Contract-Level Informational Efficiency in Prediction Markets: Measurable Ex Post, Not Predictable Ex Ante — Evidence from 244 Kalshi Markets

I study informational efficiency at the level of the individual contract in prediction markets, using a frozen dataset of 244 resolved Kalshi markets (21 distinct events, 2024–2026) in high-emotion categories. Three sets of results emerge. First, naive trading strategies fail: contrarian mean-reversion loses significantly even under event-clustered inference (mean return per unit staked −0.38, 95% CI [−0.67, −0.05]; −0.49 with taker fees and bid–ask execution), momentum is statistically indistinguishable from zero yet ruinous under Kelly compounding, and the market price itself is a strong benchmark (Brier 0.072 at two-thirds of contract life versus 0.156 for the climatological forecast; skill +0.54). Second, efficiency is measurable ex post: I construct a composite contract-level efficiency index (convergence skill, variance-ratio resistance, random-walk-normalized signal-to-noise), whose category ranking is robust to the choice of weights (Spearman ≥ 0.95 across nine weighting schemes). Third, efficiency is not predictable ex ante: none of 13 canonical efficiency metrics predicts out-of-sample under event-grouped cross-validation and Bonferroni correction — the single apparent signal is shown to be a mechanical artifact, both by synthetic-market controls and in closed form (expected calibrated surprise equals 2√(p(1−p))) — and none of 18 pre-registered indicator–target tests for regime transitions passes. Exploratory analyses indicate that contract-level efficiency is multidimensional (first principal component: 40% of variance) and continuously distributed (no bimodality), consistent with the Adaptive Markets Hypothesis. Code, data, and the pre-registration protocol are available for full reproduction.