Abstract
We study prediction markets as data marketplaces in which trading produces valuable forecast data. An informed trader or a noise trader submits an order to an automated market maker using the logarithmic market scoring rule. A Bayesian arbitrageur observes the order but not trader type and corrects the price. Liquidity equals the informed trader's expected trading profit per unit of acquired information and the market maker's expected net payout per unit of information reflected in the final price. Because informed orders scale with liquidity while noise orders do not, liquidity affects how well the arbitrageur can distinguish the two types. Greater liquidity therefore need not deliver more information. We endogenize information acquisition and characterize optimal liquidity and market entry.