FactorEngine (FE) is a novel program-level factor mining framework for quantitative investment that addresses key limitations of existing symbolic and neural approaches. Unlike prior methods that rely on bounded symbolic expressions or black-box neural networks, FE represents factors as Turing-complete Python programs, enabling complex control flows and higher-order feature interactions. The framework introduces a macro-micro co-evolution mechanism: macro-level logic mutations are guided by LLM agents using chain-of-experience (CoE) from historical trajectories, while micro-level parameter optimization is handled by Bayesian search on local computational resources. This separation of concerns significantly improves efficiency and avoids local optima. Additionally, FE incorporates a knowledge-infused bootstrapping module that extracts factor ideas from unstructured financial reports through a multi-agent pipeline (PDF processing, factor extraction, code generation), transforming domain knowledge into executable seed factors.

Extensive experiments on real-world CSI300 and CSI500 market data demonstrate that FE achieves state-of-the-art predictive and portfolio performance. With 400 iterations, FE-report achieves an IC of 0.0474 and excess annual return of 18.99% on CSI300, outperforming baselines like AlphaAgent and RD-Agent by substantial margins. FE also produces more diverse factors with lower redundancy and exhibits slower alpha decay over time. Ablation studies confirm the effectiveness of Bayesian micro-search, multi-island evolution, and CoE guidance. The framework is computationally efficient, with 99% executable ratio and low API costs. Overall, FactorEngine provides a scalable, interpretable, and high-performing solution for automated alpha factor discovery.