AI agents replicate classic market bubble dynamics in simulated auctions
A study by Shumiao Ouyang and Pengfei Sui finds that LLM agents in simulated open-call auction markets exhibit classic behavioral finance patterns — including the disposition effect and recency-weighted beliefs — that aggregate into market bubble dynamics matching landmark experimental findings.
Score breakdown
Practitioners building or deploying LLM-based trading agents should note that prompt design directly influences behavioral biases and can significantly amplify or dampen market bubble dynamics.
- 01Authors Shumiao Ouyang and Pengfei Sui study LLM agents trading in a simulated open-call auction market.
- 02AI agents exhibit a pronounced disposition effect and recency-weighted extrapolative beliefs.
- 03Individual-level behavioral patterns aggregate into market-level bubble dynamics.
Shumiao Ouyang and Pengfei Sui investigate how autonomous Large Language Model (LLM) agents behave in simulated experimental asset markets, using an open-call auction populated entirely by AI agents. The paper documents three core findings: AI agents exhibit a pronounced disposition effect (the tendency to sell winners too early and hold losers too long) alongside recency-weighted extrapolative beliefs, both of which are well-established patterns in human behavioral finance.
At the aggregate level, these individual biases compound into market-level dynamics that closely mirror the classic experimental results of Smith et al.
At the aggregate level, these individual biases compound into market-level dynamics that closely mirror the classic experimental results of Smith et al. (1988) — a landmark study of human-driven asset market bubbles. Specifically, the simulated markets reproduce the predictive power of excess demand for future prices and the positive relationship between trader disagreement and trading volume.
The paper's third contribution is methodological: the authors analyze agents' reasoning text through a twenty-mechanism scoring framework, then apply targeted prompt interventions to causally amplify or suppress specific behavioral mechanisms. This approach demonstrates that prompt design can meaningfully alter the magnitude of market bubbles, with direct implications for how LLM-based trading agents might be configured or constrained in practice.
Key facts
- 01Authors Shumiao Ouyang and Pengfei Sui study LLM agents trading in a simulated open-call auction market.
- 02AI agents exhibit a pronounced disposition effect and recency-weighted extrapolative beliefs.
- 03Individual-level behavioral patterns aggregate into market-level bubble dynamics.
- 04The simulated results replicate classic findings from Smith et al. (1988) on experimental asset markets.
- 05Excess demand is shown to have predictive power for future prices in the simulated markets.
- 06A positive relationship between trader disagreement and trading volume is observed.
- 07Targeted prompt interventions, analyzed via a twenty-mechanism scoring framework, can causally amplify or suppress specific behaviors and alter bubble magnitude.