FactorMiner is a self-evolving agent framework for automated discovery of interpretable formulaic alpha factors in quantitative finance. It addresses three key challenges: the vast combinatorial search space of factor expressions, the lack of knowledge accumulation across mining sessions, and the need for interpretable signals. The framework combines a modular skill architecture that encapsulates domain knowledge (60+ financial operators, multi-stage validation pipeline) into reusable tools, with a structured experience memory that distills historical mining trials into successful patterns and forbidden regions. The agent operates via the Ralph Loop paradigm—retrieve, generate, evaluate, and distill—iteratively using memory priors to guide exploration, reducing redundant search while focusing on promising directions. A lightweight GPU-accelerated evaluation engine ensures computational efficiency.

Experiments on multiple datasets (CSI500, CSI1000, HS300, and cryptocurrency) demonstrate that FactorMiner constructs diverse, high-quality factor libraries with superior predictive performance. On CSI500, FactorMiner achieves a top-40 factor IC of 8.25% and ICIR of 0.77, outperforming baselines including AlphaAgent (5.90% IC) and GPLearn (6.04% IC). The factor library maintains low redundancy (average pairwise correlation 0.203) even as it scales to 110 factors. FactorMiner also shows strong cross-market generalization, consistently achieving the highest IC and ICIR across all tested markets. Ablation studies confirm the critical role of experience memory in improving search efficiency and factor quality. The released open library of 110 interpretable A-share equity factors provides both research insights and practical value for quantitative trading.