PaperlayerView exact revision on arXiv

Source-linked AI summary

EvoQuantLab: Training-Free Dual-Loop Evolution for Auditable Quantitative Research

Haolin Yang, Zhiqiang Qian, Chutian Chen, Jipeng Zhang, Zhenghao Zhu, Tangjia Deng, Jian Xie, Shuaishuai Gong, Yike Guo, Sirui Han

arXiv:2610.17130v2cs.AI

TL;DR

Automated factor discovery faces adaptive backtest overfitting as repeated experiments feed results into later searches. EVOQUANTLAB combines genetic programming with cross-cycle diagnosis and reports out-of-sample gains across daily equities and hourly cryptocurrency.

  • Problem

    Repeated evaluation of many factors and variants can make apparently strong signals reflect noise or adaptive selection, while using LLM calls for every symbolic variation is costly.

    Source: §1 · p. 1

  • Method

    EVOQUANTLAB pairs LLM-guided hypotheses with diversity-preserving genetic programming and an outer diagnostic loop that updates future research protocols.

    Source: §1 · pp. 1–2

  • Results

    Across daily equities and hourly cryptocurrency, EVOQUANTLAB outperforms representative agent baselines; its hourly cryptocurrency annualized return is 37.91%.

    Source: §1 · p. 2, §6 · p. 10

  • Takeaways & Limitations

    The framework provides an auditable research workflow under immutable task contracts and one-way out-of-sample governance.

    Source: §6 · pp. 9–10

Abstract

from arXiv · show

Automated alpha factor discovery is critical for modern quantitative research. While large language models (LLMs) have increasingly been integrated into quantitative workflows, existing agent-based approaches remain highly susceptible to severe backtest overfitting. To address this challenge, we propose EvoQuantLab, a training-free, dual-loop self-evolving framework for robust factor discovery. Instead of relying entirely on LLM generation, EvoQuantLab deploys an inner genetic-programming (GP) loop that dramatically improves token efficiency while evolving typed factor expressions under structural and diversity constraints. An outer diagnosis loop then systematically audits evidence across cycles to drive protocol-level improvements. Furthermore, our evaluation pipeline applies multi-metric validation to rigorously filter out spurious factors before deployment. Extensive out-of-sample experiments across both daily and hourly frequencies -- covering equities and cryptocurrency -- demonstrate the robustness of EvoQuantLab. Specifically, on daily CSI 500 and S&P 500 tasks, EvoQuantLab achieves leading annualized returns of 18.34% and 21.30%, consistently outperforming existing agent baselines. On high-turnover hourly cryptocurrency data, EvoQuantLab delivers a 37.91% annualized return while effectively avoiding execution collapse.

1 INTRODUCTION

EVOQUANTLAB addresses adaptive backtest overfitting and costly factor exploration with a training-free dual-loop framework. It reports stronger out-of-sample results across daily equities and hourly cryptocurrency tasks.

1 INTRODUCTION

  • Repeatedly feeding back development results across many factor variants raises the risk that strong signals reflect noise or adaptive selection rather than persistent predictive structure.The paper also notes that LLM calls for every candidate are costly when many variants are symbolic transformations.Source: §1 · p. 1
  • The framework combines LLM-generated economic hypotheses with genetic programming to evolve typed expressions through selection, crossover, and mutation while preserving structural and population diversity.The GP loop reduces reliance on repeated LLM inference for symbolic factor exploration.Source: §1 · p. 2
  • An outer loop audits completed cycles for overfitting and other risks, then updates protocols for future cycles alongside multi-metric filtering of unstable, redundant, or economically inconsistent candidates.Protocol changes do not retroactively alter completed evaluations.Source: §1 · p. 2
  • 18.34% on CSI 500 and 21.30% on S&P 500 annualized returns exceed strongest baselines’ 14.02% and 15.23%, respectively; hourly cryptocurrency return reaches 37.91%.The paper also reports GP efficiency for large-scale exploration and improved robustness across successive research cycles.Source: §1 · p. 2

2 RELATED WORK

Prior work uses symbolic search, learning-based methods, and LLM agents to automate factor discovery and broader quantitative research. EVOQUANTLAB differs by evolving research knowledge and symbolic search under a frozen evidence protocol.

2 RELATED WORK

  • Classical GP and learning-based approaches search or optimize factor expressions, but typically keep their operators and search procedures fixed; EVOQUANTLAB evolves validated knowledge and search vocabulary across cycles.Examples include AutoAlpha [Zhang et al., 2020] and AlphaGen [Yu et al., 2023].Source: §2 · p. 2
  • LLM agents automate hypothesis generation, implementation, evaluation, and refinement, but repeated exposure to noisy backtest feedback can misguide planning in open-ended trial-and-error loops.EVOQUANTLAB separates LLM updates to reusable research knowledge from deterministic, budget-accounted GP factor search.Source: §2 · p. 3
  • Self-evolving-agent research adapts workflows and procedural skills, while EVOQUANTLAB applies this system-level adaptation to research knowledge and symbolic search without allowing held-out outcomes to influence later evolution.The cited line of work includes program and heuristic search [Wang et al., 2023] and workflow adaptation [Hu et al., 2025; Zhang et al., 2025a;b].Source: §2 · p. 3

3 EVOQUANTLAB

EVOQUANTLAB separates fast factor-expression search from slower protocol evolution across research cycles. A frozen evaluation contract and staged validation pipeline govern candidates, while held-out outcomes remain outside search and diagnosis.

3 EVOQUANTLAB

  • The framework addresses symbolic-space exploration and adaptive overfitting through a fast GP loop within a frozen protocol and a slow loop that audits evidence across cycles to update protocol rules.These loops operate on distinct timescales.Source: §3 · p. 3

3.1 FORMAL SETUP & IMMUTABLE TASK CONTRACT

  • Each cycle freezes a protocol specifying the task and data contract, terminals, typed operators, evaluation rules, and search and trial budgets; candidate selection uses development evidence, while the holdout remains inaccessible during search.The protocol also enforces point-in-time availability and prefix replay invariance; candidates are scored across development folds with a complexity penalty.Source: §3.1 · p. 4

3.2 THE FROZEN L0–L6 EVIDENCE PIPELINE

  • The L0–L6 pipeline moves candidates from data and typed-expression search through economic falsification, stress testing, admission, and portfolio construction.Falsification checks mechanism consistency and incremental value after ablations and residualization; stress tests cover regimes, execution delays, asset clusters, and matched-budget permutations before admitted implementations are frozen.Source: §3.2 · p. 4

3.3 LOOP A: INNER GP SEARCH & KNOWLEDGE INGESTION

  • Knowledge ingestion admits proposed primitives only after reference and prefix-causality checks, while GP evolves typed trees through selection, crossover, and mutation under diversity safeguards.Near-duplicate or highly correlated expressions are replaced with immigrants from underrepresented families, and family-stratified archives preserve structural diversity.Source: §3.3 · p. 4

3.4 LOOP B: DIAGNOSIS-DRIVEN PROTOCOL EVOLUTION

  • After a cycle barrier, the outer loop audits execution traces for overfitting risks and compares real-search results with budget-matched permutation pseudo-searches to diagnose whether search finds alpha beyond noise.Gate 1 can flag model or selection overfitting, suspected leakage, metric fragility, or insufficient evidence.Source: §3.4 · p. 5
  • Diagnosis routes review toward the implicated protocol components, and bounded patches affect only the next protocol version; completed-cycle scores, selections, and evaluation rules remain unchanged.For example, selection overfitting routes review to the search space and trial budget.Source: §3.4 · p. 5

4 MAIN EXPERIMENTS

The experiments compare predictive and execution-aware portfolio performance across daily equity and hourly cryptocurrency tasks, then audit whether development-selected factors retain out-of-sample performance without retuning.

4.1 TASKS AND METRICS

  • Evaluation spans next-day CSI 500 and S&P 500 prediction and next-hour cryptocurrency prediction, measuring predictive IC/ICIR and execution-aware portfolio returns, IR, and drawdown.The portfolio metrics include annualized and cumulative return, information ratio, and maximum drawdown.Source: §4.1 · pp. 5–6

4.2 DAILY AND HOURLY RESULTS

  • On CSI 500, EVOQUANTLAB leads with IC 0.0281, ICIR 0.1211, IR 1.613, and cumulative return 5.31%; on S&P 500, its annualized return is 21.30%.The reported S&P 500 results also include IC 0.0233, IR 1.614, and MDD -5.82%.Source: §4.2 · p. 6, Table 2 · p. 6
  • On Crypto, EVOQUANTLAB posts the highest IR (3.709), cumulative return (7.17%), and lowest drawdown (-3.23%), while RD-Agent's highest IC (0.0444) accompanies IR -1.870.Figure 4 compares predictive and portfolio metrics alongside OOS cumulative excess-return trajectories.Source: §4.2 · p. 6, Figure 4 · p. 7

4.3 DEV-TO-OOS GENERALIZATION AND OVERFITTING AUDIT

  • Under frozen, no-retuning transfer, EVOQUANTLAB averages OOS IR 2.521 and retains positive alpha in 100% of tasks; several baselines suffer IR collapses and sign reversals.The audit selects each method's single best factor using development IR, freezes it, and tests it on held-out OOS data across three markets.Source: §4.3 · pp. 6–7, Table 3 · p. 7

5 ABLATION STUDIES

Ablations, cycle trajectories, resource comparisons, and backbone tests examine which components and search behaviors accompany stronger factor discovery and portfolio performance.

5.1 COMPONENT ABLATION STUDY

  • Directed GP updates were necessary for filter-passing alphas, while dynamic L1 expansion, outer diagnosis, and family diversity control were critical to downstream alpha quality.Without GP updates, no candidates passed the multi-stage filters; the ablations use frozen out-of-sample evaluations and a matched candidate-evaluation budget.Source: §5.1 · pp. 7–8, Table 4 · p. 7

5.2 EVOLUTION ACROSS CYCLES

  • Cycle 5 added three L5-approved coverage-tail factors and raised cumulative IR from 2.67 to 3.21; Cycle 6 reached peak portfolio IR 3.62 and arithmetic return 8.52%.Earlier momentum- and spread-based evolution raised cumulative IR from 1.57 at Cycle 0 to 2.93 by Cycle 2; Cycle 3's highly overlapping candidates (ρ = 0.881) added negligible gain.Source: §5.2 · p. 8
  • Later cycles showed diminishing returns: portfolio IR fell from 3.62 to 3.59 at Cycle 7, and deduplication reduced Cycle 10's corresponding IR from 3.39 to 3.08.The trajectory associates stronger gains with opening new factor families, while later cycles revisit correlated structures and approach a plateau.Source: §5.2 · p. 8

5.3 EVOLUTION EFFICIENCY ANALYSIS

  • EVOQUANTLAB ran for 22h08m, with its first valid candidate at 2h45m; despite the longer run, it used 50 LLM calls and cost $9.000.QuantaAlpha used 94 API calls and cost $15.224 during its 2-hour run, with over 767K input and 455K output tokens.Source: §5.3 · p. 9

5.4 BACKBONE MODEL ANALYSIS

  • Claude Opus5 achieved the strongest portfolio-level performance, while Qwen3.8-27B had the highest standalone IC but lower admission and a larger predictive-to-portfolio gap.Claude Sonnet 5 showed competitive intermediate performance across dimensions.Source: §5.4 · p. 9

6 CONCLUSION

The conclusion presents EVOQUANTLAB as a dual-loop framework for reducing overfitting and trial multiplicity, supported by evaluations across equities and digital assets.

6 CONCLUSION

  • Across daily equities and high-frequency digital assets, EVOQUANTLAB outperformed representative agent baselines in realized, risk-adjusted returns; ablations and cycle analyses support the roles of directed evolution, knowledge acquisition, and multi-metric falsification.The conclusion also reports a modest LLM token and financial footprint.Source: §6 · pp. 9–10

A DUAL-LOOP EXECUTION PROCEDURE

  • The outer-cycle procedure freezes each completed inner GP artifact, audits finalist evidence, admits evidence-backed patches only to the next protocol, and checkpoints state atomically.Patches lacking evidence or regression tests, touching immutable fields, or attempting to change the completed cycle are rejected.Source: A DUAL-LOOP EXECUTION PROCEDURE · p. 13

B GATE 1 EVIDENCE SCHEMA

  • Gate 1 thresholds are calibrated per task before OOS reads; the protocol manifest records calibrated values and their calibration data.The manifest also records the commit preceding all OOS reads.Source: Table 6 · p. 13, B GATE 1 EVIDENCE SCHEMA · p. 13

C EVALUATION DETAILS

The evaluation defines factor-level predictive metrics and portfolio-level performance measures, then describes separate execution-aware strategies for equities and hourly cryptocurrency. The case studies illustrate three research-evolution behaviors, with decisions based only on training and development evidence.

C.1 METRICS

  • IC measures cross-sectional correlation between predicted scores and subsequent realized returns, while ICIR captures predictive stability over time.IR measures risk-adjusted excess return and corresponds to the annualized Sharpe ratio under the dollar-neutral, zero-benchmark setup; MDD captures peak-to-trough downside, and cumulative and annualized returns summarize portfolio gains.Source: §C.1 · p. 14

C.2 TRADING STRATEGY

  • Equity strategies use daily-bar LightGBM on pre-screened factors, while cryptocurrency uses an equal-weight hourly factor portfolio with smoothed, inverse-volatility-scaled positions.Equity factors are selected using development data; daily reported returns are gross of execution fees and slippage. Crypto signals have one-hour execution latency and incur 5 bps per one-way turnover, excluding perpetual funding costs.Source: §C.2 · p. 15

D CASE STUDIES OF RESEARCH EVOLUTION

  • Three Crypto case studies cover redirecting an unproductive search, rejecting attractive development candidates, and testing model-suggested knowledge before search-space admission.The decisions used training and development evidence only; Table 7 presents two high-development candidates rejected at different validation stages.Source: D CASE STUDIES OF RESEARCH EVOLUTION · pp. 15–16, Table 7 · p. 16

D.1 CASE 1: REDIRECTING SEARCH AFTER A FAILED CYCLE

After Cycle 4 produced no admitted factors, the outer review changed the next-cycle search policy. Cycle 5 showed lower family concentration and renewed admissions, although the trace does not establish that the policy change caused the improvement.

D.1 CASE 1: REDIRECTING SEARCH AFTER A FAILED CYCLE

  • Cycle 4 yielded no residual-validation survivors or L5 admissions, so the outer review rejected all 240 candidates and issued 15 cross-layer terminal bans for Cycle 5.The largest family occupied 55.6% of the final Cycle 4 population.Source: §D.1 · p. 16
  • 41.9% maximum family share and 0.795 normalized family entropy accompanied three L5-approved factors in Cycle 5, versus 55.6%, 0.689, and none in Cycle 4.The three Cycle 5 factors included one whose development-fold scores were 5.881, 5.982, and 4.910.Source: §D.1 · p. 16
  • The Cycle 4–5 trace is not a controlled causal comparison: it shows the policy intervention preceded lower family concentration and renewed admission, but does not attribute the gains to that intervention alone.The outer loop can alter the search protocol after a failed cycle rather than continue optimizing in the same unproductive region.Source: §D.1 · p. 16

D.2 CASE 2: HIGH DEVELOPMENT PERFORMANCE IS NOT SUFFICIENT

The case contrasts rejection for unnecessary complexity with rejection for redundancy: strong development performance alone does not establish that a candidate should be admitted. A candidate must also provide sufficient incremental evidence and independent information.

D.2 CASE 2: HIGH DEVELOPMENT PERFORMANCE IS NOT SUFFICIENT

  • The 20-node gp-87c5cf72fc0e scored 7.192, 7.120, and 5.487 across development folds, yet was rejected because its simpler counterfactual scored 5.496 on the worst fold and its combination lift was −0.0096.Its execution-delay and residual tests passed; the combination lift was below the required 0.25.Source: §D.2 · p. 16
  • The gp-5f24c12b824c candidate passed robustness and residual checks but was rejected because its position-weight correlation of ρ = 0.726 exceeded the 0.7 deduplication threshold.The less redundant representative was retained.Source: §D.2 · pp. 16–17
  • Development performance is necessary but insufficient: candidates can be predictive yet unnecessarily complex, or robust individually yet add little independent information.The two candidates illustrate different rejection reasons despite attractive raw development performance.Source: §D.2 · pp. 16–17

D.3 CASE 3: MODEL-SUGGESTED KNOWLEDGE IS TESTED, NOT TRUSTED

Model-assisted reference extraction can propose knowledge for testing, but does not directly place it in the active search space. In this case, a paired-seed counterfactual failed the admission threshold and the next-cycle policy excluded the proposal.

D.3 CASE 3: MODEL-SUGGESTED KNOWLEDGE IS TESTED, NOT TRUSTED

  • The signed-volume mechanism entered as PROPOSED NOT ACTIVE, and external knowledge changed the search space only after validation under the frozen development protocol.The mechanism was motivated by persistent imbalanced order flow.Source: §D.3 · p. 17
  • The paired-seed improvements in minimum-fold score were 0.206 and 0.243, giving a median Δpaired = 0.225 below the 0.25 threshold and a REJECTED PAIRED SEED decision.The comparison tested an aligned terminal against a deterministic time-shifted control under the same expression and seed.Source: §D.3 · p. 17
  • The rejection gave book dimb524 zero next-cycle budget, omitted its formula, and placed it in Cycle 3’s banned and removed sets.The trace ran from reference proposal through provisional terminal and paired counterfactual to rejection and next-cycle exclusion.Source: §D.3 · p. 17

E DEV-TO-OOS FACTOR-TRANSFER AND OVERFITTING AUDIT

The transfer audit reports sign survival across three markets, positive out-of-sample performance despite larger candidate pools, and strong out-of-sample ranking of development winners. Terminal lifecycle and patch-admission rules preserve trial history and require evidence and validation before protocol changes are staged for child protocols.

E DEV-TO-OOS FACTOR-TRANSFER AND OVERFITTING AUDIT

  • Sign survival was 80.0% in Crypto, 95.5% in CSI 500, and 81.1% in S&P 500, while QuantaAlpha had 0.0% survival on CSI 500 and RD-Agent and AlphaAgent had S&P 500 sign flips.Table 8 defines positive sign survival as the share of factors with positive development IR that remain positive out of sample.Source: E DEV-TO-OOS FACTOR-TRANSFER AND… · pp. 17–18, Table 8 · p. 17
  • EVOQUANTLAB evaluated 12 Crypto, 36 CSI 500, and 49 S&P 500 candidates, versus baselines evaluating one or at most four, and sustained positive out-of-sample IRs.The audit attributes this result to outer-loop diagnostic falsification insulating the search trajectory from selection bias.Source: E DEV-TO-OOS FACTOR-TRANSFER AND… · p. 18
  • The development winner ranked at the 89.4% out-of-sample percentile on S&P 500 and the 80.0% percentile on Crypto.The audit describes these as top-tier out-of-sample positions within their candidate pools.Source: E DEV-TO-OOS FACTOR-TRANSFER AND… · p. 18

F TERMINAL LIFECYCLE AND PATCH ADMISSION

  • Terminals are classified as ACTIVE, INTERACTION-ONLY, DORMANT, or BANNED by evidence and falsification status, and dormant terminals remain in the manifest to preserve trial history.BANNED is reserved for leakage, false field semantics, or repeated independent falsification.Source: F TERMINAL LIFECYCLE AND PATCH ADMISSION · p. 18
  • Patch admission rejects empty evidence lists, empty validation plans, and immutable paths; staged patches are applied only to child protocols that record their patch IDs and parent ID.Each patch includes its target layer, path, value, evidence, benefit, risks, and validation plan.Source: F TERMINAL LIFECYCLE AND PATCH ADMISSION · p. 18

G TASK-SPECIFIC CONTRACTS

The task-specific contracts define point-in-time data construction, labels, and evaluation checks separately for Hong Kong order-book, daily equity, and hourly crypto tasks.

G.1 HOURLY HONG KONG LIMIT ORDER BOOK

  • The Hong Kong LOB adapter builds hourly point-in-time states from valid two-sided observations, predicts next-hour returns, and prevents within-bar lookahead and targets crossing session boundaries.Rolling state resets at midday and end of day, and tick size follows the price-dependent market schedule.Source: §G.1 · p. 18

G.2 DAILY CROSS-SECTIONAL EQUITIES

  • Daily equity adapters use past-only OHLCV rolling operations and fold-specific cross-sectional transformations to predict normalized next-day returns, judging CSI 500 and S&P 500 separately.Reports include IC, RankIC, cross-sectional R2, worst-fold score, direction consistency, concentration diagnostics, and a costed top-K portfolio as a secondary diagnostic.Source: §G.2 · p. 18

G.3 HOURLY CRYPTO CROSS-SECTION

  • The crypto adapter uses an hourly asset panel, historical liquidity universe, and lagged participant-flow inputs, with finalist gates spanning stability, costs, ablations, residual survival, correlation, and sign-reversal controls.The out-of-sample interval is declared in the task contract and physically withheld.Source: §G.3 · p. 18

Related papers

Loading 2610.17130v2…