Propose, Don’t Judge: An Anytime-Valid Referee for LLM Agents That Mine Investment Factors

Battle Count: 0

Highly relevant. It addresses the core problem of overfitting and false discovery in automated factor mining, a major issue in quantitative finance. It provides a rigorous statistical framework for validating LLM-generated trading signals.

High complexity. Requires implementing anytime-valid statistical procedures (e-processes, online e-BH), integrating them with an LLM agent loop, and managing point-in-time data pipelines. The theoretical underpinnings are advanced.

Code, run manifests, prompts, and recorded model calls are available from the corresponding author. Market data is licensed from a commercial vendor and cannot be redistributed, but the methodology is fully described with synthetic and real-data experiments.