonstrated autonomy is bounded by a human operator who sets goals, owns the sandbox, and supervises promotion, so the results do not imply unattended operation on new markets or frequencies without re-tuning.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper introduces AQuA, a pair of independent LLM-driven research systems for quantitative investment research: one for symbolic factor discovery (Part I) and one for trainable model development (Part II). The central claim is that each system closes its own research loop by retaining validated evidence and using it to guide later proposals, thereby implementing recursive self-improvement at the level of the research process, while a sealed sandbox prevents data leakage by construction. Part I reports a combined validation Spearman IC of about 0.190 on a crypto universe; Part II reports a per-stock IC of +0.0843 on US equities and a threshold long/short held-out Sharpe of up to +2.50 at a two-leg cost, positive in every year from 2021 to 2025. The paper also reports a fully causal walk-forward Sharpe of about +2.0 for Part II.

Significance. If fully supported, the paper would make a useful conceptual contribution by separating leakage into a generation channel and a selection channel and by insisting that the metric used for ranking candidates differ from the metric reported as the result. The Part II evaluation is a genuine strength: the 2021–2025 test window is untouched by model selection, and the fully causal walk-forward procedure addresses a common failure mode in backtesting. The worked failure case in Appendix B and the candid Limitations section are also valuable. However, the principal quantitative evidence for Part I's self-improvement is the validation objective itself, and Part II's held-out result is a single final point rather than evidence of improvement over iterations. The central claim of recursive self-improvement is therefore not yet established with the evidence presented.

major comments (3)
  1. [Section 4.5, Figure 3, Section 3.1] The Part I headline result, a combined validation IC of approximately 0.190, is measured on the validation slice that the search loop uses to rank candidates and select factors (Section 3.1). No untouched test-set IC for Part I is reported anywhere. The paper's own defense against selection leakage, 'reporting a metric the loop never optimizes against' (Section 6), does not apply to Part I because the reported 0.190 is the optimized validation metric itself. The monotone improvement in Figure 3 is exactly what adaptive overfitting to a fixed validation slice would produce. Please report a Part I evaluation on a test window that is frozen before the final combined signal is fixed and never returned to the agent, and disclose the number of candidate factors and selection rounds so the multiple-testing burden can be assessed.
  2. [Section 5.1, Section 5.6, Table 3] Part II's untouched 2021–2025 test window is a genuine out-of-sample result, and the walk-forward Sharpe is a strength. However, a single final test point cannot by itself establish recursive self-improvement. The paper does not report how many configuration diffs were tried, how many autonomous iterations the loop ran, or how validation or test performance evolved over those iterations. Without this information, the result is compatible with a broad search that found one good configuration rather than an autonomous loop whose research process improves over time. Please report the iteration trajectory and the total number of configurations evaluated, and, if possible, the test-window performance of the final few iterations.
  3. [Abstract, Section 7, Section 6] The abstract and Section 6 claim that the systems are leakage-free 'by construction' and that reported results are out of sample, but Section 7 concedes that keeping the final test window out of selection 'rests on the sealed protocol and operator discipline rather than a hard technical barrier.' This caveat is especially important for Part I, where no final test window is designated at all. The unqualified language in the abstract and Section 6 should be revised to state that Part I's headline IC is a validation-slice number and that the isolation of any final test evaluation is a governance property rather than a cryptographic guarantee.
minor comments (4)
  1. [Figure 3] The legend includes a 'Fit-to-validation IC range' that is never defined in the text; please explain how this range is computed and why it is reported alongside the combined validation IC.
  2. [Sections 4 and 5] Several important experimental details are missing: the crypto universe is not specified (exchange, coin list, data vendor), the baseline implementations in Figure 3 and Table 2 are not accompanied by hyperparameters or adaptation details, and the exact feature set and convolutional configuration for Part II are explicitly left undisclosed. This limits reproducibility.
  3. [Listing 1 and Appendix A] The 'expression: withheld' entries mean that no actual factor formula is disclosed for Part I, so an independent reader cannot inspect, reproduce, or falsify the specific factor claimed to carry the signal.
  4. [Section 5.3] The sentence describing the split as 'train on 2010–2019, leave 2020 as an embargo gap' is clear, but the same paragraph says 'the exact feature set, normalization, and label construction are part of that sandbox and are not disclosed'; please clarify whether the label is the 30-minute forward return mentioned in Section 5.3 or a different construction.

Circularity Check

1 steps flagged · score 6.0 of 10

Part I's headline 0.190 combined IC is the validation score the search loop optimizes, so the main evidence for recursive improvement in factor discovery is the optimizer's own objective; Part II's untouched test window keeps the paper only partially circular.

  1. fitted input called prediction [Section 3.1 (sealed sandbox) and Section 4.5 / Figure 3]
    "During search the harness returns to the agent only a score on a validation slice fixed in advance; the score s_k that ranks candidates and the inner-validation signal that drives early stopping and checkpoint choice are computed on that slice alone. ... The combined validation information coefficient rises as the loop accumulates and reuses evidence, reaching approximately 0.190."

    The Part I headline result—combined validation IC of about 0.190—is the value of the exact score used to rank candidates and select/combine factors during search. Section 3.1 says the only score returned to the agent is computed on a fixed validation slice, and Section 4.5 reports that same 'combined validation information coefficient' rising across iterations as the evidence of improvement. The reported rise is therefore the optimizer's objective value after selection, not an untouched out-of-sample measure.

full rationale

Part I's evidence for recursive self-improvement is the combined validation IC reaching about 0.190 (Section 4.5, Figure 3). Section 3.1 establishes that this same validation slice is the only score returned to the agent during search and is the score that ranks candidates; no untouched test-set IC is reported for Part I. Hence the headline number is the optimization objective's value, and reporting its increase as evidence of research-process improvement is a fitted-input-called-prediction: adaptive selection on the fixed validation slice would produce the same curve without any genuine improvement in research quality. The paper's split-metric defense applies only where a final test window is scored once after freezing, and Part I has none. Section 7 explicitly concedes that test isolation rests on 'operator discipline rather than a hard technical barrier,' which is especially relevant when the reported number is merely the validation score. Part II is genuinely different: the 2021–2025 window is scored once after the configuration is frozen, is never returned to the agent, and the fully causal walk-forward Sharpe of about +2.0 provides independent support. Therefore the circularity is partial—the central 'recursive self-improvement' claim still has independent content through Part II—so a score of 6 rather than 8 or higher is appropriate. No load-bearing self-citation or imported uniqueness theorem appears in the paper.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central claims rest on several unverified assumptions: the existence of exploitable alpha, the unbiasedness of the repeated validation slice, the correctness of the restricted DSL's causal closure, and the realism of the 2bp cost model. The paper is open about some of these, but provides no experiments isolating them.

free parameters (3)
  • two_leg_cost_bps = 2
    Assumed trading cost per two legs; the Sharpe ratio is highly sensitive to this parameter and no sensitivity analysis is provided.
  • strategy thresholds and volatility-targeting parameters = not disclosed
    Parameters of the long/short book and the volatility-targeting overlay are tuned on a training fraction; their values are not specified in the paper.
  • Part I factor acceptance thresholds = not disclosed
    The Factor Evaluator's thresholds for IC, stability, turnover, and correlation with the existing pool are not specified, and they control how many factors survive to the combination layer.
assumptions (6)
  • domain assumption Historical price-volume data in crypto and US equities contain exploitable predictive structure at five-minute and thirty-minute horizons after costs.
    The entire system presupposes that such alpha exists and is not fully arbitraged away; Section 5.3 states the signal 'lives in their joint, nonlinear, and temporal structure.'
  • domain assumption The frozen validation slice remains unbiased after many iterations of selection on that same slice.
    Section 3.1 returns only validation scores to the agent; the paper acknowledges in Section 7 that test isolation is a governance property, not a technical guarantee.
  • domain assumption Causality is closed under composition of the operator registry, so no DSL expression can leak future information.
    Section 3.1 argues every time-series operator reads only a trailing window and every cross-sectional operator reads only the current timestamp; this assumes the registry itself has no hidden look-ahead and that the compiler implements it as specified.
  • standard math Information coefficients are an adequate measure of factor or model quality for iterative selection.
    The evaluator uses IC as the optimization target; this is a standard but reductive choice that ignores risk-adjusted return and tail behavior.
  • domain assumption The two-basis-point two-leg cost model approximates real intraday trading costs.
    No cost-model benchmark or sensitivity analysis is provided even though the Sharpe depends directly on this assumption.
  • ad hoc to paper LLM-generated factor hypotheses with stated mechanisms and falsification criteria reduce data dredging relative to bare formula search.
    The paper asserts this discipline in Section 4.2 but provides no controlled comparison against a baseline that searches formulas without hypotheses.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AQuA: Recursively Self-Improving Quantitative Trading Research Agents." pith.science (2026). https://pith.science/paper/TZGYOUZM

@misc{pith2026260812841,
  author       = {Pith},
  title        = {Pith review of: AQuA: Recursively Self-Improving Quantitative Trading Research Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TZGYOUZM}},
  note         = {Machine review of arXiv:2608.12841}
}
abstract

We study recursive self-improvement at the level of quantitative-investment research: whether an autonomous system can use evidence from earlier experiments to improve the hypotheses and candidates proposed in later iterations. We present AQuA, which comprises two separate language-model-driven research systems: one for symbolic factor discovery and one for trainable model development. The two systems do not share agents, memories, candidate spaces, or research state. Instead, each independently closes its own research loop by retaining validated evidence and using it to guide subsequent proposals. In this bounded sense, both systems implement recursive self-improvement at the level of the research process. Each system also uses its own sealed sandbox, which fixes the data splits, feature and label definitions, and evaluator while allowing the model to act only through constrained factor expressions or configuration diffs. The factor system, a manager-mediated multi-agent pipeline, discovers and combines factors into a signal that reaches a combined information coefficient of about $0.190$ on a crypto universe. The model system, a config-driven loop over a hybrid time-series architecture, reaches a per-stock information coefficient of $+0.0843$ on US equities and converts it into a threshold long/short strategy with a held-out Sharpe of up to $+2.50$ at a two-leg cost. The strategy is positive in every year from 2021 to 2025.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

55 extracted references · 27 canonical work pages

  1. [1]

    Proceedings of the 32nd International Conference on Machine Learning , series =

    The Ladder: A Reliable Leaderboard for Machine Learning Competitions , author =. Proceedings of the 32nd International Conference on Machine Learning , series =. 2015 , url =

    work page 2015

  2. [2]

    2024 , doi =

    Sycophancy to Subterfuge: Investigating Reward-Tampering in Large Language Models , author =. 2024 , doi =

    work page 2024

  3. [3]

    2025 , doi =

    Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation , author =. 2025 , doi =

    work page 2025

  4. [4]

    2026 , doi =

    Atinafu, Yonas and Cohen, Robin , publisher =. 2026 , doi =

    work page 2026

  5. [5]

    Lu, Chris and Lu, Cong and Lange, Robert Tjarko and Foerster, Jakob and Clune, Jeff and Ha, David , month = sep, year =. The. doi:10.48550/arXiv.2408.06292 , publisher =

    arXiv doi:10.48550/arxiv.2408.06292

  6. [6]

    doi:10.48550/arXiv.2602.07085 , publisher =

    Han, Jun and Zhang, Shuo and Li, Wei and Dong, Yifan and Hu, Tu and Zhu, Yumo and Yu, Xiaomin and Guo, Xin and Liu, Zhaowei and Wang, Kunyi and Liu, Jingping and Jiang, Tianyi and An, Ruichuan and Hu, Sen and Yang, Zhi and Che, Ronghao and Wang, Huacan , month = may, year =. doi:10.48550/arXiv.2602.07085 , publisher =

    arXiv doi:10.48550/arxiv.2602.07085

  7. [7]

    doi:10.48550/arXiv.1808.03668 , publisher =

    Zhang, Zihao and Zohren, Stefan and Roberts, Stephen , year =. doi:10.48550/arXiv.1808.03668 , publisher =

    arXiv doi:10.48550/arxiv.1808.03668

  8. [8]

    Kakushadze, Zura , month = mar, year =. 101. doi:10.48550/arXiv.1601.00991 , publisher =

    arXiv doi:10.48550/arxiv.1601.00991

Show all 55 references
  1. [9]

    doi:10.48550/arXiv.2002.08245 , publisher =

    Zhang, Tianping and Li, Yuanqi and Jin, Yifei and Li, Jian , month = apr, year =. doi:10.48550/arXiv.2002.08245 , publisher =

    2002 doi arXiv

  2. [10]

    doi:10.1145/3448016.3457324 , booktitle =

    Cui, Can and Wang, Wei and Zhang, Meihui and Chen, Gang and Luo, Zhaojing and Ooi, Beng Chin , month = jun, year =. doi:10.1145/3448016.3457324 , booktitle =

  3. [11]

    Generating

    Yu, Shuo and Xue, Hongyan and Ao, Xiang and Pan, Feiyang and He, Jia and Tu, Dandan and He, Qing , month = may, year =. Generating. doi:10.48550/arXiv.2306.12964 , publisher =

    doi arXiv

  4. [12]

    Notices of the American Mathematical Society , author =

    Pseudo-. Notices of the American Mathematical Society , author =. 2014 , pages =. doi:10.1090/noti1105 , number =

    2014 doi

  5. [13]

    The probability of backtest overfitting , issn =

    Bailey, David and Borwein, Jonathan and López De Prado, Marcos and Zhu, Qiji Jim , year =. The probability of backtest overfitting , issn =. doi:10.21314/JCF.2016.322 , journal =

    2016 doi

  6. [14]

    Zico and Koltun, Vladlen , month = apr, year =

    Bai, Shaojie and Kolter, J. Zico and Koltun, Vladlen , month = apr, year =. An. doi:10.48550/arXiv.1803.01271 , publisher =

    doi arXiv

  7. [15]

    Gu, Albert and Dao, Tri , year =. Mamba:. doi:10.48550/arXiv.2312.00752 , publisher =

    doi arXiv

  8. [16]

    and Kaiser, Lukasz and Polosukhin, Illia , year =

    Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob and Jones, Llion and Gomez, Aidan N. and Kaiser, Lukasz and Polosukhin, Illia , year =. Attention. doi:10.48550/arXiv.1706.03762 , publisher =

    doi arXiv

  9. [17]

    2026 , note=

    Tang, Ziyi and Yin, Xuexiong and Chen, Weixing and Chen, Zechuan and Zheng, Yongsen and Ye, Wenxuan and Wang, Keze and Lin, Liang , booktitle=. 2026 , note=

    2026

  10. [18]

    Luo, Haochen and Ko, Ho Tin and Chen, Jiandong and Sun, David and Zhang, Yuan and Liu, Chen , booktitle=

  11. [19]

    doi:10.48550/arXiv.2603.16365 , publisher =

    Lin, Qinhong and Feng, Ruitao and Feng, Yinglun and Huang, Zhenxin and Chen, Yukun and Yang, Zhongliang and Zhou, Linna and Fei, Binjie and Liu, Jiaqi and Li, Yu , month = apr, year =. doi:10.48550/arXiv.2603.16365 , publisher =

    doi arXiv

  12. [20]

    doi:10.48550/arXiv.2602.11917 , publisher =

    Guo, Taian and Shen, Haiyang and Luo, Junyu and Chen, Binqi and Ding, Hongjun and Huang, Jinsheng and Liu, Luchen and Ma, Yun and Zhang, Ming , month = feb, year =. doi:10.48550/arXiv.2602.11917 , publisher =

    doi

  13. [21]

    doi:10.48550/arXiv.2602.14670 , publisher =

    Wang, Yanlong and Xu, Jian and Zhang, Hongkang and Huang, Shao-Lun and Sun, Danny Dongning and Zhang, Xiao-Ping , month = feb, year =. doi:10.48550/arXiv.2602.14670 , publisher =

    doi

  14. [22]

    Shi, Runze and Yan, Shengyu and Cai, Yuecheng and Lv, Chengxi , month = apr, year =. Hubble:. doi:10.48550/arXiv.2604.09601 , publisher =

    doi arXiv

  15. [23]

    doi:10.48550/arXiv.2603.20247 , publisher =

    Weng, Zhangyuhua and Zhang, Shengli and Wang, Taotao and Xia, Yihan , month = mar, year =. doi:10.48550/arXiv.2603.20247 , publisher =

    doi

  16. [24]

    and Liu, Tongliang and Wang, Zhiyong and He, Fengxiang , month = may, year =

    Yu, Hang and Zheng, Zifan and Pan, Jeff Z. and Liu, Tongliang and Wang, Zhiyong and He, Fengxiang , month = may, year =. doi:10.48550/arXiv.2606.20625 , publisher =

    doi arXiv

  17. [25]

    and Li, Ying , month = may, year =

    Zhang, Lingzhe and Jia, Tong and Zhai, Yunpeng and Xie, Zixuan and Duan, Chiming and He, Minghua and Yu, Philip S. and Li, Ying , month = may, year =. From. doi:10.48550/arXiv.2605.15412 , publisher =

    doi arXiv

  18. [26]

    The Review of Financial Studies , author =

    Empirical. The Review of Financial Studies , author =. 2020 , pages =. doi:10.1093/rfs/hhaa009 , number =

    2020 doi

  19. [27]

    Review of Financial Studies , author =

    … and the. Review of Financial Studies , author =. 2016 , pages =. doi:10.1093/rfs/hhv059 , number =

    2016 doi

  20. [28]

    Neural Computation , author =

    Long. Neural Computation , author =. 1997 , pages =. doi:10.1162/neco.1997.9.8.1735 , number =

    1997 doi

  21. [29]

    Learning

    Cho, Kyunghyun and Merrienboer, Bart van and Gulcehre, Caglar and Bahdanau, Dzmitry and Bougares, Fethi and Schwenk, Holger and Bengio, Yoshua , month = sep, year =. Learning. doi:10.48550/arXiv.1406.1078 , publisher =

    doi arXiv

  22. [30]

    doi:10.48550/arXiv.2405.04517 , publisher =

    Beck, Maximilian and Pöppel, Korbinian and Spanring, Markus and Auer, Andreas and Prudnikova, Oleksandra and Kopp, Michael and Klambauer, Günter and Brandstetter, Johannes and Hochreiter, Sepp , month = dec, year =. doi:10.48550/arXiv.2405.04517 , publisher =

    doi arXiv

  23. [31]

    Nature , author =

    Mathematical discoveries from program search with large language models , volume =. Nature , author =. 2024 , pages =. doi:10.1038/s41586-023-06924-6 , number =

    2024 doi

  24. [32]

    Proceedings of the AAAI Conference on Artificial Intelligence , author =

    Navigating the. Proceedings of the AAAI Conference on Artificial Intelligence , author =. 2026 , pages =. doi:10.1609/aaai.v40i2.37069 , number =

    2026 doi

  25. [33]

    doi:10.1109/TSP.2025.3576781 , journal =

    Zhao, Junjie and Zhang, Chengxi and Qin, Min and Yang, Peng , year =. doi:10.1109/TSP.2025.3576781 , journal =

    2025

  26. [34]

    doi:10.1145/3711896.3736838 , booktitle =

    Tang, Ziyi and Chen, Zechuan and Yang, Jiarui and Mai, Jiayao and Zheng, Yongsen and Wang, Keze and Chen, Jinrui and Lin, Liang , month = aug, year =. doi:10.1145/3711896.3736838 , booktitle =

  27. [35]

    doi:10.1109/ICASSP55912.2026.11463591 , booktitle =

    Wu, Yihong and Lou, Chengcan and Zhang, Jiayuan and Chen, Sirong and Yang, Yan , month = may, year =. doi:10.1109/ICASSP55912.2026.11463591 , booktitle =

    2026

  28. [36]

    Chain-of-

    Cao, Lang , month = aug, year =. Chain-of-. doi:10.48550/arXiv.2508.06312 , publisher =

    doi

  29. [37]

    Learning from

    Zhao, Junjie and Zhang, Chengxi and Wang, Chenkai and Yang, Peng , month = jul, year =. Learning from. doi:10.48550/arXiv.2507.20263 , publisher =

    doi arXiv

  30. [38]

    Chen, Qizhao and Kawashima, Hiroaki , month = dec, year =. Multi-. doi:10.1109/BigData66926.2025.11400963 , booktitle =

    2025

  31. [39]

    Exploring the

    Guo, Tian and Hauptmann, Emmanuel , month = nov, year =. Exploring the. doi:10.48550/arXiv.2510.15691 , publisher =

    doi

  32. [40]

    doi:10.48550/arXiv.2606.13662 , publisher =

    Xin, Amy and Siow, Jiening and Wang, Junjie and Yao, Zijun and Zhang, Fanjin and Song, Jian and Hou, Lei and Li, Juanzi , month = jun, year =. doi:10.48550/arXiv.2606.13662 , publisher =

    doi arXiv

  33. [41]

    Autonomous

    Zhou, Lianhao and Ling, Hongyi and Fu, Cong and Huang, Yepeng and Sun, Michael and Yu, Wendi and Wang, Xiaoxuan and Li, Xiner and Su, Xingyu and Zhang, Junkai and Chen, Xiusi and Liang, Chenxing and Qian, Xiaofeng and Ji, Heng and Wang, Wei and Zitnik, Marinka and Ji, Shuiwang...

    doi arXiv

  34. [42]

    Miyazaki, Kunihiro and Kawahara, Takanobu and Roberts, Stephen and Zohren, Stefan , month = jun, year =. Toward. doi:10.3905/jfds.2026.008 , journal =

    2026 doi

  35. [43]

    doi:10.2139/ssrn.5580590 , author =

    An. doi:10.2139/ssrn.5580590 , author =

    doi

  36. [44]

    2025 , pages =

    Sci , author =. 2025 , pages =. doi:10.3390/sci7010007 , number =

    2025 doi

  37. [45]

    Journal of Risk and Financial Management , author =

    A. Journal of Risk and Financial Management , author =. 2026 , pages =. doi:10.3390/jrfm19030203 , number =

    2026 doi

  38. [46]

    Forecasting , author =

    From. Forecasting , author =. 2025 , pages =. doi:10.3390/forecast7040055 , number =

    2025 doi

  39. [47]

    Deep learning and machine learning models for portfolio optimization:

    Ashrafzadeh, Mahdi and Sadrani, Mohammad and Zolfani, Sarfaraz Hashemkhani , month = sep, year =. Deep learning and machine learning models for portfolio optimization:. doi:10.1016/j.rineng.2025.106263 , journal =

    2025

  40. [48]

    Enhancing

    Wang, Tianye , year =. Enhancing. doi:10.1016/j.procs.2025.12.015 , journal =

    2025 doi

  41. [49]

    doi:10.2139/ssrn.6085266 , author =

    Time. doi:10.2139/ssrn.6085266 , author =

    doi

  42. [50]

    doi:10.2139/ssrn.5166656 , author =

    Large language models in finance: estimating financial sentiment for stock prediction , url =. doi:10.2139/ssrn.5166656 , author =

    doi

  43. [51]

    doi:10.2139/ssrn.6906675 , author =

    Self-. doi:10.2139/ssrn.6906675 , author =

    doi

  44. [52]

    From Hypotheses to Factors: Constrained LLM Agents in Cryptocurrency Markets , url =

    Huang, Yikuan and Fan, Zheqi and Hu, Kaiqi and Ye, Yifan , month = apr, year =. From Hypotheses to Factors: Constrained LLM Agents in Cryptocurrency Markets , url =. doi:10.48550/arXiv.2604.26747 , publisher =

    doi arXiv

  45. [53]

    doi:10.48550/arXiv.2607.26642 , publisher =

    Yi, Jingyang and Yang, Jian and Jin, Yifei and Li, Yuqi and Li, Jian , month = jul, year =. doi:10.48550/arXiv.2607.26642 , publisher =

    doi arXiv

  46. [54]

    Cognitive Alpha Mining via LLM-Driven Code-Based Evolution , url =

    Liu, Fengyuan and Huang, Yi and Luo, Sichun and Wang, Yuqi and Yang, Yazheng and Li, Xinye and Hu, Zefa and Feng, Junlan and Liu, Qi , month = nov, year =. Cognitive Alpha Mining via LLM-Driven Code-Based Evolution , url =. doi:10.48550/arXiv.2511.18850 , publisher =

    doi arXiv

  47. [55]

    Towards Autonomous Formulaic Alpha Discovery: An Evolutionary Computation Perspective , url =

    Yu, Xinwei and Fu, Yiyang and Fan, Mingcheng and Li, Enqi and Gao, Yilin and Xu, Shugong , month = aug, year =. Towards Autonomous Formulaic Alpha Discovery: An Evolutionary Computation Perspective , url =. doi:10.48550/arXiv.2608.01789 , publisher =

    doi arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.