QuantaAlpha: An Evolutionary Framework for LLM-Driven Alpha Mining

Battle Count: 73

This paper is directly and highly relevant to quantitative trading. It addresses the core problem of alpha factor mining, proposes a novel evolutionary framework for systematic factor discovery, demonstrates strong performance on major equity indices (CSI 300, CSI 500, S&P 500), includes practical backtesting with transaction costs, addresses factor crowding and alpha decay, and shows cross-market transferability. The framework is designed for real-world quantitative research workflows with interpretable factor expressions and portfolio construction strategies.

The framework involves multiple components: multi-agent orchestration, symbolic AST-based factor construction, LLM-based consistency verification, evolutionary operators (mutation/crossover), diversified planning initialization, complexity/redundancy gating, and integration with Qlib backtesting. It requires access to multiple LLM APIs, a standardized operator library, and careful parameter tuning. The 20-hour runtime and 1.8M token consumption per run indicate significant computational requirements. However, the modular design and available GitHub code reduce some implementation barriers.

Code is available on GitHub (https://github.com/QuantaAlpha/QuantaAlpha). The paper provides detailed experimental setup including data splits, backtesting configuration, operator library, factor constraints, and trading strategy parameters. However, full reproducibility depends on access to specific LLM APIs (GPT-5.2, DeepSeek-V3.2, etc.) and the Qlib framework. Computational cost is reported (~20 hours, ~1.8M tokens per run). The paper includes comprehensive ablation studies and cross-seed variance analysis.