Persona Design and Bubble Formation in LLM-Populated Experimental Asset Markets

Asset price bubbles in laboratory experiments are robust but difficult to study at scale; human sessions are expensive, slow, and limited to short horizons. We ask whether markets populated by large language model (LLM) agents can reproduce the shape of human bubble dynamics, comparing three ways of building agent "personas": theory-driven archetypes curated by the authors, data-derived personas clustered from real human traders, and literature-grounded personas extracted from canonical bubble theories. Data-derived personas, the closest analogue to "what humans actually do," fail to bubble, even though the humans they are drawn from bubble to 2--3$\times$ fundamental value. Literature-grounded personas inflate dramatically but never crash. Only the theory-driven archetypes produce the canonical inverted-U of a bubble and crash, and the amplitude depends on the model. On the stronger Claude models the peak sits inside the human range ($\approx 2$--$3\times$ FV) and the market crashes cleanly, while on GPT-5.4 the same personas inflate several times higher and often do not crash within the horizon. Within the theory-driven family, a sweep over population composition separates the bubble's rise from its fall. Bubbles form readily, climbing to roughly $20\times$ FV even when the population contains a rational arbitrageur meant to sell against overpricing, whose corrective pressure is not enough to stop them. The crash is what requires a specific ingredient. A market returns to fundamentals only when some persona is instructed to exit before the terminal date, and without one the bubble never comes down. Extending to horizons infeasible with humans ($T=50$ and $T=100$), the crash at $T=100$ arrives well before the horizon reminder we inject, suggesting attention to the endpoint may emerge on its own. We position LLM markets not as a substitute for human experiments but as a low-cost sandbox for testing and pre-registering behavioral hypotheses before committing to expensive human runs.