Research Project

Hypothesis Discovery with AI Research Agents
in Accounting and Finance

Project Overview

This project introduces an autonomous AI research agent for quantitative signal discovery inside a human-designed research laboratory. Using structured accounting data, a constrained hypothesis language, and a fixed empirical testing pipeline, the system generates interpretable asset-pricing signals, evaluates them on historical data, and iteratively refines subsequent generations of ideas based on the evidence.

More broadly, this research agenda asks how AI research agents can be embedded in rigorous scientific architectures to expand the frontier of finance and accounting research. The agenda focuses on building transparent, auditable systems that make discovery more systematic while preserving economic reasoning, empirical discipline, and human judgment.

Key Themes

Agentic Hypothesis Generation

We study how autonomous AI research agents can generate, test, and iteratively refine interpretable financial hypotheses inside a structured discovery process.

Accounting and Finance Domain

Grounded in fields that combine rich economic structure, structured and textual data, and well-defined empirical tests that make disciplined AI-guided discovery possible.

Human–AI Collaboration

Human researchers design the laboratory—defining the hypothesis space, data, and evaluation rules—while AI research agents work autonomously inside that environment to explore, test, and refine ideas.

Paper

“Can AI Do Financial Research? LLM-Guided Hypothesis Discovery in Asset Pricing”
Working Paper
with Huan Liu, Zhizhe Liu, and Danqing Mei