AlphaSchema Enhances LLM-Based Alpha Mining with Structured Semantic Exploration

Automated alpha mining, which involves discovering predictive trading factors, increasingly utilizes large language model (LLM) agents. However, existing LLM-based systems often allow agents to both construct and search for factors without a clear exploration space or a systematic navigation mechanism. This leads to implicit and difficult-to-control exploration. To address this, AlphaSchema proposes a novel framework that constructs and explores a structured space of trading semantics for alpha mining. AlphaSchema defines each point in this semantic space as a "schema plan," comprising elements like Event, Context, Qualities, Direction, and Output, which explicitly specify the semantics of a potential trading factor before it is implemented. This approach separates the exploration of factor ideas from their actual implementation. An LLM then translates selected schema plans into executable factors, and the evaluated rewards are used to train a surrogate model over the semantic space. An iterative selection mechanism balances global exploration, surrogate-guided exploitation, and local mutation to discover high-performing factor pools. Experiments on the Chinese stock market demonstrate AlphaSchema's effectiveness in finding factors with strong predictive and portfolio performance, showing that its semantic search navigates diverse regions while focusing on high-reward areas.