Challenges the smart-versus-dumb money dichotomy in prediction markets by synthesizing research from Snowberg/Wolfers/Zitzewitz, INSEAD's BIN model, and Wharton's cognitive search framework. Argues that noisy traders fund the probability space rather than serve as exit liquidity, and compares how binary CLOBs versus continuous probability markets decompose and harness noise differently.

Reviews Philip Tetlock's Superforecasting and draws a direct line from the book's core thesis — that forecasting skill is measurable, trainable, and outperforms expert punditry — to Polymarket's success during the 2024 US election. Explains Tetlock's key concepts (foxes vs hedgehogs, the Good Judgment Project, Brier scores, calibration) and argues that Polymarket effectively operationalized Tetlock's framework at scale by converting crowd forecasting into a liquid financial market.

Fits a Bayesian hierarchical model to 292 million trades across 327,000 contracts on Kalshi and Polymarket to decompose calibration errors into structured components: universal horizon effects, domain-specific biases, and trade-size scale effects, which together explain 87.3% of variance on Kalshi. Finds persistent underconfidence in political markets where prices compress toward 50%, and shows that large trades amplify this pattern on Kalshi but not on Polymarket, pointing to platform-specific microstructure differences.

Questions whether prediction markets are capturing the right signal. Argues binary yes/no markets flatten complex beliefs into coin flips, losing the precision that separates superforecasters from average predictors. Uses the 2024 French trader whale ($30M moving election odds) and a Vanderbilt study (PredictIt's 93% accuracy vs 67% on high-volume platforms) to argue that more liquidity doesn't mean better signal.

Presents ForecastBench, a benchmark tracking how well LLMs forecast real-world outcomes against superforecasters and crowd forecasters. The best LLM (GPT-4.5) achieves a Brier score of 0.101 versus superforecasters' 0.081, with LLMs improving roughly 0.016 Brier points per year, projecting parity by late 2026. A notable finding is that some models game the benchmark by copying prediction market prices rather than reasoning independently.