Reproduce “Agentic Empirical Asset Pricing: Methodological Foundations” on OpenResearch

About this paper

Yingjian Pan, Xiaowei Ding, Kay Giesecke

This research provides a foundational framework for Agentic Empirical Asset Pricing (AEAP), defining its core components and proposing a multi-axis evaluation methodology that assesses not only the performance of discovered factors but also the adaptivity of the autonomous discovery system itself. The work introduces SEADS, a concrete AEAP system, which demonstrates how such systems can achieve competitive out-of-sample Sharpe ratios and highlights the critical role of continuous re-discovery for sustained performance.

Key insights

• Evaluating the autonomous discovery *system* through rolling re-execution, rather than just its static outputs, is crucial for assessing adaptivity and ensuring reliability in finance, with SEADS's rolling arm outperforming its static counterpart significantly.
• The Planner and Self-evolution mechanisms within AEAP systems are load-bearing, significantly contributing to the quality and quantity of admitted factors, with their removal leading to substantial performance degradation.
• Agentic systems like SEADS can identify more complex factor structures (e.g., multiplicative interactions, ratios) than some baselines, suggesting a pathway to discovering genuinely novel and non-redundant signals that add incremental value to portfolios.

Summary by alphaXiv.