Agent Policy-Value Audit: Separating Transition Composition from Event Selection in Financial LLM Agents

Battle Count: 50

Highly relevant for quantitative researchers evaluating LLM-based trading agents. It provides a rigorous statistical framework to distinguish between alpha generated by market exposure (beta/style) and true stock-picking skill (alpha), addressing a critical flaw in current end-to-end backtesting methods.

The core audit logic (randomization inference with fixed transition counts) is conceptually straightforward but requires careful implementation of pooling, stratification, and clustered bootstrapping. Integrating it into an existing LLM agent pipeline requires logging detailed transition types and eligible event pools.

The paper states that all reported tables and figures are reproducible from de-identified intermediate data. Upon acceptance, the authors will release the benchmark generator, analysis code, frozen prompt hashes, OSF preregistration record, deviation log, and per-decision traces. WRDS inputs are not redistributed due to licenses.