The paper introduces CogAlpha, a novel framework for automated alpha mining that combines large language model (LLM) reasoning with evolutionary search. Unlike prior methods that rely on formulaic or neural approaches with limited search scope, CogAlpha represents alphas as executable code and uses a seven-level agent hierarchy to explore diverse financial concepts—from market structure to geometric patterns. A multi-agent quality checker ensures code validity, logical consistency, and economic interpretability. The framework then applies thinking evolution, where LLMs perform mutation and crossover operations on alpha codes, guided by fitness metrics (IC, ICIR, RankIC, RankICIR, Mutual Information) and adaptive generation that incorporates feedback from previous iterations.

Experiments on five datasets from three stock markets (China, US, Hong Kong) demonstrate that CogAlpha consistently discovers alphas with superior predictive accuracy, robustness, and interpretability compared to 21 baselines including machine learning models, deep learning models, existing alpha libraries, and other LLM-based methods. On the CSI300 dataset, CogAlpha achieves an IC of 0.0591 and IR of 1.8999, significantly outperforming all competitors. The framework also produces interpretable alphas with detailed comments explaining their economic rationale, and shows strong generalization across different markets, training methods, and prediction horizons. The work opens a new direction for cognitive alpha mining by aligning evolutionary optimization with LLM-based reasoning.