acceptodds
Under review as a conference paper at ICLR 2027

CosmosAlpha: Multi-Relational Graph Search with Pareto-Guided Exploration for LLM-Driven Alpha Mining

Abstract

Discovering predictive alpha factors from noisy financial data is a fundamental yet challenging problem in quantitative finance. LLM-driven formulaic alpha mining offers an attractive paradigm by combining interpretable symbolic expressions with iterative generation and refinement guided by quantitative feedback. However, existing generation-centric and tree-based search strategies primarily organize factors through generation lineage, leaving structural and statistical relationships across search branches underutilized. Consequently, information remains localized along parent–child trajectories, which can confine exploration to redundant regions. Moreover, effective factor discovery must account for multiple distinct objectives that cannot be adequately captured by a single search criterion. We introduce , a graph-based LLM factor-mining framework that represents discovered factors as a multi-relational graph encoding generation lineage, expression-structure similarity, and cross-sectional exposure correlation. A Graph Flow mechanism propagates search utilities across the graph, allowing each factor to be assessed using relational evidence from nearby factors rather than in isolation. Pareto-guided selection jointly considers quality, fertility and diversity, explicitly modeling their trade-offs within a multi-objective search framework. Across CSI 300 and CSI 500, CosmosAlpha-Graph achieves the best mean rank over seven predictive and portfolio metrics and outperforms its controlled CosmosAlpha-MCTS counterpart on every reported metric, showing a consistent advantage for relational graph search in LLM-driven factor mining.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
large language models, alpha factor mining, graph search, multi-objective optimization, quantitative finance agent

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Wn4uOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes CosmosAlpha, an LLM-driven formulaic alpha-mining framework that organizes previously evaluated factors as a multi-relational graph rather than only through generation lineage. The graph contains generation, expression-structure, and empirical exposure-correlation relations among complete factors. The method defines three search utilities, namely Quality, Fertility, and Diversity, propagates these utilities through the graph using Graph Flow, and then applies Pareto-guided parent selection before LLM-based factor mutation.

The method is evaluated on CSI 300 and CSI 500 using a common downstream LightGBM and portfolio evaluation pipeline. The main comparisons include handcrafted factors, genetic programming, several recent formulaic alpha-mining methods, multiple LLM-based systems, and a controlled CosmosAlpha-MCTS baseline. The paper also provides component and relation ablations, LLM-backbone sensitivity experiments, and several independent runs. The reported results show that CosmosAlpha-Graph achieves strong predictive and portfolio metrics relative to the considered baselines.

Critical strengths
  1. The paper proposes a clear and reasonably well-motivated search representation. Instead of preserving only generation lineage, CosmosAlpha connects complete evaluated factors using generation, expression-structure, and empirical exposure-correlation relations. The combination of Graph Flow with Pareto-guided selection over Quality, Fertility, and Diversity gives the paper a concrete algorithmic contribution rather than relying only on prompt engineering or a generic LLM search loop.

  2. The experimental section is relatively comprehensive. The authors compare against a broad set of formulaic alpha-mining and LLM-driven baselines, include a controlled CosmosAlpha-MCTS comparison, provide relation- and utility-level ablations, test multiple LLM backbones, and report independent runs. In particular, the ablation study is useful because removing Graph Flow, individual objectives, or individual relation types consistently degrades the reported metrics.

Critical weaknesses
  1. I am not yet sufficiently convinced that the reported improvement is robust to overfitting or adaptive search effects. The paper uses a chronological train/validation/test protocol, which is a meaningful strength, and the test period is held out from factor-expression search. However, formulaic alpha mining evaluates a large number of candidate hypotheses and search decisions, while the headline evidence is still based primarily on one temporal split and two closely related A-share universes. This leaves open the possibility that some of the reported advantage is specific to the chosen market regime, temporal partition, or experimental design. I would like to see additional robustness evidence, such as alternative chronological cutoffs, rolling or walk-forward evaluation, subperiod results, or another genuinely untouched market/period. This is important because the central claim is not simply that the search procedure fits the training period well, but that it discovers factors with robust out-of-sample predictive value.

  2. The economic evaluation is incomplete without a clear transaction-cost sensitivity analysis. The paper reports substantial differences in ARR, IR, and MDD, but these metrics alone do not establish that the advantage remains economically meaningful after trading frictions. This matters especially for automatically mined factors because different factor pools may induce very different turnover. I would like the authors to report turnover and net performance under reasonable transaction-cost and slippage assumptions, preferably across several cost levels. If CosmosAlpha retains its relative advantage after such costs, this would significantly strengthen the empirical contribution.

As for writing, the paper is generally understandable and the high-level method is reasonably well organized. However, I found several presentation issues that noticeably reduce readability:

Most importantly, Figure 1 is visually poor in the submitted PDF. The text and graphical elements appear heavily rasterized/compressed, and some details are difficult to read at normal zoom. I have no objection to the use of AI-assisted figure generation or design tools, but the final paper should provide a publication-quality figure, preferably as vector graphics, with text and diagram elements that remain sharp and readable in the PDF.

The layouts of Tables 1 and 2 could also be improved. The large number of metrics, narrow columns, and multi-line headers make the tables visually crowded and harder to scan than necessary. Reformatting the predictive and portfolio metric groups, improving spacing, and reducing header clutter would substantially improve presentation quality.

Questions for authors
  1. Could the authors provide stronger evidence that the reported improvement is not specific to the current temporal split or the particular A-share evaluation setting?

I appreciate that the 2021--2025 test period is held out from factor search. However, given the highly adaptive nature of formulaic alpha mining, a single chronological split does not completely resolve my concern about search overfitting or regime-specific discovery. Could the authors report at least one additional robustness experiment, for example:

  • alternative chronological train/test cutoffs,
  • rolling or walk-forward evaluation,
  • performance broken down by test subperiod or market regime,
  • or evaluation on an additional genuinely untouched market or period?

I would also find it useful to see whether the relative ranking between CosmosAlpha and the strongest baselines remains stable across these splits. Convincing additional evidence here would materially improve my assessment.

  1. How does the portfolio advantage change after realistic trading costs?

Please report the turnover of the resulting strategies and net portfolio performance after commissions/slippage. A simple sensitivity analysis using several reasonable one-way cost assumptions, for example 5, 10, and 20 bps, would already be informative. The important question for me is not the exact cost model, but whether CosmosAlpha's improvement in ARR and IR remains after accounting for differential turnover across methods.

A convincing transaction-cost robustness result would materially increase my confidence in the practical significance of the reported portfolio improvements.

  1. Could the authors improve the camera-ready presentation of Figure 1 and Tables 1--2?

Figure 1 is difficult to read in the current PDF because the visual appears strongly rasterized/compressed. Please consider regenerating it as a vector figure with readable text at normal zoom. Tables 1 and 2 would also benefit from cleaner spacing and layout, as the current multi-level headers and dense metric columns make them unnecessarily difficult to scan. This is primarily a presentation concern rather than a technical one.

3 comments
AuthorsOpenReview ↗

Response to W1–W3/Q1–Q3: Rolling Robustness, Transaction Costs, and Presentation

Response to Reviewer 3

Rolling out-of-sample robustness (W1; Q1)

Thank you. We added a five-window CSI 300 rolling evaluation against Alpha158, MCTS and Chain. Both factor mining and LightGBM training roll forward annually. The first window trains on April 2010–May 2019, validates on June 2019–December 2020 and tests on 2021. All three segments then move forward one year, through the 2025 test year. In each window, Graph, MCTS and Chain mine factors using only that window's training segment, which also trains LightGBM; expressions are then fixed for validation and testing. Alpha158 retains its predefined expressions. LightGBM hyperparameters are shared and fixed; validation selects the boosting iteration.

These supplemental rolling results are reported separately from the original Table 1 runs.

Annual test RankIC. Bold indicates the highest value within each year.

Test yearAlpha158ChainMCTSGraph
20210.05600.05150.05620.0573
20220.05430.04370.04870.0562
20230.04610.02580.03230.0481
20240.04650.02840.03780.0445
20250.05460.05080.05720.0611

Annual test ARR (%), before transaction costs.

Test yearAlpha158ChainMCTSGraph
20219.4910.4811.479.56
2022-11.83-12.65-12.57-8.49
2023-0.02-6.66-8.16-2.61
202420.5621.8020.7623.84
202511.0213.6015.2116.66

Graph's RankIC exceeds MCTS and Chain in all five years and Alpha158 in four. Its ARR exceeds each baseline in four of five years. Graph ranks first in RankIC in four years and ARR in three. These results support the temporal robustness of both factor mining and downstream integration across different chronological cutoffs.

Turnover and transaction-cost sensitivity (W2; Q2)

Thank you. We replayed saved-model CSI 300 predictions over 2021–2025, separately from the rolling study above. We charge symmetric proportional costs of 0/5/10/20 bps on both buys and sells as combined commission/slippage scenarios. Predictions and strategy settings are fixed; portfolios are replayed at each cost level.

Daily turnover is total buy-plus-sell notional divided by previous account value; annualized turnover is its daily mean times 252. ARR and IR use the common Qlib annualization; signed MDD is drawdown of cumulative summed daily returns. The uncharged CSI 300 price-index reference covers the same 1,212 dates.

Factor setCost/side (bps)ARR (%)IRMDD (%)Daily turnover (%)Annual turnover (x)
Graph08.440.5332-25.0719.1148.15
Graph56.280.3897-26.5419.0948.10
Graph104.010.2485-30.8819.0748.06
Graph20-0.53-0.0330-46.0019.0347.96
MCTS05.930.3546-26.5019.9650.30
MCTS53.550.2126-32.7519.9450.26
MCTS101.180.0707-40.1219.9350.22
MCTS20-3.55-0.2124-57.3519.9050.15
Alpha15806.320.3945-29.7319.9350.23
Alpha15853.970.2480-32.5819.9150.16
Alpha158101.620.1012-37.6719.8850.10
Alpha15820-3.07-0.1919-53.6819.8249.95
Chain03.490.1839-39.2719.3148.67
Chain51.180.0623-44.5419.3048.65
Chain10-1.12-0.0590-50.6519.2948.61
Chain20-5.70-0.3002-67.8819.2648.54
CSI 300 index—-0.78-0.0445-56.00——

Graph has the lowest turnover and leads all factor baselines on ARR, IR and MDD at every cost level. At 10 bps, its net ARR is 4.01% (IR 0.2485). At 20 bps, all four net ARRs are negative. Even at this highest tested cost, Graph still outperforms the uncharged CSI 300 index in ARR (-0.53% vs. -0.78%), IR (-0.0330 vs. -0.0445) and MDD (-46.00% vs. -56.00%).

Presentation (W3; Q3)

We will provide a vector version of Figure 1 and improve the spacing, headers and metric grouping in Tables 1–2 in the revised manuscript.

Thanks for the additional experiments. The new ablations and cost-matched comparison address several of my earlier concerns, particularly regarding component contributions and evaluation fairness. The development-quota experiment is also interesting, as more feedback improves dev IC but hurts test IC.

My main remaining concerns are the lack of repeated runs for the ablations and the limited evidence for scaling efficiency under a fixed compute budget. I also think the paper could be more careful in distinguishing positive test IC from statistically significant discoveries. Overall, the rebuttal has improved my assessment

AuthorsOpenReview ↗

Thank You for Your Constructive Follow-up

Dear Reviewer,

Thank you for your thoughtful follow-up and for taking the time to carefully reconsider our work in light of the additional experiments. We sincerely appreciate your constructive feedback and your recognition of the improvements made during the rebuttal.

We acknowledge your remaining concerns regarding the reproducibility of the ablation results and the limited evidence for scaling efficiency across different compute budgets. These are important considerations for assessing the robustness and generalizability of the proposed search framework. We will clarify the scope and limitations of our current experimental evidence in the revised manuscript.

We also appreciate your suggestion to distinguish positive out-of-sample IC from statistically significant factor discoveries. We agree that this distinction is particularly important in adaptive factor mining, and we will refine the relevant statements to ensure that our conclusions accurately reflect the strength of the statistical evidence.

Thank you again for your valuable feedback and constructive suggestions. They have been very helpful in improving the rigor and clarity of our work.

Reviewer_VcYxOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3(+1)
Soundness
2
Read review
Paper summary

CosmosAlpha organizes evaluated formulas into a graph with lineage, structural-similarity, and exposure-correlation edges. It propagates quality, fertility, and diversity scores, then selects parents through Pareto ranking for LLM mutation. The resulting factors are combined with LightGBM. Experiments on CSI300 and CSI500 compare this search with recent factor miners, an MCTS variant, and component ablations.

Critical strengths
  1. I find the use of exposure correlation to guide search well motivated. Formulas that look different can still produce nearly identical signals, so expression structure alone is an incomplete guide to where further exploration is worthwhile. Bringing both relationships into parent selection gives the graph a useful role in the mining process.

  2. The experiments make a reasonable effort to test what drives the gains. Removing propagation or either similarity relation hurts performance, and the advantage over the MCTS variant persists across the tested LLMs and repeated runs. This gives some confidence that the design is useful, even though the comparison of search costs remains unresolved.

Critical weaknesses
  1. I am not convinced that the search comparison is fair. The methods are matched by the number of factors they eventually produce, but reaching that number may require very different amounts of generation, repair, and evaluation. Sharing the factor language and downstream model does not resolve this. The paper reports LLM usage for Graph without equivalent accounting for the competing methods, so it is hard to tell how much of the improvement comes from better search and how much might come from greater search effort. A comparison under the same evaluation or token budget would make the central claim much more convincing.

  2. The overall ranking overstates how clearly the method improves on strong baselines. Several predictive results are very close to the best alternatives, including the fixed Alpha158 library. Averaging ranks obscures the size of those differences and gives separate credit for closely related metrics. The repeated runs are useful, but cover the internal search variants on CSI300, leaving uncertainty about the main comparisons. I would put more weight on the portfolio gains if the closest competitors were also repeated and the test-period differences were assessed with temporal dependence taken into account. At present, the evidence for beating the internal MCTS variant is stronger than the evidence for a broader advance.

  3. The main figure is totally AI-generated.

Questions for authors

See weaknesses.

3 comments
AuthorsOpenReview ↗

Response to W1–W3: Search Fairness, Statistical Robustness, and Presentation

Response to Reviewer 2

Search cost and matched Token budgets (W1)

Thank you. Matching approximately 150 evaluable factors does not equalize generation, repair or evaluation costs. That target balanced mining cost and model input size near Alpha158. We added a CSI 300 comparison of Graph, MCTS and Chain with a common 4,000,000 input-plus-output Token cap, common initial factors, DeepSeek-V4-Flash, data split, validation/repair and LightGBM settings.

The cap covers all requests, including retries, failures, repair and Feedback. Missing usage is conservatively charged at the reserved upper bound. Search stops when the next request cannot fit; final factor counts are unconstrained. These new runs are separate from the original Table 1 runs.

CSI 300 test performance. ARR and MDD are percentages.

MethodICICIRRankICRankICIRARRIRSigned MDD
Graph0.03410.38170.05270.60138.290.5823-23.92
MCTS0.02930.33990.04440.53486.130.4494-25.20
Chain0.02140.23010.03460.39943.080.1709-34.09

Search and LLM accounting. Expansion cycles exclude initialization; evaluations include it. Counts at different workflow stages can overlap.

MeasureGraphMCTSChain
Evaluated factors: accepted / rejected176: 136 / 40225: 183 / 4295: 56 / 39
Expansion cycles218260113
Evaluations: events / unique factors176 / 176226 / 22596 / 95
Unsuccessful43 recorded attempts + 4 pre-record cycles34 attempt keys14 no-row cycles + 5 failed factors
Duplicate proposals / evaluations1 / 04 / 11 / 1
Total LLM calls683800468
Repair calls1032
Reported input Tokens1,993,9992,282,4572,734,050
Reported output Tokens1,685,1361,484,3411,224,876
Reported total3,679,1353,766,7983,958,926
Missing-usage requests / charge6 / 272,0004 / 184,0000 / 0
Charged total3,951,1353,950,7983,958,926
Remaining budget48,86549,20241,074

At the common Token cap, Graph leads both alternatives on all seven test metrics, with fewer evaluations and LLM calls than MCTS. This single-run comparison supports its advantage under a matched search budget.

Effect sizes, temporal dependence and repeated runs (W2)

We agree that Mean Rank is descriptive: it obscures effect sizes, and correlated metrics are not independent evidence. We therefore added a five-window CSI 300 rolling study, testing on 2021–2025. In each window, Graph, MCTS and Chain rerun LLM factor mining using only that window's training segment, which also trains LightGBM; expressions are then fixed for validation and testing. Alpha158 retains its predefined expressions. LightGBM hyperparameters are shared and fixed; validation selects the boosting iteration. These supplemental results are separate from the original Table 1 runs.

Graph minus baseline with 95% stationary-bootstrap confidence intervals. ARR differences are percentage points.

BaselineRankIC difference [estimated 95% CI]ARR difference [estimated 95% CI]
Alpha158+0.00194 [-0.00069, 0.00458]+1.95 [-1.71, 5.66]
MCTS+0.00700 [0.00447, 0.00953]+2.45 [-1.96, 6.99]
Chain+0.01340 [0.00987, 0.01685]+2.48 [-2.68, 7.59]

We use 5,000 paired stationary-bootstrap replicates within each test year, with mean block length 20 trading days. The same sampled date blocks are used across methods, preserving pairing and local temporal dependence. Intervals are the 2.5th and 97.5th percentiles of the bootstrap estimates. Mean block lengths 10 and 40 give the same qualitative conclusions.

Graph's annual RankIC exceeds Alpha158 in four years and MCTS and Chain in all five; its ARR exceeds each baseline in four years. In confidence analysis, RankIC intervals against MCTS and Chain exclude zero. Intervals including zero reflect uncertainty, but their upper bounds are about 2.8–6.6 times the absolute lower bounds, placing zero in their lower third and directionally favoring Graph.

Figure presentation (W3)

We will provide a clear vector version of the main figure in the revised manuscript.

Thanks for your careful response, and the main figure looks much better now. I will update my score accordingly.

AuthorsOpenReview ↗

Thank You for Your Feedback

Dear Reviewer,

Thank you for your positive feedback and thoughtful reconsideration. We sincerely appreciate your constructive suggestions, which have helped us improve the manuscript.

Reviewer_p7rkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CosmosAlpha searches for formulaic alpha factors using an LLM and a graph of previously evaluated candidates. The graph contains three types of relations: generation lineage, expression-structure similarity, and cross-sectional exposure correlation. A Graph Flow step propagates factor quality, fertility, and diversity through these relations, and Pareto sorting selects parents for further LLM-guided mutation. Discovered factor pools are evaluated through a common LightGBM and backtesting pipeline on CSI 300 and CSI 500. The Graph variant obtains the best mean rank over seven reported metrics in both universes and outperforms the authors' MCTS variant on every reported metric. The paper also includes component ablations, comparisons with three LLM backbones, and a three-run analysis on CSI 300.

Critical strengths

S1. Search across generation branches. By connecting evaluated factors through lineage, structural similarity, and exposure correlation, CosmosAlpha can incorporate evidence from related factors outside a node's ancestry. The distinction between the intrinsic quality archive and the propagated utilities used for parent selection is clearly specified (Section 3; Appendix A). This is a coherent design for avoiding redundant local search. S2. Competitive results across several evaluations. The Graph variant has the best mean rank on both CSI 300 and CSI 500 and exceeds the authors' MCTS variant on all seven reported metrics (Tables 1–2). The ablations favor the full model, and the three independent CSI 300 runs consistently favor Graph over MCTS and Chain (Table 20). These results support the usefulness of the overall pipeline.

Critical weaknesses

W1. The Graph–MCTS comparison does not fully isolate the contribution of relational search. Section 4.2 treats this comparison as evidence for the benefit of changing the search organization. Yet Graph and MCTS also use different factor-design prompts and feedback. Graph receives archive-relative gaps and correlation-neighborhood guidance, whereas MCTS uses target-dimension guidance and different refinement instructions (Appendices A.6 and D.3–D.4; Table 11). These choices may be necessary for the respective systems, but they make it difficult to assign the observed improvement specifically to the multi-relational graph and Graph Flow. The methods are also matched on approximately 150 evaluable output factors, rather than on actual expansion attempts, LLM calls, or token consumption (Appendix C.2). Thus, the reported results establish an advantage for the complete Graph pipeline under the chosen protocol, while the independent contribution of relational propagation and selection remains less clear. A targeted control with more comparable proposal conditions and resource use, or a narrower claim supported by detailed budget accounting, would address this concern. W2. Clipped normalization can obscure improvements to the raw quality frontier in the fertility reward. The Quality Pareto Archive uses the raw three-objective quality vector, but SelfQuality and hypervolume gains use fixed, clipped normalized coordinates (Appendix A.4, Eqs. 11–16). These representations can induce different dominance relations. In Table 17, E1 and E3 are non-dominated in raw quality space: E1 has higher absolute RankIC and direction-aligned Top-20 return, while E3 has lower RankIC variation. After the specified normalization, their coordinates are approximately (1, 0.4043, 1) and (1, 0.4355, 1), respectively; E3 therefore dominates E1 in the hypervolume space. A factor that extends the raw archive can consequently contribute no additional normalized hypervolume if its clipped point is already covered. This matters because hypervolume gain defines expansion outcomes and hence the Fertility signal used to select subsequent parents. Clipping may be an intentional way to limit extreme values, but the paper does not report how often frontier improvements are lost in the reward or how sensitive the search is to the normalization anchors. This example establishes a discrepancy between the two quality representations, not a demonstrated error in the actual birth-time reward or a measured loss of final performance.

Questions for authors

Q1. Search attribution and budget. Can the authors compare Graph and MCTS under matched initial factors and LLM-call or token budgets, with proposal prompts and accessible feedback made as comparable as the algorithms allow? Please provide the actual admitted pool sizes, expansion attempts, evaluated candidates, unsuccessful/duplicate attempts, repair calls, total LLM calls, and token counts for the main comparisons. If the prompts cannot be matched, which existing or targeted controls distinguish graph-based propagation and selection from the effects of proposal guidance? Please also specify the MCTS selection and backup rules. Finally, does the 185-node graph shown in Figure 3 belong to the Table 1 run, or to a separate diagnostic run? Q2. Normalization and fertility. What fraction of admitted factors reaches the clipping bounds in Eqs. 11–13, and how often does an extension of the raw Quality Pareto Archive yield zero normalized hypervolume gain? Please explain the choice of normalization anchors and report whether the search trajectory or final factor-pool performance changes under reasonable alternative anchors or a less-saturating normalization. Q3. Fertility in Figure 2. The figure describes its points as newly generated factors with nonzero Fertility, while Eq. 16 assigns zero raw Fertility to a node without prior expansion attempts. Does the panel show the parent's previous Fertility, the child's birth-time hypervolume gain, a propagated utility, or a retrospective quantity? Please identify the plotted variable and confirm that parent selection uses Eq. 16. Q4. Relation ablations. Appendix H indicates that removing the Structure or Correlation relation changes both Graph Flow and the AverageDistance diversity metric. Can the authors separate these two effects, for example by holding the diversity computation fixed while removing the corresponding propagation edges?

4 comments
AuthorsOpenReview ↗

Response to W1/Q1: Matched Generation Conditions and Token Budgets

Response to Reviewer 1 - Part 1

Matched generation conditions and Token budgets (W1 / Q1)

Thank you. The approximately 150-factor target balanced mining cost and feature scale near Alpha158, without matching search effort.

Estimated cost at the paper's 150-factor scale. Input+output Tokens: Graph 3.98M, MCTS 3.24M and Chain 7.84M, including conservative missing-usage charges.

Matched control. All methods start from the same two factors:

mul(mul(sub(close, low), inv(sub(high, low))), ts_rank(volume, 20))
rank(mul(sub(div(close, vwap), 1), div(turn, div(ts_sum(turn, 20), 20))))

Shared FactorDesign prompt (all three methods).

SYSTEM
You are the CosmosAlpha Factor Design Agent. Produce exactly one concise, testable daily equity-factor design. Use only supplied raw fields, registered operators and supported windows. Do not write DSL or executable code. When a parent expression is supplied, design a useful refinement grounded in the hypothesis and the parent raw quality metrics. Consider predictive strength, stability and top-20 return together. Return only the exact JSON object requested.

USER
Factor design request JSON:
${request_json}

Use `request_json.universe` as the research stock-pool scope. Do not replace it with another universe or use current or future constituent information.

Runtime capability view JSON:
${prompt_view_json}

Return a designs array containing exactly one item. Each design must contain exactly: name, description, mechanism_mapping, signal_definition, variables (objects with name and role), formulation, window_choices (objects with window and rationale), expected_direction, failure_conditions. Every variables[].name must exactly match one field name in Runtime capability view JSON.catalog.fields; use only its raw catalog name and do not list intermediate variables. variables[].role should describe the field's input role. Intermediate quantities belong only in formulation or signal_definition. Every window_choices[].window must be one of the supported_windows shown by the catalog. formulation must be mathematical or natural-language construction guidance, not DSL.

Graph/MCTS share Hypothesis/Design fields: parent expression, raw RankIC mean/std and aligned Top-20 return, without architecture-specific guidance. Chain retains history/Feedback, with its cost counted.

All use DeepSeek-V4-Flash (temperature 0), common stage limits, data split, validation/repair and LightGBM settings.

Budget. The 4,000,000 input+output Token cap covers failures, retries, repair and Feedback. Missing usage is charged at its reserved upper bound. Search stops when the next request cannot fit. These accounts cover the new experiment.

CSI 300 test performance. ARR/MDD in %.

MethodICICIRRankICRankICIRARRIRSigned MDD
Graph-G00.03410.38170.05270.60138.290.5823-23.92
MCTS-G00.02930.33990.04440.53486.130.4494-25.20
Chain-native0.02140.23010.03460.39943.080.1709-34.09

Search accounting. Expansion cycles exclude seeds; evaluations include them.

MeasureGraph-G0MCTS-G0Chain-native
Structure / CSV173 graph nodes; archive 18227 node occurrences (220 unique); AlphaZoo 18100 rows
Evaluated pool: accepted / rejected176: 136 / 40225: 183 / 4295: 56 / 39
Expansion cycles / controller steps218 / 214 attempts260 / 102 iterations113 / 113 serial cycles
Evaluations: events / unique176 / 176226 / 22596 / 95
Unsuccessful43 attempts (including 2 duplicates) + 4 pre-record cycles34 attempt keys14 no-row cycles + 5 failed factors
Duplicate proposals / evaluations1 / 04 / 11 / 1

Graph omits 3 low-date-coverage factors: 176−3=173176-3=173. MCTS omits 5 and retains 7 repeat-expression occurrences: 225−5+7=227225-5+7=227. Its 2 failed rows create no nodes; Chain excludes 5 failed rows. Counts can overlap across stages.

LLM and Token accounting. Calls include retries and repair. Charged total = reported total + missing-usage charge.

MeasureGraph-G0MCTS-G0Chain-native
Calls: success / failure / parse failure683: 654 / 20 / 9800: 780 / 6 / 14468: 455 / 12 / 1
Repair calls1032
Reported input1,993,9992,282,4572,734,050
Reported output1,685,1361,484,3411,224,876
Reported total3,679,1353,766,7983,958,926
Missing-usage requests / charge6 / 272,0004 / 184,0000 / 0
Charged total3,951,1353,950,7983,958,926
Remaining budget48,86549,20241,074

In this single run under matched generation conditions and a common Token cap, Graph leads MCTS on all seven metrics with fewer evaluations and calls.

AuthorsOpenReview ↗

Response to Q1: MCTS Selection, Backup, and Figure 3 Clarification

Response to Reviewer 1 - Part 2

MCTS selection and backup rules (Q1)

Thank you. Our controller adapts UCT, virtual expansion and maximum-reward backup from AlphaJungle (Shi et al., AAAI 2026), with the following scheduling and scoring rules.

Reward. Let Z−Z^- be the pre-admission union of the current-tree AlphaZoo and shared AlphaZoo. The pool is internally deduplicated by factor ID; an existing entry with candidate xx's ID is not excluded. Let μ\mu and II be mean daily RankIC and RankICIR. For Z−≠∅Z^-\ne\varnothing, at birth dx=1−max⁡z∈Z−∣ρ(x,z)∣d_x=1-\max_{z\in Z^-}|\rho(x,z)| and

pm(x;Z−)=1∣Z−∣∑z∈Z−1[m(z)≤m(x)],r(x)=0.5pμ+0.25pI+0.25pd.p_m(x;Z^-)=\frac{1}{|Z^-|}\sum_{z\in Z^-}\mathbf{1}[m(z)\le m(x)],\qquad r(x)=0.5p_\mu+0.25p_I+0.25p_d.

Reference entries use stored birth metrics; r(x)r(x) and dxd_x remain frozen. For an empty reference pool (Z−=∅Z^-=\varnothing), dx=1d_x=1 and r(x)=0.5g(μx)+0.25clip⁡(Ix,0,1)+0.25r(x)=0.5g(\mu_x)+0.25\operatorname{clip}(I_x,0,1)+0.25, where g(μ)=0g(\mu)=0 for μ≤0\mu\le0 and g(μ)=[1+exp⁡(−80(μ−0.025))]−1g(\mu)=[1+\exp(-80(\mu-0.025))]^{-1} otherwise.

Correlation is absolute pooled Pearson in the inspected legacy source, and the absolute value of mean valid daily cross-sectional Pearson in 4M.

UCT. Every parent selection in the 4M run used UCT. Start at the root: return a leaf, or choose an unvisited child uniformly. When all children have been visited, compare

U(s′)=Bs′+clog⁡NsNs′.U(s')=B_{s'}+c\sqrt{\frac{\log N_s}{N_{s'}}}.

V(s)=Bs+clog⁡Ns/(1+∣C(s)∣)1+κs+0.5c(us/3)1+0.3hs.V(s)=\frac{B_s+c\sqrt{\log N_s/(1+|C(s)|)}}{1+\kappa_s}+\frac{0.5c(u_s/3)}{1+0.3h_s}.

Here BB is the maximum backed-up reward, NN visits, CC direct children, κ\kappa cooling (initially zero), hh depth, uu the number of effectiveness/stability/diversity action labels absent among C(s)C(s), and c=1.6c=1.6 for multi-tree search. Expand ss if V(s)>max⁡U(s′)V(s)>\max U(s'); otherwise descend to the maximizing child, breaking child ties by insertion order.

After each child, increment κs\kappa_s if r(x)<0.7r(s)r(x)<0.7r(s), reset it if r(x)>1.05r(s)r(x)>1.05r(s), and otherwise retain it. UCT expansion allows up to two attempts, or three when ∣μs∣>0.04|\mu_s|>0.04.

Each attempt samples a label with P(j∣s)∝exp⁡(−aj)P(j\mid s)\propto\exp(-a_j), where a=(g(μs),clip⁡(Is,0,1),ds)a=(g(\mu_s),\operatorname{clip}(I_s,0,1),d_s). G0 retains these labels for coverage but omits dimension-specific generation instructions.

Backup. Initialize the new child with Nx=0N_x=0 and Bx=r(x)B_x=r(x). Immediately after insertion, visit each node from xx to the root exactly once: increment NvN_v and set Bv←max⁡(Bv,r(x))B_v\leftarrow\max(B_v,r(x)). Thus NxN_x becomes one, and each ancestor is incremented once. No-child attempts update neither visits, backup nor cooling; their token costs remain charged.

We will document these rules, Zoo updates and tree budgets in the revised appendix.

Source of the 185 nodes in Figure 3 (Q1)

Figure 3 (Appendix G) comes from a separate qualitative diagnostic run, rather than the run reported in Table 1. It illustrates the complete generation lineage and terminal search Pareto fronts on CSI 300. Its 185 nodes comprise 2 initial factors and 183 subsequently generated, unique, technically admissible factors, including those below the publication-quality threshold. Thus, 185 is the full graph size, rather than the accepted-factor count or a Pareto-front size. This diagnostic run exhausted 100 expansion attempts, each allowing up to two children. The nominal target of approximately 150 evaluable factors in Appendix C.2 was not enforced as a hard node-count cap for this run.

AuthorsOpenReview ↗

Response to W2/Q2/Q3: HV Normalization and Fertility Diagnostics

Response to Reviewer 1 - Part 3

Clipped normalization and the Quality Pareto Archive (W2 / Q2)

Thank you for raising this question. The Quality Pareto Archive applies dominance to the unclipped vector of absolute mean RankIC, negative RankIC standard deviation, and direction-aligned Top-20 return. SelfQuality and HV use the normalized coordinates in Eqs. 11–13.

Why normalize and clip? Normalization puts the three metrics on common dimensionless scales for SelfQuality and our HV calculation in the unit cube. The saturation thresholds are empirical reference levels for strong single-factor performance. Beyond these levels, further training-metric gains can have diminishing downstream value and increase overfitting risk if pursued excessively. Capping their reward encourages balanced improvement across the three dimensions, while the raw archive preserves improvements beyond the caps.

Clipping frequency in the three appendix runs. Across the three Graph runs reported in the appendix, the denominator is 467 factors:

Coordinate at 0 or 1Count / nodesPercentage
Absolute RankIC23/4674.93%
Reversed RankIC standard deviation14/4673.00%
Aligned Top-20 return56/46711.99%
At least one coordinate70/46714.99%

The remaining 397/467 factors (85.01%) lie strictly inside all three intervals and retain their raw Pareto dominance comparisons. Boundary contact alone does not imply a changed comparison. Your E1/E3 example correctly illustrates the saturation trade-off; raw archive membership remains determined by the original metrics.

Zero-HV gains and alternatives. In these same three runs, 106 new factors improved the raw Pareto frontier. Their birth gains were evaluated independently against the frozen pre-round archive. Under the original normalization, 8/106 (7.55%) received zero HV gain; the other 98/106 (92.45%) received positive credit.

We evaluated the same 106 improvements with wider IC/Std/return intervals: [0, 0.075]/[0.04, 0.10]/[0, 0.0012] for 1.5 times the original widths, and [0, 0.10]/[0.03, 0.11]/[0, 0.0016] for twice the widths. We also used a strictly monotone smooth map of the original affine coordinate zz, reversing it for Std:

T(z)=12+arctan⁡(π(z−1/2))π.T(z)=\frac{1}{2}+\frac{\arctan(\pi(z-1/2))}{\pi}.

NormalizationZero HV among 106 raw improvementsPercentage
Original clipping8/1067.55%
Wider intervals, 1.5 times3/1062.83%
Wider intervals, 2 times0/1060.00%
Smooth atan map0/1060.00%

Fresh search with wider intervals. Normalization can affect rewards and subsequent parent selection. We therefore reran the full search with the twice-wider intervals above and evaluated the resulting factor pool with LightGBM on the CSI 300 test set:

MetricOriginal Graph (Table 1)Wider intervals, 2 times
IC0.03240.0332
ICIR0.34410.3313
RankIC0.04940.0512
RankICIR0.56700.5538
ARR8.44%7.86%
IR0.53320.5254
Signed MDD-25.07%-26.12%

Performance is comparable to the original setting and falls within or close to the ranges observed across the three Graph runs in the appendix's Independent-Run Stability Study. We will consider wider bounds and smooth normalization in future refinements.

Meaning of Fertility in Figure 2 (Q3)

Thank you for noting the ambiguity. The three diagnostics describe complementary aspects of factor search: Quality combines predictive strength, stability and aligned return; Fertility summarizes the ability to generate frontier-improving children from past expansion gains; AverageDistance measures structural and correlation novelty relative to high-quality factors. Each can be computed from either method's factor evaluations and expansion history.

Figure 2 reports Flow-propagated versions of all three metrics, sampled for each new factor after its generation batch. A newborn factor has zero raw Fertility under Eq. 16, but can receive positive propagated Fertility from neighboring factors.

For MCTS, we replay its recorded expansions in order, construct generation, structure and correlation relations from the observed lineage, expressions and aligned factor values, and apply the same metric definitions and Flow settings (γ=0.75\gamma=0.75, equal relation weights). Birth-time HV gains use the frozen pre-event archive; each point shows the factor's own post-batch propagated coordinates. These offline diagnostics never enter MCTS selection or backup.

In Graph search, Eq. 16 supplies raw Fertility; parent selection uses the propagated utilities. We will clarify the axes, caption and text, including carried-forward values at steps without a node coordinate.

AuthorsOpenReview ↗

Response to Q4: Separating Relation Propagation from Diversity

Response to Reviewer 1 - Part 4

Separating relation propagation from diversity (Q4)

Thank you for pointing out this confound. The original Structure/Correlation ablations removed the named relation from both Graph Flow and AverageDistance. We added four controls to separate these roles:

  • Flow only: remove the named relation from propagation, retaining the full AverageDistance definition.
  • Distance only: remove the named relation's contribution to AverageDistance, retaining the full Flow operator and its weights.

Other settings follow the common CSI 300 ablation protocol. For each relation, Full, Flow-only removal, Distance-only removal and joint removal form a 2-by-2 comparison. The table reports LightGBM test metrics; ARR and signed MDD are percentages. Original rows come from the paper's ablation table.

Variant (removed relation / role)ICICIRRankICRankICIRARRIRSigned MDD
Full Graph (original)0.03240.34410.04940.56708.440.5332-25.07
Structure: Flow only (new)0.03030.32510.04640.51885.120.3145-29.01
Structure: Distance only (new)0.03110.33010.04690.53195.490.3283-28.19
Structure: Flow + Distance (original)0.02760.30220.04530.52894.370.2659-29.27
Correlation: Flow only (new)0.03030.33320.04760.53377.750.4715-26.22
Correlation: Distance only (new)0.03070.33160.04610.51187.520.4791-26.23
Correlation: Flow + Distance (original)0.02950.32900.04470.53047.280.4401-26.06

Full leads all seven metrics. Each single-role control exceeds its joint-removal counterpart on most metrics, including ARR and IR. These single-run controls support the contribution of each relation through both propagation and diversity. We will add this table and clarify the ablation definitions in the revision.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Revisions to the Manuscript

Dear Reviewers and Area Chair,

We have uploaded a revised manuscript incorporating the clarifications and additional experiments discussed in our rebuttal. The main revisions are summarized below.

  1. Methodological clarifications (Sections 3–4 and Appendices A, C, F–G): We clarified the distinction between the Quality Pareto Archive and search Pareto selection, specified the MCTS selection and backup procedures, and explained the search diagnostics and the provenance of Figure 3.

  2. Search fairness and ablations (Appendix K.1–K.3): We added the matched 4M-token comparison, detailed search-resource accounting, normalization sensitivity experiments, and separate tests of relation propagation and diversity.

  3. Temporal robustness and statistical analysis (Appendix K.4): We added five rolling-window factor-mining evaluations and paired stationary-bootstrap confidence intervals.

  4. Transaction costs (Appendix K.5): We added turnover statistics and net portfolio performance under multiple transaction-cost assumptions.

  5. Presentation improvements (Figure 1 and Tables 1–2): We reconstructed Figure 1 as a vector diagram and improved the readability of the main result tables.

We thank the reviewers for their constructive feedback, which has helped us improve the clarity, reproducibility, and empirical evaluation of the manuscript.

Area_Chair_rSoFOpenReview ↗

Engaging with the authors

Dear reviewers,

The authors have posted several replies to your review. Please have a look and interact with the authors as needed. Reviewer Wn4u has left a comment and how about Reviewer p7rk and VcYx?

Thank you,

Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.