A new arXiv paper builds a stock market entirely out of AI trading bots - and finds it can flip from calm to chaotic with little warning.
Researchers built a simulated limit order book populated only by reinforcement-learning trading agents, then tested how the system behaves as the number of agents and the depth of visible orders change. They found distinct phase boundaries: below certain thresholds, trading stays orderly with normal price discovery, but cross them and the market tips into volatile cascade states. The team also studied market impact - how much a trade moves the price - and found it does not follow the classical square-root relationship traders typically assume. Instead, depending on feedback loops between agents, impact falls into one of three regimes: dissipative, balanced, or non-dissipative, where price moves compound instead of fading.
This matters because more real trading volume now comes from automated, increasingly AI-driven strategies, and knowing when a multi-agent system tips from stable to unstable is a practical risk question, not just an academic one. The breakdown of the standard square-root impact model also matters for anyone using it to estimate trading costs or set risk limits.
This is a simulation of agents trained against each other, not live markets, so the specific thresholds are a lab result, not a trading signal - but the phase-transition framing is a useful reminder that crash-like behavior can emerge from agent interaction alone, with no external shock required.