QuantaAlpha: An Evolutionary Framework for LLM-Driven Alpha Mining

In one sentence

Summary

Financial alpha mining is hindered by market non-stationarity and noisy backtest feedback, which often cause LLM-based agents to drift into spurious correlations or local optima. QuantaAlpha addresses this by treating each mining run as a trajectory and applying evolutionary operators to refine factors systematically. The framework enforces semantic consistency between hypotheses, symbolic expressions, and code, while constraining complexity to prevent crowding. This approach aims to provide controllable, traceable, and robust factor discovery that adapts to regime shifts. The method employs a multi-agent system for hypothesis generation, symbolic factor construction, and backtesting. Evolution occurs via mutation, which localizes and rewrites suboptimal trajectory segments, and crossover, which recombines high-reward segments from parent trajectories. Experiments on the CSI 300 dataset (2016-2025) compare QuantaAlpha against traditional ML, deep learning, and other LLM agents like AlphaAgent and RD-Agent. The system uses GPT-5.2 as the backbone and evaluates factors using Information Coefficient (IC) and strategy metrics like Annualized Return (ARR) and Maximum Drawdown (MDD). Cross-market transfer to CSI 500 and S&P 500 is also tested. QuantaAlpha achieves an IC of 0.0472 and ARR of 4.68% on CSI 300, outperforming baselines. Ablation studies confirm that mutation drives exploration while crossover aids reuse. The framework demonstrates strong robustness during the 2023 market regime shift, maintaining performance where baselines failed due to alpha decay. Factors mined on CSI 300 transfer effectively to CSI 500 and S&P 500, yielding significant cumulative excess returns. Limitations include the need for transaction cost modeling in live deployment and potential diminishing returns after ~15 iterations.

Research problem

Existing LLM-based alpha mining agents suffer from fragile controllability, limited trustworthiness due to stochastic re-generation, and constrained exploration that leads to factor crowding. The paper addresses the challenge of mining robust alpha factors in noisy, non-stationary financial markets where backtest feedback is unreliable and regime shifts are common.

Core contributions

• Proposes QuantaAlpha, an evolutionary framework that treats alpha mining as trajectory optimization rather than one-shot generation.
• Introduces trajectory-level mutation and crossover operators to enable structured exploration and reuse of validated experience.
• Implements semantic consistency gates and complexity/redundancy controls to prevent drift and factor crowding.
• Demonstrates superior performance and robustness against regime shifts compared to AlphaAgent and RD-Agent.
• Shows effective zero-shot transfer of mined factors to different markets (CSI 500, S&P 500).

Method

• Multi-agent workflow: Hypothesis generation, symbolic factor construction, and backtesting.
• Symbolic representation: Factors are built using an operator library and AST for controllable generation.
• Consistency verification: LLM verifier ensures alignment between hypothesis, symbol, and code.
• Complexity control: Regularizes symbolic length, parameters, and feature usage.
• Redundancy control: Filters factors based on AST similarity and correlation.
• Trajectory mutation: Localizes and rewrites suboptimal steps in a mining trajectory.
• Trajectory crossover: Recombines high-reward segments from parent trajectories.
• Diversified planning: Initializes search with complementary hypotheses to avoid local optima.

Results & findings

• QuantaAlpha with GPT-5.2 achieves IC 0.0472, ARR 4.68%, and MDD 11.80% on CSI 300.
• Outperforms AlphaAgent by 0.0125 IC and 3.57% ARR.
• Ablation shows mutation is critical for exploration (IC drop 0.0079 without it).
• Complexity control is vital for strategy-level robustness (ARR drop 0.95% without it).
• Factors transfer to CSI 500 with ~40.28% cumulative excess return over four years.
• Factors transfer to S&P 500 with ~19.1% cumulative excess return over four years.
• Maintains performance during the 2023 market regime shift where baselines failed.

Data & experiments

• Datasets: CSI 300 (training 2016-2020, validation 2021, testing 2022-2025), CSI 500, and S&P 500.
• Baselines: Traditional ML (XGBoost, LightGBM), Deep Learning (Transformer, LSTM), Factor Libraries (Alpha158), and LLM Agents (AlphaAgent, RD-Agent).
• Metrics: IC, Rank IC, ARR, MDD, and Information Ratio.
• Experiments: Main performance comparison, ablation of evolutionary components, ablation of consistency/complexity controls, and cross-market transfer analysis.
• Case study: Detailed trace of hypothesis and factor evolution over 5 iterations.

Limitations & caveats

• Results are based on backtesting; real-world deployment requires transaction cost modeling.
• Performance may diminish after ~15 iterations due to redundancy.
• Dependence on high-quality LLM backbones (e.g., GPT-5.2) for optimal performance.
• Cross-market transfer assumes standard OHLCV data and daily normalization.
• Complexity of the multi-agent system may increase computational costs.

Curator notes

Discovery status: institutional_trading_round3_promoted. Source links: AlphaAgent: LLM-Driven Alpha Mining with Regularized Exploration to Counteract Alpha Decay; Can Large Language Models Mine Interpretable Financial Factors More Effectively? A Neural-Symbolic Factor Mining Agent Model; Navigating the Alpha Jungle: An LLM-Powered MCTS Framework for Formulaic Factor Mining