Self-improving alpha mining for quantitative trading via multi-agent large language models with knowledge base accumulation

Quantitative trading strategies depend on discovering novel alpha factors that capture market inefficiencies, yet traditional approaches rely on manual feature engineering, which is time-consuming and requires substantial domain expertise. This paper proposes an automated framework that leverages large language models (LLMs) to systematically generate, refine, and evaluate trading alpha strategies through a multi-agent pipeline comprising a WriterAgent, JudgeAgent, and BacktestEngine, organized in a nested-loop architecture. The key contribution of this work is the systematic integration of WorldQuant 101 Formulaic Alphas as a structured knowledge prior, rooted in proven formulaic patterns rather than unconstrained free-form synthesis. This design significantly narrows the search space, reduces invalid code generation, and accelerates convergence of the refinement loop. Experiments on Vietnamese stock market data demonstrate that the proposed framework consistently generates alpha factors with statistically significant Information Coefficient (IC) values and positive Sharpe ratios within a short generation time. Furthermore, the incorporation of WorldQuant 101 Formulaic Alphas consistently outperforms LLM-based generation without constraints in both alpha quality and convergence efficiency, establishing a scalable, robust, and domain-based approach to automated alpha mining.