Using 41.6 million trades, we measure adverse selection in prediction markets. Adapting Kyle’s λ and the Glosten-Harris decomposition, we show single-name markets exhibit greater informed price impact than broad-based markets. Despite this, effective spreads are only modestly wider, and market makers earn twice as much per contract. A frequency-magnitude decomposition resolves this puzzle: traders systematically overbet YES in markets that predominantly settle NO, generating a behavioral surplus that cross-subsidizes adverse selection. Adapting the VPIN toxicity metric, we find that one-sided order flow predicts maker losses in single-name markets but not broad-based markets. Our research suggests a new microstructure equilibrium concept for bilateral settlement markets.