PaperlayerView exact revision on arXiv

Source-linked AI summary

QuantaAlpha: An Evolutionary Framework for LLM-Driven Alpha Mining

Jun Han, Shuo Zhang, Wei Li, Yifan Dong, Tu Hu, Yumo Zhu, Xiaomin Yu, Xin Guo, Zhaowei Liu, Kunyi Wang, Jingping Liu, Tianyi Jiang, Ruichuan An, Sen Hu, Zhi Yang, Ronghao Che, Huacan Wang

arXiv:2602.07085v3q-fin.STcs.AIq-fin.CP

TL;DR

Alpha mining in noisy, non-stationary markets lacks controllable multi-round search and reliable reuse of validated experience. QuantaAlpha evolves end-to-end mining trajectories through mutation and crossover while constraining factor generation. It reports stronger performance across markets, including an IC of 0.0472 on CSI 300 and positive transfer to CSI 500 and the S&P 500.

  • Problem

    Existing agentic alpha-mining systems face fragile controllability, limited inheritance of validated rationales, and constrained exploration under noisy, non-stationary markets.

  • Method

    QuantaAlpha treats each end-to-end mining run as a trajectory and improves it through diversified initialization, targeted mutation, complementary crossover, and generation constraints.

  • Results

    QuantaAlpha consistently outperforms strong baselines across Chinese and U.S. equity markets, achieving an IC of 0.0472 on CSI 300 and positive transfer performance on CSI 500 and the S&P 500.

  • Takeaways & Limitations

    The results support agentic evolution as a promising paradigm for interpretable factor discovery in high-noise, non-stationary domains.

Abstract

from arXiv · show

Financial markets are noisy and non-stationary, making alpha mining highly sensitive to backtest noise and regime shifts. While recent agentic frameworks improve automation, they often lack controllable multi-round search and reliable reuse of validated experience. To address these challenges, we propose QuantaAlpha, an evolutionary alpha mining framework that treats each end-to-end mining run as a trajectory and improves factors via trajectory-level mutation and crossover. QuantaAlpha localizes suboptimal steps for targeted revision and recombines complementary high-reward segments to reuse effective patterns, enabling structured exploration and refinement across iterations. During factor generation, it enforces semantic consistency across hypothesis, factor expression, and executable code, and constrains the complexity and redundancy of the generated factor to mitigate crowding. Extensive experiments on CSI 300 show consistent gains over strong baselines and prior agentic systems. Using GPT-5.2, QuantaAlpha achieves an IC of 0.0472 with ARR of 4.68% and MDD of 11.8%. Moreover, factors mined on CSI 300 transfer effectively to CSI 500 and the S&P 500, delivering about 40.28% and 19.1% cumulative excess return over four years, respectively, which indicates strong robustness under market distribution shifts.

1 Introduction

QuantaAlpha addresses controllability, trustworthiness, and exploration limitations in agentic alpha mining through trajectory-level evolutionary search and constrained factor generation.

  • Existing agent frameworks automate hypothesis generation, factor construction, and backtesting refinement, reducing trial-and-error costs while preserving interpretability.
  • Noisy backtest feedback can induce semantic drift, steering iterative updates toward spurious correlations and away from intended economic mechanisms.
  • Stochastic regeneration often fails to inherit validated rationales, leaving factor lineage less traceable, auditable, and trustworthy.
  • Local search can over-exploit neighborhoods around initial seeds, causing redundancy and crowding while limiting broader hypothesis-space coverage.
  • QuantaAlpha evolves complete mining trajectories through diversified initialization, targeted mutation, complementary crossover, and gates for semantic consistency, complexity, and redundancy.

2 Related Work

Related work has expanded automated financial reasoning and agentic factor research, while self-evolving agents are moving systems toward feedback-driven evolution; trading constraints remain insufficiently intrinsic.

  • Financial LLMs and evaluation benchmarks have broadened automated financial reasoning toward decision-oriented workflows such as factor discovery and trading analysis.
  • Alphafin and AlphaEval establish standardized, task-oriented protocols for reproducible evaluation of factor-related systems.
  • Trading constraints such as turnover and complexity remain largely post-hoc filters, limiting generalization and interpretability in live trading.
  • Self-evolving agents shift from static instruction-following toward autonomous learning through environmental interaction and feedback-driven evolution.

3 Problem Setup

The paper formulates alpha mining as constrained prediction over market features and models each end-to-end multi-agent run as a trajectory whose terminal reward guides policy learning.

  • Alpha Mining: Alpha mining maps each time slice of a market feature tensor to next-period cross-sectional returns through an alpha factor.
  • Alpha Mining: The optimization balances predictive effectiveness against regularization that encourages factor-expression simplicity and novelty.
  • Alpha Mining Trajectory: A mining trajectory is an ordered sequence of states and actions beginning with mining context and ending with the evaluated result.
  • Alpha Mining Trajectory: Trajectory quality is measured by the terminal reward associated with the factor produced by that trajectory.
  • Objective: The objective is to learn a trajectory-generation policy that maximizes expected terminal reward over trajectories induced from the initial state.

4 Method

QuantaAlpha frames alpha mining as a multi-agent workflow that converts hypotheses into evaluated factors and iteratively improves complete mining trajectories. It combines semantic and implementation controls with diversified initialization, trajectory-level evolution, and redundancy filtering.

  • Workflow: The workflow coordinates hypothesis generation, factor construction, implementation, and backtesting evaluation within complete mining trajectories.Each trajectory begins from market context and optional seed factors and ends in an evaluated factor.
  • Factor Realization: An intermediate operator-library expression and AST bridge market hypotheses to executable code, while verification checks consistency across hypothesis, semantics, expression, and code.Failed consistency checks trigger rewriting or repair of inconsistent components.
  • Factor Realization: Complexity and redundancy constraints reject or rewrite overly complex or structurally similar factors, followed by correlation filtering that retains the higher-RankIC factor.Structural similarity is measured through common AST subtrees, while output correlation filtering addresses functional equivalence.
  • Initialization: Diversified initialization generates complementary hypotheses across signal sources, time scales, and mechanism types to broaden the search frontier.The resulting trajectories explore multiple promising regions in parallel, reducing premature convergence risk.
  • Self-Evolution: Mutation and crossover generate improved trajectory demonstrations that provide imitation-learning priors for subsequent trajectory generation.Mutation revises one trajectory, whereas crossover recombines complementary sub-trajectories; repeated iterations form progressively improved factors.
  • Self-Evolution: Trajectory-level mutation diagnoses a suboptimal decision node and revises a localized action, preserving the fixed prefix while regenerating the remaining steps coherently.The rewrite may change the hypothesis, symbolic expression, compiled code, time scale, or regime conditions.

5 Experiments

Experiments show that QuantaAlpha improves predictive and strategy performance through trajectory evolution, generation controls, and robust transfer across markets and iterations.

  • Main Results: Using GPT-5.2, QuantaAlpha achieves the best IC at 0.0472, ARR of 4.68%, and MDD of 11.80% on CSI 300.Strategy results use approximately 150 validated factors evaluated by the same downstream LightGBM model.
  • Main Results: Compared with RD-Agent, QuantaAlpha improves IC by 0.0186 and ARR by 1.10% while reducing MDD by 4.96%.Compared with AlphaAgent, it further improves IC by 0.0125 and ARR by 3.57% while reducing MDD by 2.09%.
  • Evolutionary Component Ablation: Removing mutation causes the largest predictive decline, decreasing both IC and Rank IC by 0.0079 and ARR by 1.26%.Removing planning mainly harms strategy outcomes, while removing crossover produces a smaller but consistent degradation.
  • Generation Control Ablation: Disabling complexity control drops annualized excess return by 0.95% and increases maximum drawdown by 2.31%.Consistency, complexity, and redundancy controls each contribute, and disabling all three produces the largest degradation.
  • Factor Generalizability: QuantaAlpha factors transferred without re-optimization achieve roughly 40.28% cumulative excess return on CSI 500 and over 19.1% on the S&P 500.The transfer experiment evaluates factors mined on CSI 300 under market distribution shifts.
  • Evolutionary Efficiency: QuantaAlpha maintains the highest IC across five iterations, gaining rapidly early before stabilizing at a high level.The cumulative factor pool continues improving until performance begins to decay after roughly 15 iterations.

6 Conclusion

QuantaAlpha formulates factor discovery as a constrained multi-agent process and reports more stable and generalizable factors than existing baselines across Chinese and U.S. equity markets.

  • QuantaAlpha formulates factor discovery as a constrained multi-agent process for interpretable alpha mining.
  • Experiments across Chinese and U.S. equity markets show more stable and generalizable factors than existing baselines.
  • Broad hypothesis exploration and redundancy-aware evolution improve diversity while symbolic representations and constraint-aware synthesis maintain controllability.
  • The results identify agentic evolution as a promising paradigm for discovery problems in high-noise, non-stationary domains.

Impact Statement

The paper provides experimental and implementation details for evaluating QuantaAlpha, including metrics, data processing, baselines, computational cost, and factor reuse.

  • QuantaAlpha is evaluated with factor predictive-power and strategy-level metrics, including IC, Rank IC, ARR, and MDD.
  • Strategy metrics use excess returns after transaction costs, subtracting benchmark returns and transaction costs from portfolio returns.
  • Backtesting uses Qlib across CSI 300, CSI 500, and S&P 500, with OHLCV-derived features predicting next-day returns and cross-sectional rank normalization.
  • The benchmark suite spans machine-learning models, deep-learning models, classical factors, and LLM agents.
  • A complete run uses 10 parallel planning directions and 5 evolutionary iterations, takes approximately 20 hours, and requires no local GPU resources.
  • QuantaAlpha consumes approximately 1.8M tokens per run and produces a reusable symbolic factor pool for later backtesting, portfolio construction, and cross-market transfer.

A.5 Cross-Seed Variance and Evolutionary Robustness

QuantaAlpha remains robust across different initial seed combinations and reduces uncontrolled randomness through constrained self-evolution, while predictive performance remains significant over 966 trading days.

  • Different seed combinations introduce limited fluctuations without changing the overall conclusion about QuantaAlpha’s performance.
  • Across 966 trading days from 2022-01-04 to 2025-12-26, factors from Claude-4.5-Sonnet and DeepSeek-v3.2 show highly significant predictive power with p < 0.001.
  • Positive information coefficient days remain above 60% for the evaluated factors.
  • The standard deviations of IC and Rank IC are 0.0021 and 0.0024, respectively, across seed combinations.
  • Constrained self-evolution combines hypothesis guidance, trajectory-pool memory, feedback-driven iteration, and diversity-preserving mutation to preserve continuity and directional consistency.

A.6 Transaction Cost Sensitivity Analysis

The paper examines transaction-cost robustness alongside factor construction, evolutionary operations, and a representative crossover case showing how complementary trajectories produce an offspring factor.

  • Transaction Cost Sensitivity Analysis: QuantaAlpha is tested with 0.20% round-trip transaction costs and additional settings scaled to 0.30% and 0.40%.
  • Transaction Cost Sensitivity Analysis: QuantaAlpha maintains stable annualized returns and Sharpe ratios even when transaction costs are doubled.
  • Transaction Cost Sensitivity Analysis: Low turnover from TopkDropoutStrategy, which replaces about 10% of holdings at each rebalance, contributes to robustness under conservative cost assumptions.
  • Evolutionary Factor Construction: The evolutionary process uses original, mutation, and crossover phases, with crossover combining high-performing parent trajectories into offspring.
  • Evolutionary Factor Construction: The representative crossover combines complementary institutional-trend and retail-herding hypotheses rather than simply averaging factor expressions.
  • Evolutionary Factor Construction: The offspring factor achieves ARR (Excess) of 7.80% versus 5.20% for the baseline, while its MDD (Excess) is −11.37% versus −7.30%.

C.5 Trajectory Summary

The trajectory summary documents how factor updates combine improved predictive performance with increased risk exposure, leading to rejection for direct deployment. It also situates these results within the 2023 CSI 300 regime transition toward noisier, less persistent market structure.

  • Trajectory Evaluation: Improved annualized excess return, IC, and RankIC validate synthesizing institutional and retail momentum signals.The combined factor improves return and predictive metrics relative to the baseline.
  • Trajectory Evaluation: Increased maximum drawdown and marginally lower Information Ratio indicate higher volatility during certain market regimes.The offspring factor introduces greater risk exposure despite stronger predictive performance.
  • Trajectory Evaluation: The full hypothesis requires institutional momentum, retail herding reversal, and volatility-adaptive weighting to work together effectively.Without regime-adaptive weighting, the combined signals may amplify risks during turbulent periods.
  • Decision: The candidate was rejected for direct deployment because improved returns and IC were accompanied by deteriorating risk metrics.The summary identifies the factor’s risk profile as insufficient for immediate use.
  • Market Regime Context: The 2023 CSI 300 regime shifted from large-cap core assets toward small-cap and thematic stocks with greater intraday noise and weaker trend persistence.These changes challenge factors relying on stable intraday structure or fast mean reversion.

D.2 Factor semantics aligned with market microstructure

QuantaAlpha’s factor semantics align with changing market microstructure by emphasizing overnight information, volatility structure, and conditionally filtered trends. This semantic diversity is associated with broader coverage, stronger Rank IC tails, and greater robustness across regime shifts and stress periods.

  • Overnight and Auction Information: Overnight information becomes a dominant and stable signal when intraday predictability deteriorates.Gap-based factors occupy the right tail of the 2023 Rank IC distribution while maintaining near-complete coverage.
  • Volatility Structure: Volatility-structure and range-based signals remain predictive under elevated noise, producing a heavier positive Rank IC tail for QuantaAlpha than AlphaAgent.These factors capture abnormal variability rather than relying on directional trends.
  • Trend Quality and Liquidity: Conditioning trend continuation on low residual volatility and improving liquidity filters out noise-driven pseudo-trends during small-cap rotation.This helps QuantaAlpha retain a higher fraction of strong-performing factors than raw momentum proxies.
  • Reversal and Exhaustion Signals: Reversal and exhaustion hypotheses become unstable under non-stationary liquidity and thematic crowding, with weaker Rank IC tails and lower coverage.AlphaAgent and baseline methods rely primarily on volume–price exhaustion and bottom-fishing logic.
  • Robustness Under Distribution Shift: Semantic diversity through heterogeneous mutation and recombination supports a broad factor population whose predictive subsets persist across market style transitions.During the April 2025 stress period, QuantaAlpha exhibited lower maximum drawdown and faster recovery than AlphaAgent and RD-Agent.

E Iteration Study: Convergence of Factor-Pool Performance (CSI300)

The iteration study shows that QuantaAlpha’s factor-pool performance improves nonlinearly and reaches its best return–drawdown balance around iterations 11–12. Further iterations add redundant information and can reduce robustness.

  • Experimental Setup: QuantaAlpha runs for 15 iterations, with each iteration comprising Mutation followed by Crossover phases.The global factor pool is maintained greedily using RankIC-driven admission and capped at 50% of mined factors.
  • Convergence Pattern: Predictive performance does not increase linearly with iteration count as return capacity and risk control are jointly optimized.The study tracks hypothesis and factor-expression updates across mutation and crossover steps.
  • Convergence Pattern: Iterations 11–12, representing approximately 350 factors, provide the optimal trade-off between return and maximum drawdown.The strategy reaches a balanced high level of return and risk control at this point.
  • Post-Optimum Behavior: Subsequent iterations fail to produce significant improvement and may introduce redundant information that deteriorates drawdown performance.Additional factors can reduce overall strategy robustness after the apparent optimum.
Loading 2602.07085v3…