A market can be rational without looking human

Thomas Henning, Siddhartha M. Ojha, Ross Spoon, Jiatong Han, and Colin F. Camerer put language-model agents into an experimental asset market in LLM Agents Do Not Replicate Human Market Traders: Evidence From Experimental Finance. The market had a single risky asset with a known, constant fundamental value, and the traded price emerged from participants' orders rather than an external price feed.

The authors ran 30 trading periods per market. In single-model markets, 20 copies of one model traded together; a separate “Battle Royale” treatment mixed models. Six commercial model families were tested, alongside 19 human-only markets under matching market rules. Agents were asked to maximize profit and could keep notes between periods.

Similar orders did not create similar bubbles

Claude 3.5 Sonnet and GPT-4o prices stayed close to the asset's fundamental value. Other models departed from it in different ways: the authors report bubble-and-crash cycles for Grok-2, a more sustained run-up for GPT-3.5, and below-value trading before recovery for Gemini 1.5 Pro. Human markets, meanwhile, consistently produced larger bubbles. The model agents therefore did not reproduce a prominent group-level pattern from this particular human experiment, even when the models interacted with one another.

That is a meaningful difference for researchers who hope to use model agents as stand-ins for market participants. A model can follow a profit-maximizing instruction and still fail to reproduce the price dynamics generated by human groups.

This is a laboratory asset, not a live market

The paper uses one asset, a short horizon, frozen model weights, and a specified prompt. The authors note that model agents do not experience human incentives or online learning; longer horizons, short selling, and richer strategic interaction remain open questions. Its result does not predict price formation for actual financial assets.

Bibliography

  • Thomas Henning, Siddhartha M. Ojha, Ross Spoon, Jiatong Han, and Colin F. Camerer (2025). LLM Agents Do Not Replicate Human Market Traders: Evidence From Experimental Finance. arXiv:2502.15800v3 [q-fin.TR]. Full paper (PDF).