Cognitive Alpha Mining via LLM-Driven Code-Based Evolution

In one sentence

Summary

The paper addresses the challenge of discovering robust, interpretable alpha factors in high-dimensional financial data, where existing neural and symbolic methods suffer from opacity, fragility, or poor generalization. The authors propose CogAlpha, a framework that treats LLMs as adaptive cognitive agents to perform structured, human-like exploration of the alpha search space, balancing logical consistency with creative innovation. This approach aims to move beyond shallow pattern replication toward deeper reasoning and economically grounded factor discovery. CogAlpha employs a seven-level agent hierarchy to generate initial alpha candidates from OHLCV data, covering domains from market cycles to geometric fusion. A multi-agent quality checker validates code syntax, logic, and economic meaning, while a fitness evaluation filters candidates using IC, RankIC, and other metrics. The core innovation is the 'Thinking Evolution' module, which uses LLM-driven mutation and crossover operations to iteratively refine qualified alphas. Experiments on five datasets from China, the US, and HK markets demonstrate that CogAlpha consistently discovers alphas with superior predictive accuracy and robustness compared to 21 baselines, including deep learning models and other LLM-based methods. The framework achieves state-of-the-art performance, with CogAlpha-generated alphas showing significantly higher IC and RankIC values than baselines like LightGBM and Alpha158. The generated factors are accompanied by detailed comments and code, ensuring interpretability and economic grounding. However, the method relies heavily on the reasoning capabilities of large LLMs, which may incur high computational costs. The study also notes that while generalization is strong across markets, the specific performance may vary with different prediction horizons and training methods, suggesting potential sensitivity to market regime shifts.

Research problem

Discovering effective predictive signals (alphas) from financial data is difficult due to high dimensionality and low signal-to-noise ratios. Existing methods, including neural networks and genetic programming, often produce opaque, fragile, or economically ungrounded factors that generalize poorly. There is a need for a framework that enables broad, structured, and human-like exploration of the alpha search space, balancing logical consistency with creative leaps to produce robust and interpretable factors.

Core contributions

• Introduction of the Cognitive Alpha Mining Framework (CogAlpha), which combines code-level alpha representation with LLM-driven reasoning and evolutionary search.
• Development of a seven-level agent hierarchy that systematically explores diverse alpha-discovery directions from macroscopic market structures to microscopic geometric patterns.
• Design of a multi-agent quality checker that validates alpha code for syntax, logic, economic meaning, and numerical stability before inclusion in the candidate pool.
• Implementation of a 'Thinking Evolution' module that uses LLM-driven mutation and crossover to iteratively refine and recombine alpha candidates, inducing deeper reasoning.
• Demonstration of superior predictive accuracy, robustness, and generalization of CogAlpha across five stock datasets from three different global markets compared to 21 baselines.

Method

• Seven-level agent hierarchy generates initial alpha candidates from OHLCV data.
• Multi-agent quality checker validates code syntax, logic, and economic meaning.
• Fitness evaluation filters candidates using IC, RankIC, ICIR, RankICIR, and MI metrics.
• Thinking Evolution module performs LLM-driven mutation and crossover on qualified alphas.
• Adaptive generation incorporates feedback from valid and invalid alphas into prompts.
• Diversified guidance uses five paraphrasing modes to enrich exploration.
• LightGBM is used to train and evaluate the predictive power of generated alphas.

Results & findings

• CogAlpha outperforms all 21 baselines on the CSI300 dataset, achieving an IC of 0.0591 and RankIC of 0.0814.
• The framework consistently shows superior performance across different markets (China, US, HK) and prediction horizons (10-day, 30-day).
• Ablation studies confirm that all components of CogAlpha contribute to improved alpha mining effectiveness.
• Generated alphas exhibit strong interpretability, with detailed comments explaining the underlying logic and economic intuition.
• The evolutionary process successfully refines initial alphas, improving metrics like IC and RankIC through mutation and crossover.
• Closed-source LLMs (GPT-4.1, o3) performed poorly compared to open-source models and CogAlpha, highlighting the importance of the evolutionary framework.
• CogAlpha demonstrates robustness against overfitting, with alphas maintaining predictive power in out-of-sample tests.

Data & experiments

• Experiments conducted on five stock datasets: CSI300, CSI500 (China), S&P500 (US), HSI, and HSCI (HK).
• Primary prediction target is 10-day return, with additional tests on 30-day horizons.
• Comparison against 21 baselines including machine learning models (LightGBM, XGBoost), deep learning models (LSTM, Transformer), alpha libraries (Alpha158, Alpha360), and other LLM-based methods (AutoAlpha, AlphaAgent).
• Evaluation metrics include Information Coefficient (IC), RankIC, ICIR, RankICIR, Mutual Information (MI), Information Ratio (IR), and Annualized Excess Return (AER).
• Ablation studies performed to assess the contribution of each component: Adaptive Generation, Diversified Guidance, Seven-Level Agent Hierarchy, and Thinking Evolution.

Limitations & caveats

• The method relies on large LLMs, which may incur high computational costs.
• The performance may vary depending on the specific market conditions and regimes.
• The framework's effectiveness is demonstrated on stock markets; generalization to other asset classes is not explored.
• The quality of generated alphas depends on the reasoning capabilities of the underlying LLM.
• The study does not address the long-term sustainability of discovered alphas in live trading environments.

Curator notes

Discovery status: institutional_trading_round3_promoted. Source links: AlphaAgent: LLM-Driven Alpha Mining with Regularized Exploration to Counteract Alpha Decay; Navigating the Alpha Jungle: An LLM-Powered MCTS Framework for Formulaic Factor Mining