First-hand information for everyone
Researchers simulated a simple exchange populated entirely by autonomous large language model (LLM) agents to study how AI forms price expectations and trades. They report three main findings: the AI agents show well-known human behavioral patterns, these individual behaviors sum up to familiar market dynamics, and carefully written prompt changes can increase or decrease the size of asset-price bubbles.
The experiment used a controlled open-call auction with 20 main trading periods. Each agent started with 100 units of cash and 4 shares of a risky asset. Cash paid a fixed 5% interest per period. The risky asset paid a dividend that was either 0.4 or 1.0 with equal chance, so the expected dividend was 0.7. The designers also included a terminal buyout that kept the asset’s fundamental value constant at 14 units (the expected dividend divided by the 5% interest). The setup made it easier to identify when prices deviated from rational value.
At the individual level, the LLM agents reproduced two classic biases. They displayed a disposition effect — selling assets that had gone up in price while holding on to losers — and they formed recency‑weighted extrapolative beliefs, meaning recent price moves had outsize influence on their forecasts of future prices. Unlike many human studies that find a weak link between stated beliefs and actual trades, these AI agents translated their stated expectations much more directly into trading decisions, because they were not constrained by human frictions like attention lapses or transaction costs.
When many biased agents traded together, market outcomes matched classic experimental results. Measures of excess demand predicted future price changes, and higher disagreement across agents’ beliefs was associated with more trading volume. The researchers could read the agents’ internal reasoning text and found that, during bubble episodes, agents explicitly described speculative strategies such as momentum chasing and “riding the bubble.” This link between text and action helped connect cognitive mechanisms to price dynamics.
The team developed a twenty‑mechanism scoring framework to classify the agents’ reasoning. Using that framework, they ran causal prompt interventions: targeted changes to the instructions given to agents that were designed to suppress or amplify particular behavioral mechanisms. Suppressive prompts reduced the measured bubble size, while contrasting prompts that encouraged the same biases made bubbles larger. The results show that agent behavior — and aggregate market outcomes — can be altered by changing how AI agents are instructed.
Important caveats remain. The study is a laboratory simulation with a specific market design: a fixed dividend process, a terminal buyout to hold fundamental value constant, and a selected set of fourteen LLMs. Real financial markets include many other actors, information flows, costs, and institutional rules that this experiment does not model. The authors also note that LLM biases may stem from the human‑generated data used in training, a hypothesis rather than a proven mechanism. The paper itself flags broader interpretive limits and therefore stops short of claiming direct, immediate implications for live markets or regulation.