How Accurate Are Prediction Markets? What the Resolved-Market Data Shows

Prediction markets get calibration right more often than they get individual calls right, and that distinction is important to us at PredictionHero. Across large samples of resolved contracts, calibration research consistently finds that markets priced around 70% resolve "yes" close to 70% of the time. That is calibration, and it is the metric that actually matters, not whether any single favorite won.

That headline finding gets lost in most coverage of prediction market accuracy, because most coverage focuses on famous hits and misses instead of the distribution underneath them. Here is what the resolved-market record actually shows, and where it breaks down.

Methodology: What We Mean by "Accurate"

Before the numbers, the terms.

"Accuracy" for a prediction market does not mean "picked the right winner." A market that prices a team at 70% and loses did not fail. It said there was a 30% chance of exactly that outcome.

The right test is calibration: pull every resolved contract that traded near a given price band, and check how often "yes" actually happened across that band.

PredictionHero aggregates the same real-world question across various leading platforms, such as:

Polymarket, a crypto-native venue known for high-volume political, geopolitical, sports, and cryptocurrency markets;

Kalshi, the CFTC-regulated US event exchange focused on economics, politics, weather, and current-event forecasting;

Opinion, focused on event-driven forecasting and real-time public-sentiment markets; Limitless, a blockchain-based venue spanning a broad range of speculative and event-driven markets;

Predict.Fun, focused on crowd-sourced forecasting and speculative event markets. What follows draws on the public record of prediction-market calibration research rather than on any single venue.

Three limits matter for reading anything below. First, resolution rules differ by platform and by market, so "yes" on one venue is not always the identical proposition as "yes" on another for the same event.

Second, calibration studies typically bucket contracts by price band (60-70%, 70-80%, and so on) across hundreds or thousands of resolved markets, not single events, because any one market can miss without the underlying process being broken.

Third, coverage varies by category. Election and macro markets on these platforms tend to be deep and well-studied. Long-tail entertainment or niche political markets often are not, and their calibration is far less established.

What Calibration Actually Means

Say a contract is trading at $0.70. That is a 70% implied probability. If markets are well calibrated, and you gather every contract that traded at $0.70 at resolution across a large enough sample, roughly 70% of them should have resolved "yes" and roughly 30% should have resolved "no."

That is different from accuracy in the everyday sense. A well-calibrated market that says 70% and the event does not happen has not been wrong. It assigned a 30% chance to precisely that outcome, and 30% chances happen three times out of ten.

The failure mode to watch for is systematic miscalibration: contracts priced at 70% that resolve "yes" only 40% of the time, or only 90% of the time. Either direction means the market is mispricing the category, not just getting unlucky on one event.

This is also why a single wrong call, even a famous one, tells you almost nothing about whether a market works. The question is never "did the favorite win." It is "across every contract priced like this one, does the frequency match the price."

Where Markets Beat Polls

The clearest historical case for prediction markets over polling comes from the Iowa Electronic Markets, which has operated since 1988 and remains one of the longest-running natural experiments in market-based forecasting.

Research by Berg, Nelson, and Rietz comparing IEM prices against nearly a thousand polls across five presidential elections (1988–2004) found the market closer to the final result about 74% of the time, and its advantage over polls was largest on forecasts made more than 100 days before the vote.

The mechanism is straightforward: a poll captures a snapshot of stated opinion at one moment. A market aggregates information from anyone willing to back a view with money, and it updates continuously as new information arrives, rather than waiting for the next survey wave.

The structural advantage compounds close to resolution. Polls are expensive to run frequently and slow to reflect late-breaking developments. A liquid market repositions within minutes of news, because someone with better information has a direct financial incentive to trade on it before the crowd catches up.

Where Markets Miss

The record is not uniformly favorable. Prediction markets have underperformed statistical polling aggregators on races that are numerically close, where the "wisdom of crowds" effect that markets rely on is weaker because informed traders disagree sharply with each other rather than converging.

Thin liquidity is the recurring culprit: a market with few active traders and small position sizes can be moved by a handful of large bets, which distorts the price away from genuine consensus. Deep, heavily traded markets on flagship platforms like Polymarket and Kalshi are far more resistant to this than a lightly traded contract on any platform.

Markets also inherit bias from who shows up to trade them. A venue whose user base skews toward one region, one political orientation, or one asset class will price accordingly. That is not a flaw unique to any single platform, it is a structural fact about aggregating opinion from a specific population rather than a random sample, and it is one reason checking a question across all five platforms produces a more reliable read than trusting any one venue in isolation.

Resolution ambiguity is a separate failure mode entirely, and it has nothing to do with forecasting skill. When a contract's settlement language is vague, disputes over how it should resolve can undermine confidence in the result even when the underlying price was well calibrated the whole way through.

Citable Summary Stats

• Iowa Electronic Markets: operating continuously since 1988, one of the longest-running live tests of market-based election forecasting against poll averages (Berg, Nelson & Rietz, 2008).
• Calibration research across large samples of resolved contracts consistently finds prices near a given band resolve at frequencies close to that band, the standard test for whether a market is "accurate" in the statistical sense (Wolfers & Zitzewitz, 2004).
• The market edge over polls is clearest at longer forecasting horizons and in races with a decisive frontrunner; in very close contests it narrows against strong statistical aggregators, and any single election is a thin basis for claiming market superiority (Berg, Nelson & Rietz, 2008; Undark, 2026).
• Thin liquidity, not platform choice, is the single biggest predictor of a mispriced individual contract.

Caveats Worth Repeating

A market being well calibrated in aggregate says nothing about whether any specific contract you are looking at right now is priced correctly. Calibration is a population statistic. Individual markets, especially in illiquid or novel categories, can and do drift from the true probability for stretches of time before new information corrects them.

Cross-platform comparison helps here. Checking a contract on Polymarket against the same or similar question on Kalshi, Limitless, Predict.Fun, and Opinion surfaces disagreement, and disagreement is itself information: it tells you the market has not converged, which is a signal to treat the current price as provisional rather than settled.

Sources

• Berg, J. E., Nelson, F. D., & Rietz, T. A. (2008). "Prediction Market Accuracy in the Long Run." International Journal of Forecasting, 24(2), 285–300. PDF
• Iowa Electronic Markets — University of Iowa, Tippie College of Business (real-money election markets operating since 1988). iemweb.biz.uiowa.edu
• Wolfers, J., & Zitzewitz, E. (2004). "Prediction Markets." Journal of Economic Perspectives, 18(2), 107–126. NBER
• "Prediction Markets Aren’t Likely to Replace Polling." Undark, May 5, 2026. undark.org

PredictionHero aggregates publicly available prediction market data for informational purposes only. This is not financial advice. Prediction markets may not be available in all jurisdictions.