FactorEngine: A Program-level Knowledge-Infused Factor Mining Framework for Quantitative Investment

In one sentence

Summary

The paper addresses the limitations of existing alpha mining methods, which suffer from bounded expressiveness in symbolic approaches or poor interpretability and overfitting in neural models. FactorEngine (FE) casts factors as Turing-complete Python programs, enabling complex control flows and higher-order feature interactions. The framework introduces three key separations to improve efficiency: logic revision versus parameter optimization, LLM-guided directional search versus Bayesian hyperparameter search, and LLM usage versus local computation. This allows the system to scale effectively while maintaining the interpretability required for auditable quantitative strategies. The methodology comprises a bootstrapping module that transforms unstructured financial reports into executable factors via a closed-loop multi-agent pipeline, and an evolution module that performs macro-micro co-evolution. The macro level uses LLM agents to propose structural mutations guided by a chain of experience, while the micro level employs Bayesian optimization for parameter tuning. Experiments are conducted on CSI300 and CSI500 indices using OHLCV data from 2017 to 2024, comparing FE against baselines like Alpha158, GPlearn, AlphaAgent, and RD-Agent. The evaluation includes predictive metrics (IC, ICIR) and portfolio performance (AR, Sharpe, MDD) under realistic trading constraints including transaction costs and liquidity limits. Results show that FE significantly outperforms baselines, with FE-report achieving a 58% improvement in IC and a 126% increase in annual excess return compared to Alpha158. The framework demonstrates superior predictive stability and portfolio impact, particularly in the CSI500 market. However, the study acknowledges that modern LLMs may have been trained on data extending beyond the test period, a common limitation in agent-based methods. Additionally, the performance gains are contingent on the quality of the initial knowledge infusion and the computational resources required for the multi-agent evolution process.

Research problem

The paper addresses the challenge of automated alpha factor mining in quantitative investment, specifically targeting the trade-off between model expressiveness, interpretability, and computational efficiency. Existing symbolic methods are limited by predefined operator spaces, while neural networks lack transparency and are prone to overfitting. The research aims to discover predictive signals that are directly executable, auditable, and robust to regime shifts, while remaining computationally tractable at scale.

Core contributions

• Introduction of FactorEngine (FE), a program-level factor discovery framework that treats factors as Turing-complete code.
• Development of a macro-micro co-evolution mechanism that separates logic evolution (LLM-guided) from parameter optimization (Bayesian search).
• Design of a knowledge-infused bootstrapping module that converts unstructured financial reports into executable factor programs via a multi-agent pipeline.
• Implementation of an experience knowledge base that supports trajectory-aware refinement, including learning from evolutionary failures.
• Demonstration of state-of-the-art predictive and portfolio performance on real-world Chinese A-share market data.

Method

• Bootstrapping Module: LLM-based extraction of financial ideas from reports into executable Python code.
• Evolution Module: Macro-micro co-evolution with LLM agents for logic mutation and Bayesian search for parameter tuning.
• Chain of Experience: Historical trajectory data used to guide LLM idea generation and avoid redundant exploration.
• Integration Module: Elite node selection and multi-factor modeling using a weighted fitness score.
• Multi-island Evolution: Parallel evolution processes with migration to enhance diversity and efficiency.

Results & findings

• FE-report achieved an IC of 0.0474 and annual return of 18.99% in CSI300, outperforming all baselines.
• 58% improvement in IC and 126% increase in excess annual return compared to Alpha158 with report-derived factors.
• FE consistently outperformed AlphaAgent and RD-Agent across different iteration counts.
• Factors evolved from financial reports showed significantly better performance than those from manual factors.
• FE demonstrated superior predictive stability and lower maximum drawdown compared to neural and symbolic baselines.

Data & experiments

• Datasets: CSI300 and CSI500 indices using OHLCV data from Qlib, with training (2008-2014), validation (2015-2016), and testing (2017-2024) periods.
• Baselines: GPlearn, LightGBM, LSTM, Transformer, TRA, AlphaAgent, RD-Agent, and hand-crafted Alpha158.
• Experiments: Two setups (FE-alpha and FE-report) with 200 and 400 evolution iterations.
• Metrics: Information Coefficient (IC), ICIR, Rank IC, Rank ICIR, Annual Return (AR), Sharpe Ratio (SR), and Maximum Drawdown (MDD).
• Backtesting: Realistic trading constraints including transaction costs, slippage, and liquidity limits.

Limitations & caveats

• Potential data leakage as LLMs may have been trained on data beyond the test period.
• Performance is dependent on the quality and relevance of the initial financial reports.
• Computational cost of multi-agent evolution may be high for large-scale applications.
• Results are specific to Chinese A-share markets, limiting generalizability to other markets.
• LLM-generated factors may still be subject to overfitting if not properly regularized.

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