Are Prediction Markets Good Forecasters?

Prediction markets turn trader beliefs into prices that can be read as probabilities. The sources describe strong evidence that they often beat polls: the Iowa Electronic Markets, operated by the University of Iowa, were closer to actual outcomes than 74% of polls despite a $500 maximum investment. The statistical question is whether prices behave as efficient forecasts, meaning independent errors and random-walk price changes, or whether they overreact or underreact to news. Failure conditions also matter: thin liquidity, information vacuums, extreme events and systematic biases. In 2026 the largest venues, Polymarket and Kalshi, transact billions of dollars each month, yet much of that volume is sports, crypto prices or election betting, which limits how much of it informs decisions. This brief covers calibration, efficiency tests and known limits.

Key concepts

Calibration

The key accuracy metric: in a well-functioning market, contracts priced at 70% should resolve Yes approximately 70% of the time. Plotted as predicted probability against actual frequency, ideal performance lies on the diagonal. S2 · tradealgo.com

Efficient market random walk

If prices equal expected final vote share given current information, the law of iterated expectations holds and forecast errors are independent. Theory therefore predicts a unit root in the evolution of forecasts. S1 · biz.uiowa.edu

Overreaction versus underreaction

Overreaction implies negative correlation of price changes and mean reversion, with a lagged coefficient below one. Underreaction, where markets adapt to news slowly, implies a coefficient above one. S1 · biz.uiowa.edu

Bid/ask bounce

Price movement between bid and ask quotes that causes some negative autocorrelation, creating the appearance of mean reversion in prediction market prices. S1 · biz.uiowa.edu

Information markets

Hal Varian's term for prediction markets, in which bettors are the suppliers of information. Demand from individuals, firms, media and policymakers can serve as a proxy for usefulness. S3 · asteriskmag.com

Testing forecast accuracy and standard error

1. Restrict to the last one hundred daysMarkets differ in duration and are thin and volatile early on, so only the final one hundred days before each election are analyzed, when traders and bid/ask queues are reasonably thick.
2. Model the price processEstimate whether the coefficient on lagged spread forecasts differs from one, and whether correlated errors add moving average components to the evolution of forecasts.
3. Bootstrap the confidence intervalsSmall-sample confidence intervals for AR(1) processes are hard to assess accurately, so bootstrapping is used, following Savin and Nankervis (1996).
4. Address bid/ask bounceBecause bid/ask bounce produces spurious negative autocorrelation and apparent mean reversion, the analysis applies a correction before judging overreaction.
5. Compute the forecast standard errorUsing the standard deviation of the one step ahead forecast error and the forecast horizon, build intervals of +/-1.96 standard errors around the forecast at the election date.

Evidence from the Iowa Electronic Markets

The Iowa Electronic Markets (IEM), operated by the University of Iowa's Tippie College of Business, is described as the longest-running and most-studied prediction market. Berg, Nelson and Rietz (2008) are cited for the finding that IEM beat polls 74% of the time over 16 years of data; the same source states that IEM prices were closer to actual outcomes than 74% of polls released during the same period. The advantage was most pronounced for election-eve predictions, comparing final market prices with final polls. The IEM operated with a $500 maximum investment and minimal liquidity, which the source takes as a sign that the informational advantage is structural rather than a function of scale. The same source adds that Polymarket priced Trump's victory in 2024 weeks before polling averages converged, and that a PNAS study used prediction markets to forecast whether psychology studies would replicate. Calibration remains the key test of whether prices can be read as probabilities.

Are prices efficient forecasts?

A statistical treatment of prediction market accuracy starts from the efficiency requirement that errors be independent. If errors are independently and identically distributed, the mean and standard deviation can be estimated from sample data, and the forecast and its error at the election date can be computed. The standard error of the forecast depends on the standard deviation of the one step ahead forecast error and the horizon, which allows +/-1.96 standard error confidence intervals. The source stresses there is no guarantee that markets are efficient random walks. De Bondt and Thaler (1985, 1987 and 1990) argue that ordinary and expert traders overreact, implying mean reversion. Abarbanell and Bernard (1992) and Chan, Jegadeesh and Lakonishok (1996) argue that markets adapt slowly, implying underreaction. The test is whether coefficients differing from one can be distinguished and, if so, whether forecasts and standard errors should be adjusted accordingly.

Volume, demand and usefulness in 2026

Institutions including Google, Microsoft, the CIA, the wider U.S. intelligence community and British government intelligence analysts have experimented with internal prediction markets, but all trials were small, and theory and practice both show that more bettors make markets more accurate. In 2026, Polymarket and Kalshi transact billions of dollars in volume each month, but the author notes that roughly 90% of Kalshi's trading volume is sports betting and that over 80% of Polymarket's volume is concentrated on sports, cryptocurrency prices or election betting. Because supply of forecasts does not guarantee value, the author looks at demand from individuals, firms, media and policymakers as a proxy for usefulness. Vitalik Buterin's framing captures the dual nature: for a bettor the platform is a betting site, and for a reader of the charts it is a news site. The theoretical roots run from Hayek in 1945 to Hanson's futarchy proposal in 2000.

Price dynamics under three hypotheses

Timeline

1. 1945
   Hayek argues markets aggregate dispersed, local and tacit knowledge through prices.asteriskmag.com
2. 2000
   Robin Hanson proposes futarchy, using markets to evaluate whether policies deliver.asteriskmag.com
3. 2007
   Arrow, Kahneman and other scholars publish a statement backing prediction markets for decision-making.asteriskmag.com
4. 2008
   Berg, Nelson and Rietz study finds IEM beat polls 74% of the time.tradealgo.com
5. 2024
   Polymarket prices Trump's victory weeks before polling averages converge.tradealgo.com
6. 2026
   Polymarket and Kalshi transact billions of dollars in volume each month.asteriskmag.com

Common pitfalls

• Early in a market's life, volume is low and volatility is excessive because the market is thin, so early prices are not comparable with late ones.
• Bid/ask bounce creates negative autocorrelation, so apparent mean reversion may be a quoting artifact rather than trader overreaction.
• Markets fail in identifiable conditions: thin liquidity, information vacuums, extreme events and systematic biases. Outperforming polls does not mean being perfect.
• Most volume sits in sports, crypto prices and elections, so large markets do not imply that most traded questions yield decision-useful information.

Key takeaways

• Prediction markets have outperformed polls in studied elections; IEM prices were closer to outcomes than 74% of polls, even with a $500 cap.
• Check calibration and price dynamics, not just headline accuracy: efficient prices follow a random walk, while overreaction or underreaction changes forecast error estimates.
• Liquidity and question type matter: thin markets and sports-dominated volume weaken the case that prices always carry useful forecasting information.

FAQ

How this page is made. Built from the full text of 3 sources found with the AISA Tavily API (advanced search), refreshed weekly. A language model organizes what those sources state into the blocks above; each block links the sources it draws on, and figures that do not appear in the sources are removed.