LLM Agents Do Not Replicate Human Market Traders: Evidence from Experimental Finance

Thomas Henning, Siddhartha Ojha, Ross Spoon, Jiatong Han, Colin F Camerer

Keywords: Experimental Finance, Behavioral Finance, Strategic Behavior, Financial AI, Experimental Economics, Finance

TL;DR: We compare LLM performance to human performance in a experimental trading paradigm.

Abstract: In this study, we compare Large Language Models (LLMs) with human traders in a classic experimental-finance paradigm where prices are determined endogenously. Using a well-established asset-trading design, we run homogeneous markets with single-model LLM agents and heterogeneous “battle-royale” markets with multiple LLM models. Our findings reveal that LLMs generally exhibit a “textbook-rational” approach, pricing the asset near its fundamental value and showing only a muted tendency toward bubble formation, while humans deviate substantially and generate bubbles consistently. Additional treatments, including dividend shocks and repeated-exposure/experienced runs, show that these differences persist across various experimental settings. Further analyses of LLM-generated strategy text indicate lower variance, reduced bias, and stronger reliance on fundamentals relative to humans’ more heuristic-driven trading. These results highlight the risk of using LLM-only agents to model human-driven market phenomena, as key behavioral features such as large, emergent bubbles are not reproduced.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 15485