Asks whether large language models can outperform prediction market consensus prices and argues the more tractable framing is using LLMs as updaters rather than predictors. Distinguishes cold prediction (generating a probability estimate without prior context) from updating (revising an existing estimate as new information arrives), and considers what each role implies for AI tools deployed alongside human traders in live markets.

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.

Frames prediction markets as a real-time information layer that complements traditional journalism by aggregating probabilistic forecasts from financially-incentivized participants. Argues that skin-in-the-game accountability produces more accurate signals than commentary-based analysis, with price movements often anticipating news before official announcements. Uses Polymarket and Kalshi as examples and acknowledges COVID-19 as a case where markets underperformed.

Analysis of 149 CPI prediction markets on Kalshi from 2021 to 2026 finds that trading volume explains less than 1% of variance in forecast accuracy, challenging the assumption that more liquidity improves market quality. Introduces Minimum Viable Liquidity (Cost of Expertise divided by Price Gap) as a framework for determining the threshold of liquidity needed to attract informed traders. Argues platforms should prioritize breadth over depth, running many thin markets rather than concentrating volume in few contracts.

Six-month empirical analysis of political prediction market quality across Kalshi and Polymarket. Finds only 1.3% of political markets are liquid enough to be manipulation-resistant, bid-ask spreads exceed 20% on most contracts, and only 53% of resolved US elections appeared on both platforms. Proposes a four-part blueprint: stock relevant questions, cross-subsidize political liquidity from sports profits, deploy AI market makers where human interest is insufficient, and standardize contract definitions across platforms.

Provides a quantitative framework for distinguishing gambling from systematic trading on prediction markets, including a five-point diagnostic and three trader archetypes classified by profitability. Explains why Polymarket's CLOB creates renewable structural arbitrage by design, and covers Kelly position sizing, adverse selection measurement via fill quality, and probability term structure as tools for building a repeatable edge.

Examines early results from the Prediction Arena experiment, where six AI models trade real money on Kalshi. Five of six are underwater after three weeks, suggesting that raw information processing isn't enough to generate edge. Argues that successful prediction market traders profit from embodied, local knowledge (monitoring flights, calling embassies) rather than synthesizing public information, a domain where AI remains fundamentally constrained.

Draws a parallel between prediction markets and Nielsen ratings to argue that coordination value matters more than accuracy. Points to Polymarket's Golden Globes and WSJ partnerships and Kalshi's CNN deal as signs that prediction markets are shifting from external forecasting tools to embedded institutional infrastructure. Once adopted as the shared reference point, displacement becomes nearly impossible regardless of methodological superiority.

Manifesto arguing binary yes/no prediction markets are incomplete—they flatten nuanced beliefs into coin flips and pay the same whether you were barely right or sharply right. Proposes distribution-native markets that reward precision: pay more for being closer to the actual outcome. Cites 130x volume growth from early 2024 to late 2025 as the category's credibility moment.

Examines Clinton and Huang's research on 2024 election market accuracy, finding PredictIt at 93%, Kalshi at 78%, and Polymarket at 67%, while also documenting significant cross-platform price divergences for identical contracts near Election Day. Raises concerns about Kalshi's media partnerships with CNN and CNBC, arguing they create incentives for sensational coverage of market movements and potential manipulation of thin markets.

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.

Challenges the conception that prediction markets serve as effective policy tools. Argues that conditional prediction markets proposing 'Will X lead to consequence Y?' fail to translate probability assessments into meaningful policy guidance despite being their central theoretical appeal.

Compares prediction markets with traditional polls and expert commentary along two axes: grassroots vs top-down and expertise density. Uses the 2024 Biden-Trump race to show how Polymarket priced in Biden's withdrawal probability while polls measured only head-to-head support.

Argues that prediction markets' unpopularity isn't due to regulation but fundamental demand-side issues. Markets need savers, gamblers, or sharps to function, but prediction markets attract none: they're zero-sum (no savers), have long resolution times (no gamblers), and are too small for professional traders (no sharps).

Investigates combinatorial prediction markets, which extend the standard model to support forecasts on conditional events (e.g., A given B) and Boolean combinations of events rather than only base events. Reports experimental results comparing combinatorial versus flat market structures on forecasting accuracy and calibration. Co-authored by Robin Hanson, whose LMSR underpins most automated prediction markets.