AlphaSchema: Exploring the Space of Trading Semantics for LLM-Based Alpha Mining

Jian Li; Jian Yang; Jingyang Yi; Yifei Jin; Yuqi Li

REVIEW 4 major objections 6 minor 30 references

AlphaSchema: Exploring the Space of Trading Semantics for LLM-Based Alpha Mining

T0 review · 4 major / 6 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read AlphaSchema searches over explicit trading-semantic plans to mine alpha factors, and reports factor pools that outperform code-centric LLM mining baselines on Chinese stocks.

desk verdict Genuinely new search abstraction with a credible held-out setup, but the performance headline rests on single numbers from a large selection surface and needs variance reporting. read the letter →

arxiv 2607.26642 v1 pith:NGQGNQMX submitted 2026-07-29 cs.AI

Jingyang Yi , Jian Yang , Yifei Jin , Yuqi Li , Jian Li This is my paper · ORCID

classification cs.AI

keywords alphaminingtradingsemanticsschemaplanLLMagentfactordiscoverysurrogatemodeliterativesearchChinesestockmarket

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

AlphaSchema tries to establish that the right search space for automated alpha mining is not formulas or code but structured trading semantics: each candidate factor is first described as a plan p = (event, context, qualities, direction, output), and an LLM is only called in to turn the chosen plan into executable code. A surrogate reward model trained on evaluated plan–reward pairs then guides iterative selection, so the search can be optimized, controlled, and reused without depending on the discovery instincts of a particular language model. On the CSI300 universe over 2023–2025, the OHLCV factor pool reaches IC 0.0382 and ICIR 0.2374—the best among compared machine-learning, deep-learning, factor-library, and agentic-mining baselines on those metrics—and the fundamental-augmented pool reaches IR 1.0877 and annualized excess return 11.94%. The paper also claims that the same schema plan produces comparable-quality factors across different LLM backends, and that each of the five schema dimensions carries complementary signal. If these claims hold, alpha mining becomes a semantic-space optimization problem rather than a black-box prompting problem, with a cheaper and more interpretable path to factor discovery.

What carries the argument

The load-bearing object is the schema plan p = (e, c, Q, d, o) — a structured tuple of Event, Context, Qualities (0–3), Direction, and Output that states the trading logic before code exists. Search moves through the combinatorial space of these plans, and the machinery that makes the space navigable is the surrogate reward model: a gradient-boosted tree trained on accumulated plan–reward pairs, using one-hot schema components and pairwise interactions as features. Each evaluation feeds back a scalar reward r = α·RankIC + β·RankICIR − λ·Δlag (with Δlag penalizing the drop in RankIC when the signal is delayed one day), and an adaptive quota selector blends novelty-based exploration, surrogate

What would settle it

Take the same search pipeline and replace the plan rewards with random noise before pool selection, keeping the validation and test windows unchanged; if the noise-selected pools still reach IC ≈ 0.038 and ICIR ≈ 0.237, then the surrogate and reward signal are not carrying the result. Alternatively, compute the Spearman correlation between each exported factor's training/validation reward and its 2023–2025 IC; a correlation near zero across the 120/150 factors would falsify the claim that in-sample reward ranks generalize out of sample.

Watch

Extended reading notes

Core claim

The central claim is that alpha mining can be reformulated as search over a structured semantic space before implementation. Every candidate is a schema plan p = (e, c, Q, d, o): an event being captured, a context in which it is interpreted, zero to three quality constraints, a directional hypothesis about future returns, and an output form. These plans are the search object; a language model realizes selected plans as executable factors, which are validated, backtested, and scored by r = α·RankIC + β·RankICIR − λ·Δlag. The scores accumulate in a buffer that trains a gradient-boosted tree surrogate over schema features, and an adaptive quota selector allocates each round's budget among struc

Load-bearing premise

The central claim stands or falls on whether rewards computed in the training/validation window—with hand-set weights (10,1,2)—are noise-free enough that greedily picking the highest-reward, least-correlated factors produces a pool that still predicts in 2023–2025; if in-sample reward ranking is largely selection noise, the reported IC/ICIR gains are inflated.

Editorial extensions

If this is right

If the semantic-space formulation is right, factor mining no longer needs to be tied to a single frontier LLM; the paper reports that seven different backends produce factors with |RankIC| in a narrow range, so weaker models mostly hurt success rates, not signal quality.
The explicit plan vocabulary turns discovery into a reusable prior: evaluated reward–plan pairs accumulate, so later search rounds can exploit what earlier rounds learned about which trading semantics pay off.
The ablation results imply that all five schema dimensions carry complementary information; removing any one lowers retained |RankIC| to 69–79% of the full plan, even though most incomplete plans still produce valid code.
The budget analysis supports a single-realization protocol: repeated rollouts improve plan-quality estimates but cut top-plan recall per unit of budget, so broad coverage beats per-plan precision in large-scale mining.
With the reported factor pools, a Top50/Drop5 portfolio on the fundamental-augmented 150-factor pool achieves IR 1.0877 and 11.94% annualized excess return net of costs on the 2023–2025 test window.

Reading between the lines

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

Editorial extension: the explicit semantic vocabulary is a natural substrate for injecting domain priors—a fund could encode its known edge as new schema records and let the surrogate search around it; the paper does not demonstrate targeted insertion, but the architecture allows it.
Editorial extension: if plan-level reward structure transfers, the surrogate trained on one market could seed discovery in a new universe as a warm start, turning the paper's out-of-universe CSI500 check into a transfer benchmark; the paper leaves this shortcut untested.
Editorial inference: the robustness of factor quality across LLMs suggests the frontier of the system is the semantic vocabulary and reward signal, not the code-generating model; investing in better schema records and reward definitions may yield larger gains than switching to a costlier LLM.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

AlphaSchema is worth reading. The one genuinely new thing is the search object: a five-field semantic plan (Event, Context, Qualities, Direction, Output) that sits between the trading idea and the code. Prior LLM alpha-mining systems search formulas, programs, or trajectories; this one searches a structured description of the trading mechanism and learns a LightGBM surrogate over schema features to guide exploration, exploitation, and mutation. That is a real abstraction change, and the paper supports it with unusually complete appendices: the worked example is causal and specific, the held-out protocol (selection only on 2016–2022, test 2023–2025) is credible, and the vocabulary and mutation details are all in the appendix.

The soft spots are in the quantitative claim, not the architecture. Table 1 gives a single number per metric even though the paper says there were five independent runs. The headline IC margin over LSTM is 0.0002, and the OHLCV pool has the worst MDD and a lower IR than RD-Agent. Since the final pool is the top-reward set from roughly 6,400 evaluated plans, with no multiple-testing adjustment and no reported run-to-run variance, the IC/ICIR leads could easily be selection noise. I would want a permutation test or seed-wise error bars before taking 'strongest' literally.

Two smaller issues. The LLM-robustness claim is stronger than the evidence: mean |RankIC| across backends ranges 0.0116–0.0168, a 45% spread, and Pass@1 varies from 49% to 97%. That is 'comparable but not identical,' not 'largely robust.' And the semantic-navigation analysis (Table 4) shows the search moving into high-reward regions, but that is the exploitation mechanism itself—evidence the selector works as designed, not a separate discovery.

Overall, the central argument holds up: explicit semantic plans are a learnable, controllable, and cheaper search space. The paper deserves serious peer review; it needs error bars and a toned-down headline before the performance conclusion is solid.

I'd bring it to the reading group for a discussion of what 'search space' should mean in LLM-based discovery.

Referee Report

4 major / 6 minor

Summary. AlphaSchema proposes to replace implicit, code-level search in LLM-based alpha mining with an explicit semantic plan space p = (e, c, Q, d, o). A LightGBM reward surrogate is trained on evaluated plan–reward pairs, and a quota selector allocates each round's implementation budget among structural exploration, surrogate-guided exploitation, and local mutation. An LLM translates selected plans into executable factors, which are subjected to execution, numerical, and look-ahead checks before reward computation. On CSI300 with selection on 2016–2022 and a held-out 2023–2025 period, the OHLCV pool reports the strongest IC (0.0382) and ICIR (0.2374) in Table 1, while the +Fundamental variant reports the best IR (1.0877) and AER (11.94%). Additional analyses examine semantic component ablation, search trajectory, realization budget, and LLM-backend robustness.

Significance. The central idea — searching an explicit semantic space of trading mechanisms rather than over generated code — is timely and practically attractive: it can reduce reliance on expensive frontier LLMs and makes exploration more controllable. The manuscript has credible strengths: the held-out protocol is described clearly; the CSI500 run is an independent out-of-universe check; Appendix D contains a concrete causal schema-to-code example; Appendix E includes a shuffled-reward control for schema-level reward predictability. If the performance claims survive a proper treatment of selection noise, this would be a useful advance for automated factor discovery.

major comments (4)

[Table 1, §4 (Main Results)] The paper states that five independent runs were conducted, but Table 1 reports a single number per metric with no standard deviation, confidence interval, or per-run breakdown. The headline OHLCV IC is 0.0382 versus 0.0380 for LSTM (Δ = 0.0002), and ICIR is 0.2374 versus 0.2269. Such margins are within plausible cross-run sampling noise. A claim of being 'the strongest' requires error bars and a paired significance test, especially since the paper highlights the best method across multiple metrics and ten systems. Please report per-run results and appropriate tests.
[Eq. (5), final-pool selection, Appendix A] The final factor pool is chosen greedily by in-sample reward from roughly 6,400 plan evaluations (5 runs × 16 plans × 80 rounds) using the reward of Eq. (4), with no multiple-testing adjustment and no report of selection instability across runs. The top in-sample rewards are likely the most noise-inflated, so the reported 2023–2025 performance is a single draw from a selected set. A permutation test (e.g., comparing against shuffled rewards under the same selection protocol) or a seed-level variance analysis of the full mining pipeline is needed to establish that the improvement is not a selection artifact.
[Table 4 and §5 (Semantic Navigation)] The claim that the search 'reallocates evaluations toward high-reward semantic regions' is partly circular: the quota selector is explicitly designed to direct more budget to surrogate-ranked high-reward candidates, so observing increasing mass in high-reward manifolds is a property of the mechanism rather than an independent empirical finding. The shuffled-reward control in Appendix E is helpful for reward predictability, but the navigation analysis needs a null baseline (e.g., random selection or a static exploration schedule) to demonstrate that the learned surrogate, rather than the allocation rule itself, drives the observed concentration.
[Figure 6, Table 8, §5 (LLM Realization Robustness)] The conclusion that 'alpha mining quality is largely robust to the choice of LLM' is not supported with statistical comparisons. Table 8 shows mean |RankIC| ranging from 0.0116 (GPT-5.4) to 0.0168 (DeepSeek-V4-Flash), but no error bars or tests are given; the absence of a monotonic relationship cannot be established from point estimates. In addition, Pass@1 is measured before repair while final RankIC is computed after one repair attempt, so the comparison conflates one-shot success with post-repair quality. Please report per-model distributions and clearly separate the two quantities.

minor comments (6)

[Table 1] The 'Fund. Schema' column uses '7' and '3' symbols that are never defined in the text or caption. Please replace with 'No'/'Yes' or add an explicit legend.
[Eq. (2) and Eq. (4)] These two equations are identical in substance. Consider defining the reward once and referencing it later, to avoid confusion between the preliminary definition and the experimental instantiation.
[§4 Implementation Details] The 'fast and slow time scales' are mentioned but not defined. Appendix D gives example periods (20 and 100), but the general rule for selecting these scales should be stated explicitly and the formal definition aligned with Eq. (2).
[Figure 3] The two bars labelled '|Rank IC| drop' and 'Validity drop' share a percentage axis but measure very different quantities. Please use separate axes or a normalized display, and add error bars over the 100 sampled plans.
[Figure 6 and Table 8] The figures/tables report point estimates only. Since the analysis is over 100 plans, a simple bootstrap interval or cross-model paired test would be more informative than the single means shown.
[Appendix E (repeated realization)] The reference quality q_i^RIC is defined as a finite average over available realizations; the text acknowledges this is not a noiseless oracle. Consider stating explicitly how many of the 100 plans had fewer than eight successful realizations and whether results change under a coverage filter.

Circularity Check

2 steps flagged · score 2.0 of 10

Central benchmark claims are externally evaluated and not circular; two supporting analyses are mildly self-referential, but they do not drive the headline results.

other [Appendix E, 'Realization-Budget Efficiency Details' (Figure 5 / Table 5)]
"For plan i, let ai,j denote the absolute Rank IC of its j-th successful realization. We define an approximate reference quality by averaging all available successful realizations, qRIC_i = (1/m_i) Σ_j a_i,j, where m_i is the number of successful realizations for that plan. For each budget k ∈ {1, . . . ,5}, we repeatedly sample k realizations per plan, average them to estimate plan quality, and compare the induced plan ranking with the ranking from qRIC_i."

The reference quality qRIC_i is the average over all available realizations, including exactly the k sampled realizations used to form the estimate being ranked. Thus the reported improvement in Spearman correlation (0.445 to 0.919) partly measures the mechanical overlap between the estimate and its own reference, not independent evidence that extra realizations improve plan-quality estimation. The paper notes the reference is not a noiseless oracle, but it does not disclose this self-overlap. This affects only an internal budget-efficiency diagnostic, not the held-out benchmark claims.
self definitional [Section 5, 'Semantic Navigation During Search'; Section 3, 'Adaptive Quota-Based Plan Selection']
"The search process initially covers diverse semantic regions and later allocates more evaluations toward high-reward regions, while maintaining multiple active areas of exploration. ... Both exploration and surrogate-guided exploitation select plans from the sampled candidate pool: the former favors structural novelty, while the latter ranks candidates by r̂t(p)."

Allocating a growing share of evaluations to high-reward candidates is the designed behavior of the quota selector: exploitation ranks sampled plans by the LightGBM predicted reward, and mutation starts from high-reward evaluated plans. Therefore the observed trajectory toward 'reward-enriched regions,' with mean reward rising from 1.415 to 2.516, confirms the selector's own allocation rule rather than providing independent evidence about the semantic space. It is a mechanism trace; the non-circular support for learnable reward structure is the separate held-out predictability experiment with shuffled-reward control.

full rationale

The paper's central claims — factor-pool IC/ICIR and portfolio metrics on the 2023–2025 held-out period, plus CSI500 as an out-of-universe check — are genuinely external evaluations: factor selection uses only training/validation rewards, and the test window is not used for selection. The LightGBM reward surrogate is trained on accumulated plan–reward pairs and evaluated with a canonical-key group split and shuffled-reward control, which is a valid held-out protocol rather than circular fitting. The self-citations to prior work by overlapping authors (e.g., Autoalpha, MCTS alpha mining) appear only as background references in the related-work narrative and are not load-bearing for any derivation. The two flagged items are secondary analyses: the realization-budget efficiency comparison uses a reference that contains the sampled realizations, and the semantic-navigation observation merely traces the algorithm's own exploitation mechanism. Neither undermines the main external-benchmark comparisons, so the overall circularity is minor; the headline results are not reduced to their inputs by construction.

Assumptions & free parameters 5 free parameters · 5 assumptions · 1 invented entities

The paper's contribution is mostly engineering: it constructs an explicit discrete space of trading semantics and learns a surrogate over it. The honest ledger: five groups of hand-set free parameters (reward weights, exploration schedule, budget, correlation threshold, vocabulary composition), several domain assumptions about stationarity and noise that any factor-mining protocol needs, and one representational invented entity (the plan) with falsifiable backtest handles. No physical entities, no new conserved quantities.

free parameters (5)

Reward coefficients (α, β, λ) = (10, 1, 2)
Hand-set in Appendix A; they define plan reward r = 10·RankIC + 1·RankICIR − 2·Δlag and therefore shape the surrogate, the exploration/exploitation allocation, and final pool selection. Different weights would change the discovered pools.
Exploration schedule (ρmin, ρmax, τ) = not reported in text
Appendix A mentions the schedule but gives no experimental values; it controls how quickly the search shifts from exploration to surrogate-guided exploitation across the 80 rounds.
Budget and cold-start parameters = 10 pure-exploration rounds; 80 rounds × 16 plans; 5 runs
Chosen by hand; the number of evaluated plans (~6,400) defines the selection surface and the snooping risk flagged in this review.
Correlation filter threshold = 0.7
Eq. 5; controls redundancy of the exported 120/150-factor pools; no sensitivity analysis is provided.
Schema-vocabulary composition = 40 Events / 40 Contexts / 50 Qualities / 3 Directions / 7 Outputs
Hand-built from literature and practitioner sources (Appendix C); the space is the object of search, so its composition is a design choice, not a given.

assumptions (5)

domain assumption Zero-mean realization noise: r(k)(p) = q(p) + ε(k)(p) with E[ε|p] = 0, allowing single realizations per plan to be pooled at feature level.
Appendix E, 'Why Single-Realization Rewards Remain Predictable'; load-bearing for the surrogate's validity since each plan gets only one or two (fast/slow) realizations.
domain assumption In-sample (2016–2022) reward ranking transfers to the 2023–2025 test period without regime-driven inversion.
Implied by using the training/validation search reward for final pool selection; the only support is the test-period factor-decay diagnostic against Alpha158.
domain assumption The 5-day forward close-to-close return (Eq. 3) is the right objective, and Qlib Top50/Drop5 with ~0.2% friction is an adequate tradability model.
Appendix A; all portfolio-level claims depend on this protocol.
ad hoc to paper The heuristic look-ahead leakage screening catches all look-ahead in LLM-generated code.
Section 3, 'Plan Realization and Execution Validation'; the Appendix D example is causal, but one example does not validate the screening procedure against adversarial LLM outputs.
domain assumption LLM realizations of the same plan, once they pass execution validation, are exchangeable across backends.
Section 5 and Appendix E; the robustness conclusion rests on 7 backends × 100 plans with |RankIC| in 0.0116–0.0168 after repair.

invented entities (1)

Structured semantic plan p = (e, c, Q, d, o) independent evidence
purpose: Search object replacing factor code: a five-part description of a trading mechanism (event, context, qualities, direction, output) evaluated before code realization.
Not a physical entity but a representational abstraction. Its falsifiable handle is the realized factor's backtest reward under the paper's published realization and evaluation protocol; it gains independence only if a third party re-implements the vocabulary on new data.

how reviews work

0 comments

Cite this review

Pith. "Pith review of AlphaSchema: Exploring the Space of Trading Semantics for LLM-Based Alpha Mining." pith.science (2026). https://pith.science/paper/NGQGNQMX

@misc{pith2026260726642,
  author       = {Pith},
  title        = {Pith review of: AlphaSchema: Exploring the Space of Trading Semantics for LLM-Based Alpha Mining},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NGQGNQMX}},
  note         = {Machine review of arXiv:2607.26642}
}

read the original abstract

Automated alpha mining has increasingly adopted large language model (LLM) agents for factor generation and iterative discovery. However, existing LLM-based systems often delegate both factor construction and search decisions to the agent itself, without an explicit exploration space or a principled mechanism for navigating that space. As a result, exploration remains largely implicit and difficult to control or optimize systematically. We introduce AlphaSchema, which constructs and explores a structured space of trading semantics for alpha mining. Each point in this space is a schema plan composed of Event, Context, Qualities, Direction, and Output, specifying the semantics of a candidate factor before implementation. AlphaSchema decouples exploration from implementation: an LLM translates selected schema plans into executable factors, while evaluated rewards are accumulated to learn a surrogate model over the semantic space. An iterative selection mechanism uses this model to balance global exploration, surrogate-guided exploitation, and local mutation. Experiments on the Chinese stock market show that AlphaSchema discovers factor pools with strong predictive and portfolio performance. Further analyses show that the semantic search process navigates diverse regions while increasingly allocating evaluations toward high-reward regions, and that implementations of the same schema plans by different LLMs exhibit comparable predictive quality, suggesting that alpha mining quality is largely robust to the choice of LLM within our framework.

Figures

Figures reproduced from arXiv: 2607.26642 by the authors.

**Figure 1.** Overview of ALPHASCHEMA. ALPHASCHEMA performs iterative alpha mining over a structured trading-semantic space. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** Held-out NAV trajectories against representative [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Component-wise schema ablation. component leads to a degradation in factor quality, while the implementation success rate decreases only moderately. Here, successful implementation requires passing execution checks including data-contract compliance, numerical validity, and look-ahead leakage screening. The relatively small drop in validity suggests that the LLM can still construct executable factors even from inc… view at source ↗

Figures from the paper (5 more)

**Figure 5.** Figure 5: Plan quality estimation and top-plan discovery ef [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: Effect of LLM realization models on implementa [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

**Figure 7.** Figure 7: CSI500 held-out cumulative-return curves over the 2023–2025 test period. The factor pool is independently mined and [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]

**Figure 8.** Figure 8: Test-period factor-decay analysis. We plot the test [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

**Figure 9.** Figure 9: Reward predictability from structured schema fea [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 7 linked inside Pith

[1]

2022 IEEE 20th International Conference on Industrial Informatics (INDIN) , pages=

Genetic algorithm based quantitative factors construction , author=. 2022 IEEE 20th International Conference on Industrial Informatics (INDIN) , pages=. 2022 , organization=

2022
[2]

Advances in Neural Information Processing Systems , volume=

R&D-Agent-Quant: a multi-agent framework for data-centric factors and model joint optimization , author=. Advances in Neural Information Processing Systems , volume=
[3]

arXiv preprint arXiv:2602.07085 , year=

Quantaalpha: An evolutionary framework for llm-driven alpha mining , author=. arXiv preprint arXiv:2602.07085 , year=

arXiv
[4]

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

Cognitive alpha mining via llm-driven code-based evolution , author=. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=
[5]

arXiv preprint arXiv:2506.13131 , year=

Alphaevolve: A coding agent for scientific and algorithmic discovery , author=. arXiv preprint arXiv:2506.13131 , year=

arXiv
[6]

The journal of finance , volume=

Capital asset prices: A theory of market equilibrium under conditions of risk , author=. The journal of finance , volume=. 1964 , publisher=

1964
[7]

the Journal of Finance , volume=

The cross-section of expected stock returns , author=. the Journal of Finance , volume=. 1992 , publisher=

1992
[8]

Journal of financial economics , volume=

Common risk factors in the returns on stocks and bonds , author=. Journal of financial economics , volume=. 1993 , publisher=

1993

Show all 30 references

[9]

2017 , institution=

Replicating anomalies , author=. 2017 , institution=

2017
[10]

Wilmott , volume=

101 formulaic alphas , author=. Wilmott , volume=. 2016 , publisher=

2016
[11]

Statistics and computing , volume=

Genetic programming as a means for programming computers by natural selection , author=. Statistics and computing , volume=. 1994 , publisher=

1994
[12]

arXiv preprint arXiv:2002.08245 , year=

Autoalpha: an efficient hierarchical evolutionary algorithm for mining alpha factors in quantitative investment , author=. arXiv preprint arXiv:2002.08245 , year=

2002 arXiv
[13]

Proceedings of the 2021 International conference on management of data , pages=

Alphaevolve: A learning framework to discover novel alphas in quantitative investment , author=. Proceedings of the 2021 International conference on management of data , pages=

2021
[14]

Proceedings of the 29th ACM SIGKDD conference on knowledge discovery and data mining , pages=

Generating synergistic formulaic alpha collections via reinforcement learning , author=. Proceedings of the 29th ACM SIGKDD conference on knowledge discovery and data mining , pages=
[15]

Proceedings of the AAAI conference on artificial intelligence , volume=

Alphaforge: A framework to mine and dynamically combine formulaic alpha factors , author=. Proceedings of the AAAI conference on artificial intelligence , volume=
[16]

Findings of the Association for Computational Linguistics: ACL 2024 , pages=

Can large language models mine interpretable financial factors more effectively? a neural-symbolic factor mining agent model , author=. Findings of the Association for Computational Linguistics: ACL 2024 , pages=

2024
[17]

Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V

AlphaAgent: LLM-driven alpha mining with regularized exploration to counteract alpha decay , author=. Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2 , pages=
[18]

Proceedings of the AAAI Conference on Artificial Intelligence , volume=

Navigating the alpha jungle: An llm-powered mcts framework for formulaic alpha factor mining , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=
[19]

arXiv preprint arXiv:2603.16365 , year=

FactorEngine: A Program-level Knowledge-Infused Factor Mining Framework for Quantitative Investment , author=. arXiv preprint arXiv:2603.16365 , year=

arXiv
[20]

arXiv preprint arXiv:2604.09601 , year=

Hubble: An llm-driven agentic framework for safe, diverse, and reproducible alpha factor discovery , author=. arXiv preprint arXiv:2604.09601 , year=

arXiv
[21]

arXiv preprint arXiv:2509.25055 , year=

AlphaSAGE: Structure-Aware Alpha Mining via GFlowNets for Robust Exploration , author=. arXiv preprint arXiv:2509.25055 , year=

arXiv
[22]

arXiv preprint arXiv:2009.11189 , year=

Qlib: An ai-oriented quantitative investment platform , author=. arXiv preprint arXiv:2009.11189 , year=

2009 arXiv
[23]

Advances in neural information processing systems , volume=

Lightgbm: A highly efficient gradient boosting decision tree , author=. Advances in neural information processing systems , volume=
[24]

Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining , pages=

Xgboost: A scalable tree boosting system , author=. Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining , pages=
[25]

Advances in neural information processing systems , volume=

Tree of thoughts: Deliberate problem solving with large language models , author=. Advances in neural information processing systems , volume=
[26]

Nature , volume=

Mathematical discoveries from program search with large language models , author=. Nature , volume=. 2024 , publisher=

2024
[27]

arXiv preprint arXiv:2601.07348 , year=

Controlled self-evolution for algorithmic code optimization , author=. arXiv preprint arXiv:2601.07348 , year=
[28]

Advances in neural information processing systems , volume=

Attention is all you need , author=. Advances in neural information processing systems , volume=
[29]

Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP) , pages=

Learning phrase representations using RNN encoder--decoder for statistical machine translation , author=. Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP) , pages=

2014
[30]

Supervised sequence labelling with recurrent neural networks , pages=

Long short-term memory , author=. Supervised sequence labelling with recurrent neural networks , pages=. 2012 , publisher=

2012

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [1]

2022 IEEE 20th International Conference on Industrial Informatics (INDIN) , pages=

Genetic algorithm based quantitative factors construction , author=. 2022 IEEE 20th International Conference on Industrial Informatics (INDIN) , pages=. 2022 , organization=

2022

[2] [2]

Advances in Neural Information Processing Systems , volume=

R&D-Agent-Quant: a multi-agent framework for data-centric factors and model joint optimization , author=. Advances in Neural Information Processing Systems , volume=

[3] [3]

arXiv preprint arXiv:2602.07085 , year=

Quantaalpha: An evolutionary framework for llm-driven alpha mining , author=. arXiv preprint arXiv:2602.07085 , year=

arXiv

[4] [4]

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

Cognitive alpha mining via llm-driven code-based evolution , author=. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

[5] [5]

arXiv preprint arXiv:2506.13131 , year=

Alphaevolve: A coding agent for scientific and algorithmic discovery , author=. arXiv preprint arXiv:2506.13131 , year=

arXiv

[6] [6]

The journal of finance , volume=

Capital asset prices: A theory of market equilibrium under conditions of risk , author=. The journal of finance , volume=. 1964 , publisher=

1964

[7] [7]

the Journal of Finance , volume=

The cross-section of expected stock returns , author=. the Journal of Finance , volume=. 1992 , publisher=

1992

[8] [8]

Journal of financial economics , volume=

Common risk factors in the returns on stocks and bonds , author=. Journal of financial economics , volume=. 1993 , publisher=

1993