When Prediction Markets Break: Composition-Driven Systemic Risk in Large-Scale Agent Simulations
Chuan August Sun
Prediction markets promise to turn dispersed beliefs into useful probabilities, but that promise depends on the price-formation process remaining informative. We ask when it breaks in a market populated by automated agents that follow local rules and do not collude. The core measurement problem is that live markets do not reveal interim truth. We therefore build a controlled binary-market simulator with a known latent probability path, spanning a microstructure limit order book and scalable PyTorch batch auctions. The paper makes three contributions. First, it provides a testbed for measuring prediction-market failure against ground truth. Second, it shows that composition, not scale alone, is the clearest risk variable in the current evidence: as trend-following agents become more prevalent, prices steadily lose alignment with fundamentals, while matched scale controls rule out a simple “bigger is worse” story. Third, it introduces population-level diagnostics for market safety. Pressure-balance logs show how corrective arbitrage can be overwhelmed at the moments it is needed most; resonance, concentration, and synthetic crash-path tests show how markets can remain active while becoming less informative. The resulting evidence hierarchy reframes automated-market evaluation around strategy mix, synchronization, stabilizing capacity, concentration, and mechanism design, not only individual-agent behavior.