Prediction Markets Are Not New
The idea of using markets to aggregate predictions is not a 21st-century innovation. Political betting markets existed in the United States as early as the 1860s, when Wall Street curb exchanges took bets on presidential elections. These markets were widely followed and often more accurate than contemporary straw polls. The Iowa Electronic Markets, launched in 1988, provided some of the first academic evidence of prediction market accuracy in modern form.
The Academic Track Record
Berg, Nelson, and Rietz at the University of Iowa published multiple studies comparing Iowa Electronic Markets predictions to polling data across presidential elections from 1988 through 2004. Markets outperformed polls 74% of the time when compared on the same dates. The advantage was particularly pronounced when forecasts were made more than 100 days before the election.
A 2023 meta-analysis in the International Journal of Forecasting by Atanasov and colleagues examined prediction markets, polls, and expert forecasters across multiple domains. They found that the best forecasts generally combine all three sources, with markets performing best when the question is well-defined, the resolution criteria are clear, and liquidity is adequate.
Where Markets Have Failed
Prediction markets are not infallible, and their failures are instructive. Markets systematically underestimate the probability of extreme events (black swans) because the base rate for extreme events is very low and most market participants have not experienced them. Markets also struggle with events where resolution depends on a single decision-maker whose thought process is genuinely private (will a specific CEO resign, will a specific judge rule a certain way).
The 2024 US election provided a mixed case. Polymarket correctly identified Trump as the likely winner when polls showed a toss-up, but the 57% probability still meant there was a 43% chance of being wrong. The market was right in this instance, but a 57% prediction that turns out correct does not mean the prediction was perfect. It means the market assessed the uncertainty correctly, which is a different and more subtle claim.
Calibration Is the Right Metric
The correct way to evaluate prediction market accuracy over time is calibration: do events that the market prices at 30% actually happen 30% of the time? Do events priced at 70% happen 70% of the time? Perfect calibration means the market is not systematically overconfident or underconfident at any probability level.
Studies of prediction market calibration generally show good performance in the 20-80% range but some overconfidence at the extremes (events priced at 95% happen slightly less than 95% of the time). This is consistent with the theoretical expectation: extreme probabilities are harder to price accurately because small absolute changes in probability require large percentage changes in position to correct.
Explore these tools on Blockcircle: Prediction Markets Mispricing Engine