PaperlayerView exact revision on arXiv

Source-linked AI summary

AlphaAgent: LLM-Driven Alpha Mining with Regularized Exploration to Counteract Alpha Decay

Ziyi Tang, Zechuan Chen, Jiarui Yang, Jiayao Mai, Yongsen Zheng, Keze Wang, Jinrui Chen, Liang Lin

arXiv:2502.16789v2cs.CEcs.AI

TL;DR

Alpha mining faces alpha decay associated with overfitting and crowded factors, while existing LLM approaches can rely on established market knowledge. AlphaAgent combines regularization with an autonomous workflow, and experiments report excess returns in Chinese and U.S. markets across different regimes.

  • Problem

    Alpha mining must address alpha decay: overfitting can produce spurious factors, while crowding and LLM reliance on established factors are concerns.

    Source: §1 · p. 2

  • Method

    AlphaAgent combines originality enforcement, hypothesis alignment, and complexity control in an autonomous, iterative LLM-driven factor-mining framework.

    Source: §1 · p. 2

  • Results

    From January 2021 to December 2024, average annual excess returns were 11.0% (IR=1.5) in CSI 500 and 8.74% (IR=1.05) in S&P 500.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The authors present AlphaAgent as a direction for alpha mining that adapts to market evolution while maintaining alpha sustainability.

    Source: §5 · p. 9

Abstract

from arXiv · show

Alpha mining, a critical component in quantitative investment, focuses on discovering predictive signals for future asset returns in increasingly complex financial markets. However, the pervasive issue of alpha decay, where factors lose their predictive power over time, poses a significant challenge for alpha mining. Traditional methods like genetic programming face rapid alpha decay from overfitting and complexity, while approaches driven by Large Language Models (LLMs), despite their promise, often rely too heavily on existing knowledge, creating homogeneous factors that worsen crowding and accelerate decay. To address this challenge, we propose AlphaAgent, an autonomous framework that effectively integrates LLM agents with ad hoc regularizations for mining decay-resistant alpha factors. AlphaAgent employs three key mechanisms: (i) originality enforcement through a similarity measure based on abstract syntax trees (ASTs) against existing alphas, (ii) hypothesis-factor alignment via LLM-evaluated semantic consistency between market hypotheses and generated factors, and (iii) complexity control via AST-based structural constraints, preventing over-engineered constructions that are prone to overfitting. These mechanisms collectively guide the alpha generation process to balance originality, financial rationale, and adaptability to evolving market conditions, mitigating the risk of alpha decay. Extensive evaluations show that AlphaAgent outperforms traditional and LLM-based methods in mitigating alpha decay across bull and bear markets, consistently delivering significant alpha in Chinese CSI 500 and US S&P 500 markets over the past four years. Notably, AlphaAgent showcases remarkable resistance to alpha decay, elevating the potential for yielding powerful factors.

1 Introduction

AlphaAgent addresses alpha decay associated with overfitting and factor crowding by constraining LLM-driven alpha mining. Its regularization and closed-loop workflow accompany reported returns across Chinese and U.S. markets.

1 Introduction

  • Alpha decay is linked to overfitting that produces spurious backtest factors and crowding that can accelerate decay and trigger reversals during market stress.The introduction points to size-factor underperformance in China’s A-share market in early 2024 as an example.Source: §1 · p. 2
  • The introduction says GP and RL approaches prioritize historical performance over financial rationale, while LLMs may over-rely on established factors and exacerbate crowding.The authors identify a lack of effective constraints as a limitation of current LLM-based frameworks.Source: §1 · p. 2
  • AlphaAgent regularizes factor mining through AST-based originality checks, LLM-evaluated hypothesis alignment, and complexity control.Its AST-based similarity measure compares generated factors with existing alpha libraries, while consistency scoring aligns hypotheses, descriptions, and expressions.Source: §1 · p. 2
  • The autonomous workflow proposes hypotheses, constructs and evaluates factors, backtests them, and uses feedback to guide refinement.The idea agent begins from market insights provided by domain experts; the factor agent seeks parsimonious, original factors, and the evaluation agent checks executability and numerical stability.Source: §1 · p. 2
  • From January 2021 to December 2024, average annual excess returns were 11.0% (IR=1.5) in CSI 500 and 8.74% (IR=1.05) in S&P 500.The authors report that these results outperform traditional and existing LLM-based approaches in performance and persistence.Source: §1 · p. 2
  • AlphaAgent achieved an 81% higher effective factor ratio (hit ratio) while consuming 30% fewer tokens.The authors also report stable factor performance across different market regimes.Source: §1 · p. 2

2 Related Work

Related work covers GP and RL factor-search methods, LLM-based mining approaches, and their reported challenges with alpha decay.

2 Related Work

  • GP research extends factor search with parameter learning and matrix operations or sparsity constraints, while retaining explicit formulas or guiding search toward lower-complexity factors [7, 20, 25, 28, 43, 46].AlphaEvolve adds parameter learning and matrix operations; Shen et al. use sparsity constraints during mutation.Source: §2 · p. 2
  • RL methods use policy learning, performance rewards, or factor combinations, but can overemphasize historical optimization and produce over-complex, overfitted, or poorly reasoned factors [31] [31] [45].The cited examples include AlphaGen’s performance reward and an approach that discards the critic network and shapes rewards for profitable, stable alphas [45].Source: §2 · p. 3
  • LLM approaches include adaptive generation, knowledge-guided prompting, dynamic factor combination, and feedback loops, yet the literature gives limited attention to alpha decay and lacks effective regularization [44] [36] [41].The authors identify over-reliance on historical patterns and existing market knowledge as a continuing concern.Source: §2 · p. 3

3 AlphaAgent

AlphaAgent frames alpha mining as predictive optimization constrained by factor simplicity, novelty, and alignment with market hypotheses. It operationalizes those constraints through AST-based factor construction and a closed loop of LLM agents that generate, assess, and refine factors.

3.1 Problem Formulation

  • The objective maximizes predictive effectiveness while regularizing factor expressions, with λ balancing performance against simplicity or novelty.The LLM-based version adds hypothesis alignment, factor novelty, and expression complexity to the regularization term.Source: §3.1 · p. 3

3.2 Factor Generation Modeling

  • An operator library standardizes factor-building operations to address code-generation barriers and bridge high-level market hypotheses with implementation details.The paper identifies data-format incompatibilities, package-version inconsistencies, and semantic-coherence difficulties as challenges for code-based factor construction.Source: §3.2 · p. 3, §3.2.1 · pp. 3–4

3.2.1 Factor Parsing with Abstract Syntax Trees.

  • Text hypotheses are parsed into ASTs by mapping phrases to operators, assigning parameters, and assembling trees whose leaves are raw features and whose internal nodes are operators.Factors can be bound to data fields such as price and volume.Source: §3.2.1 · p. 4

3.2.2 Interpretability and Complexity Control.

  • Regularization penalizes symbolic length and free parameters, while originality uses largest-common-subtree similarity and LLM checks of hypothesis-description and description-expression consistency.The resulting regularizer combines complexity, originality, and hypothesis alignment.Source: §3.2.2 · pp. 4–5

3.3 Autonomous Multi-Agent Framework

  • The idea agent proposes hypotheses, the factor agent constructs candidate expressions, and the eval agent backtests them and feeds performance findings into the next iteration.The agents also draw on evolving knowledge of successful and failed implementations and evaluation history.Source: §3.3 · pp. 5–6

4.1 Experiment Settings

The experiments assess predictive precision, risk-adjusted performance, returns, and drawdowns across Chinese and U.S. stock markets, using financial data and a common forecasting setup to compare AlphaAgent with established baselines.

4.1.1 Metrics.

  • IC and RankIC measure forecasting precision, ICIR and IR assess consistency and risk-adjusted excess returns, and AR and MDD capture annualized excess return and peak-to-trough decline.Persistent IC/RankIC and stable ICIR indicate sustained predictive power; AR and MDD show the practical impact of deterioration.Source: §4.1.1 · p. 6

4.1.2 Backtest Settings.

  • CSI 500 and S&P 500 OHLCV data span 2021–2024, with CSI 500 data from Baostock and S&P 500 data from Yahoo Finance.Experiments use Qlib; Table 1 gives the precise training, validation, and testing periods and trading-day counts.Source: §4.1.2 · p. 6, Table 1 · p. 6
  • A LightGBM model forecasts next-day returns from four fundamental alphas and newly proposed factors, using cross-sectional Z-score normalization and a top-k dropout strategy.Backtests select 50 top-ranked stocks, exclude the five lowest-ranked, and account for transaction fees; fees differ by market.Source: §4.1.2 · p. 6
  • Comparisons include time-series, tree-based, specialized financial, agent-based, and deep-reasoning models, including LSTM, LightGBM, AlphaForge, RD-Agent, OpenAI-o1, and DeepSeek-R1.RD-Agent uses GPT-4-turbo, while AlphaAgent uses GPT-3.5-turbo.Source: §4.1.3 · p. 6, §4.1.2 · p. 6

4.2 Overall Performance

Across four years in the CSI 500 and S&P 500, AlphaAgent leads the reported performance metrics. Its cumulative excess returns remain comparatively stable in both markets, while several alternatives show decay.

4.2 Overall Performance

  • The evaluation covers January 2021–December 2024 and scores methods on IC, ICIR, AR, IR, and MDD; RD-Agent and AlphaAgent each undergo 20 trials of five evolutionary rounds.Reasoning models Deepseek-R1 and OpenAI-o1 are prompted to mine and evolve alphas through five rounds.Source: §4.2 · pp. 6–7
  • AlphaAgent outperformed across all key metrics: AR was 11.00% versus LSTM’s 4.96% in CSI 500 and 8.74% versus Deepseek-R1’s 2.75% in S&P 500.It also recorded the highest IC, ICIR, and IR in both markets, with MDDs of -9.36% in CSI 500 and -9.10% in S&P 500.Source: Table 2 · p. 7, §4.2 · p. 7
  • Cumulative excess returns were approximately 45% in CSI 500 and exceeded 37% in S&P 500, while time-series models had negative U.S. returns from 2023-02.The comparison shows relatively stable quarterly performance for AlphaAgent; Transformer had a notable drawdown after 2024-08.Source: Figure 3 · p. 7, §4.2 · p. 7

4.3 Alpha Decay Analysis

On CSI 500, AlphaAgent’s mined alphas retain predictive performance across years, whereas Alpha158, GP, and RSI show substantial declines. The analysis links these patterns to overfitting and crowding risks.

4.3 Alpha Decay Analysis

  • The authors associate GP’s deterioration with possible overfitting and Alpha158/RSI degradation with crowding, while AlphaAgent’s sustained performance suggests exploration of less-exploited factors.The text describes AlphaAgent’s factor modeling mechanism as effectively mitigating the crowding effect affecting traditional approaches.Source: §4.3 · p. 8
  • Alpha158, GP, and RSI ICs fell from 0.022–0.036 to near zero and RankICs from 0.020–0.042 to around zero, while AlphaAgent stayed near 0.02 IC and 0.025 RankIC.Figure 4 compares yearly IC and RankIC on CSI 500, including 15 AlphaAgent alphas.Source: §4.3 · p. 7, Figure 4 · p. 8

4.4 Alpha Mining Efficiency Analysis

Across five CSI 500 test-split rounds, AlphaAgent has higher average IC than RD-Agent and AlphaForge, while its IC variance increases as rounds progress.

4.4 Alpha Mining Efficiency Analysis

  • The authors attribute AlphaAgent’s higher average IC to complexity and hypothesis-alignment constraints, and its broader exploration to originality penalties.RD-Agent’s smaller variance is described as likely reflecting less exploration incentive and more homogeneous candidates.Source: §4.4 · p. 8
  • AlphaAgent maintains higher average IC than RD-Agent and AlphaForge across rounds, while its expanding variance suggests greater factor diversity and a wider exploration space.Figure 5 shows IC evolution over the first five rounds; RD-Agent has comparatively stable, smaller variance.Source: Figure 5 · p. 8, §4.4 · p. 8

4.5 Ablation Study

AlphaAgent’s ablations report gains in hit ratio, development success, and token efficiency, while comparisons across base LLMs show that model choice affects factor quality and AlphaAgent improves over RD-Agent across variants.

4.5 Ablation Study

  • The ablation evaluates hit ratio, executable-factor success, and token efficiency; hit ratio counts alphas exceeding market-specific annualized-return thresholds.Development success measures execution without code defects or numerical errors; token efficiency is inversely related to tokens consumed per candidate generation.Source: §4.5 · p. 8
  • 0.29 versus 0.16 without factor modeling constraints yields an 81% higher hit ratio; development success is 0.83 versus 0.75, and token efficiency improves 23%.The ablation spans 100 evolution rounds split evenly between CSI 500 and S&P 500; the hit-ratio thresholds represent the top 5% of generated alphas.Source: Figure 6 · p. 8, §4.5 · pp. 8–9
  • DeepSeek-R1 achieves the highest ICIR (0.0615) and annualized return (9.19%), with the lowest drawdown (-6.50%) among the three tested base LLMs.Figure 7 compares GPT-3.5-turbo, Qwen-Plus, and DeepSeek-R1 on S&P 500 from 2021 to 2024.Source: Figure 7 · p. 9, §4.5 · p. 9
  • AlphaAgent’s IC differences from RD-Agent are statistically significant across GPT-3.5-turbo (p=0.0311), Qwen-Plus (p=0.0109), and DeepSeek-R1 (p=0.0382).Each reported p-value is below 0.05.Source: §4.5 · p. 9

4.6 Implications

The implications section argues that alpha mining must explore beyond historical patterns and established theories. It links active exploration with finding emerging inefficiencies before they dissipate.

4.6 Implications

  • Traditional statistical arbitrage faces diminishing returns in highly efficient markets as competition and market adaptation increase.The section frames this as a motivation for approaches that can go beyond fitting historical patterns and relying on established theories.Source: §4.6 · p. 9
  • The experiments indicate that active exploration mechanisms can better identify and capitalize on emerging market inefficiencies before they dissipate.The passage situates this claim in rapidly evolving markets.Source: §4.6 · p. 9
  • Future quantitative trading research should focus on frameworks balancing exploitation of known patterns with exploration of novel alpha sources.The passage relates this priority to increasingly prevalent AI-driven market analysis.Source: §4.6 · p. 9

5 Conclusion

The conclusion presents AlphaAgent as an LLM-driven framework for counteracting alpha decay through three regularization mechanisms. It says the framework produces performant, decay-resistant factors across market conditions and points to adaptive alpha mining as a future direction.

5 Conclusion

  • AlphaAgent combines originality enforcement, complexity control, and hypothesis alignment to address alpha decay.The conclusion describes it as an autonomous LLM-driven framework.Source: §5 · p. 9
  • The framework is reported to produce performant, decay-resistant alpha factors with substantial excess returns and consistency across market conditions.The conclusion also characterizes the factors as theoretically sound.Source: §5 · p. 9
  • The conclusion identifies swiftly adapting to market evolution while maintaining alpha sustainability as a promising direction for next-generation alpha mining.Source: §5 · p. 9

Related papers

Loading 2502.16789v2…