Research
Researchers put autonomous LLM agents through classic experimental asset-market auctions and found they reproduce the same disposition effect and recency bias human traders show, and that specific prompt wording can be tuned to deliberately amplify or suppress how large the resulting bubble gets
Shumiao Ouyang and Pengfei Sui's study scored twenty distinct behavioral mechanisms in LLM trading agents' reasoning and found the same excess-demand-predicts-price and disagreement-predicts-volume relationships that decades of human experimental finance has documented, with the added finding that these tendencies are not fixed but causally steerable through prompting.
Experimental asset market auctions are one of behavioral finance's oldest tools, used since the 1980s to show that human traders reliably form bubbles even when everyone in the room knows the asset's fundamental value in advance, because the bubble comes from the trading dynamic itself rather than from mispriced information. Ouyang and Sui ran the same setup with autonomous LLM agents standing in for the human subjects, and the headline finding is not that the agents avoided the trap, it is that they walked into the same one, exhibiting a pronounced disposition effect, holding losing positions too long, and recency-weighted extrapolative beliefs, overweighting whatever just happened in the last few price ticks.
The market-level result is arguably the more important one for anyone thinking about what happens when agents like this trade against each other at scale. The individual biases aggregated into the same equilibrium patterns classic experimental markets show in humans: excess demand predicting future prices, and disagreement between traders predicting how much volume changes hands. That is a specific, falsifiable claim, and it is the reason this paper is being read as more than a curiosity. It says the bubble-forming dynamic in these markets is not a peculiarity of human psychology that AI agents would sidestep by being more rational. It is a structural feature of how any boundedly-rational trader, artificial or human, behaves inside that auction mechanism.
The part that should concern anyone deploying LLM trading agents in production is the causal intervention result. The researchers built a scoring framework across twenty distinct behavioral mechanisms in the agents' own reasoning traces and then tested whether targeted prompt changes could move those scores. They could. Specific prompt interventions causally amplified or suppressed individual behavioral mechanisms and, through them, the size of the resulting market bubble. That means an LLM trading agent's propensity to chase momentum or hold losers too long is not a fixed trait determined by the underlying model. It is at least partly a function of how the agent was prompted, which is a parameter every deploying firm controls and every deploying firm can get wrong without realizing it.
Put those two findings together and the practical implication is uncomfortable: if bubble-forming behavior in LLM trading agents is prompt-steerable, then two firms running the same underlying model with different system prompts could produce meaningfully different aggregate market behavior, and neither firm would necessarily know it was happening from the outside. That is a different and harder-to-regulate risk than the usual worry about agents colluding or correlating on signal. It is agents behaving differently, and predictably steerable in that difference, purely as a function of prompt engineering choices made behind closed doors at each firm.
Read the original: arXiv (Ouyang & Sui) - Researchers put autonomous LLM agents through classic experimental asset-market auctions and found they reproduce the same disposition effect and recency bias human traders show, and that specific prompt wording can be tuned to deliberately amplify or suppress how large the resulting bubble gets. Commentary is the independent editorial view of Share Trading; the original article is credited to its publisher.