Prediction markets are accurate in the way a good weather forecast is accurate: not always right, but right about as often as they say they will be. Events priced at 70% happen roughly 70% of the time, and the markets beat opinion polls more often than not. Polymarket's markets have posted a Brier score around 0.09, a measure of forecasting error on which zero is perfect and 0.25 is what you would get by calling every question a coin flip.
So are prediction markets accurate? The honest answer has two halves. On average and over many questions, the evidence says yes, and it has said so for decades. On individual questions, markets make well-known mistakes, some of them systematic: they overprice long shots, they can be moved by thin books and partisan money, and they are less reliable far from resolution than close to it.
This guide covers what accuracy means for a probability, how it is measured, the evidence from elections and beyond, why markets tend to work, where they fail, and how to use their prices sensibly.
| Evidence | What it shows |
|---|---|
| Iowa Electronic Markets, 1988–2004 | Market prices were closer to the final result than 74% of 964 polls compared |
| Polymarket's Brier score | About 0.09, well below the 0.25 of a coin-flip forecaster |
| Favourite–longshot bias | Long shots are overpriced across betting and prediction markets |
| 2016 US election and Brexit | Markets favoured the side that lost |
| 2022 US midterms | Markets overestimated Republican gains |
| 2024 US election | Markets favoured the winner while polls showed a toss-up |
What "Accurate" Means for a Probability
A single probability cannot be right or wrong on its own. If a market prices a candidate at 70% and the candidate loses, that is exactly what should happen three times in ten. The test of a forecaster is not whether its favourites always win, but whether its probabilities match reality across many questions.
That property is called calibration. A calibrated forecaster's 70% calls come true about 70% of the time, its 20% calls about 20% of the time, and so on. The second property that matters is resolution, sometimes called sharpness: a forecaster that calls everything 50% is perfectly calibrated and useless. Good forecasts are both calibrated and willing to commit to confident probabilities when the evidence supports them.
This framing matters for how you read any single price. When the market's favourite loses, the market was not necessarily wrong. When it is wrong in the same direction again and again, it is. Our explainer on prediction market odds covers how prices translate into probabilities in the first place.
How Accuracy Is Measured: Brier Scores and Calibration
The standard measure is the Brier score: the average of the squared difference between the forecast probability and the outcome, where the outcome is 1 if the event happened and 0 if it did not. A forecast of 0.9 on something that happens scores (1 − 0.9)², or 0.01; a forecast of 0.9 on something that does not scores 0.81. Averaged across many questions, lower is better.
A quick comparison shows why the score rewards both calibration and confidence. Forecaster A says 80% on ten events, and eight happen; A's average Brier score is 0.16. Forecaster B says 50% on the same ten events; B scores 0.25, despite never being "wrong". A scores better because A committed to probabilities that the outcomes justified. Our prediction market glossary defines Brier scores, calibration and the other terms used here.
For reference, a forecaster who says 50% on every binary question scores 0.25. Polymarket's markets have posted scores around 0.09, which is strong — the kind of number usually quoted as "about 94% accurate", though that shorthand hides a lot. A score depends heavily on the mix of questions: markets full of near-certain outcomes score well easily, while markets on genuinely uncertain events cannot.
Calibration charts are the more intuitive check. Group every forecast by its probability, then see how often events in each group happened. For a well-calibrated market, the points line up along the diagonal. Most published analyses of large prediction markets show good calibration in the middle of the range and a consistent distortion at the extremes, which is the favourite–longshot bias covered below.
The Evidence From Election Markets
The longest record comes from the Iowa Electronic Markets, run by the University of Iowa since 1988 with small real-money stakes. A study of the markets' first dozen years of elections found that their prices were closer to the eventual result than 74% of the 964 polls it compared them with. That finding launched much of the academic interest in prediction markets, and our history of how prediction markets evolved traces what followed.
The modern record is more mixed. In 2016, prediction markets made Hillary Clinton a clear favourite, and in the Brexit referendum the same year, betting markets strongly favoured Remain. Both were wrong. In the 2022 midterms, markets priced a Republican wave that arrived as a ripple, with Democrats holding the Senate. In 2024, markets gave Donald Trump a better chance than most polling averages, which showed a near toss-up, and he won.
Those are four high-profile cases, which is too few to judge calibration. What they show is that markets are not a crystal ball: they aggregate the same information polls and models do, and sometimes the information is wrong. Today's 2026 midterm election odds should be read with that history in mind, and our 2026 midterms trading guide covers how those markets have behaved.
Why Markets Tend to Be Accurate
Three mechanisms do the work. The first is incentives. A trader who thinks a price is wrong can profit by correcting it, so people with good information have a reason to reveal it through their trades, and people with bad information lose money and trade less. The second is aggregation. A market price combines the views of everyone trading, weighted by how much they are willing to risk, which tends to cancel out individual errors. The economist Friedrich Hayek argued in 1945 that prices are a way of pooling knowledge no single person holds, and prediction markets are that idea applied to questions rather than goods.
The third is arbitrage. When two markets price the same event differently, traders profit by buying the cheap one and selling the dear one until the gap closes. That pulls prices toward a single consensus. PredictReport's tracker shows it clearly: on September 21, 2026, Polymarket and Kalshi priced Democratic control of the House at 91.5¢ and 91.3–91.5¢ — two separate books, with different users and different currencies, within half a cent. Our arbitrage scanner shows where that process is still working on the gaps that remain.
The Victorian scientist Francis Galton gave an early demonstration of the same idea at a county fair, where the middle estimate of a crowd guessing an ox's weight came within about 1% of the true figure. Markets refine that by letting people back their estimates with money.
Where Markets Go Wrong
Long shots are overpriced. Across horse racing, sports betting and prediction markets, contracts on unlikely outcomes trade above their true probability. A contract priced at 3¢ tends to win less than 3% of the time. Researchers have attributed the pattern to people overweighting small probabilities and enjoying the lottery-like payoff. Our expected value guide explains why this makes cheap contracts a poor default.
Thin books produce false prices. A market with little trading can be moved by a single order. PredictReport's tracker recorded a Polymarket trade putting Dak Prescott's NFL MVP chances at 13.1¢ while Kalshi's midpoint was 6.5¢, and a print of 31.4¢ for Save the Children in the Nobel Peace Prize market against 12¢ on Kalshi. Both reversed. Our NFL MVP odds and Nobel Peace Prize odds describe those episodes, and our guide to prediction market liquidity explains how to spot a price nobody is standing behind.
Crowds have biases. Election markets draw politically engaged traders, and sports markets draw fans. Enthusiasm for a candidate or a team can hold a price above fair value, especially on smaller markets. That is one reason prices that match across two venues with different user bases are more trustworthy than prices on one.
Long horizons add noise. A market two years from resolution has little information to aggregate, and its price is closer to informed speculation than to a forecast. Our 2028 Democratic nominee odds show a leader at 18%, a market that knows it does not know.
People try to move prices. In 2012, a trader on the now-defunct Intrade was reported to have spent heavily pushing up Mitt Romney's price, and markets have been targeted by people with inside information — the 2025 Nobel Peace Prize market moved from under 4% to over 70% on the eventual winner shortly before the announcement, in what the Nobel Institute later attributed to a likely cyberattack. Our analysis of prediction market insider trading regulation covers that problem.
The question can be ambiguous. A market can be "wrong" because its rules resolved differently from what traders expected. Our explainer on how prediction markets resolve covers the details that decide contracts.
Accuracy Improves as Resolution Approaches
A market's accuracy depends heavily on how far away the answer is. Close to resolution, prices reflect almost everything knowable, and calibration is typically at its best. Far from resolution, there is simply less information to aggregate, and prices lean on base rates, narratives and the preferences of whoever is trading.
That is why the same market can be excellent at one point in time and weak at another. The 2028 Republican nomination market prices J.D. Vance at about 51%, and history supports a sitting vice president's chances — our 2028 Republican nominee odds cover that base rate. But two years is long enough for events nobody is pricing, and a probability that far out should be read as today's best estimate rather than a forecast of 2028.
Markets vs Polls vs Models
Polls measure what a sample of people say they will do. Statistical models combine polls with other data, such as economic conditions and past results. Markets combine all of that with whatever else traders know or believe, and put money behind it.
In practice, the three are not rivals so much as layers. Markets watch polls and models closely, and a large poll or model update usually moves prices within minutes. Where markets add value is in speed, in incorporating information that is not in any poll, and in expressing uncertainty as a single tradeable number. Where they lose value is when the traders' shared assumptions are wrong, as in 2016 and 2022. A sensible reader treats all three as inputs and pays attention when they disagree.
Sports offer a useful comparison. A sportsbook's closing line — its final price before a game starts — is widely regarded as one of the most efficient forecasts in any market, because sharp bettors spend the whole week correcting it. Prediction markets on the same games tend to converge to similar probabilities, and without the bookmaker's margin. Our comparison of prediction markets and sports betting covers the structural differences, including why exchanges do not limit winning customers.
How to Use Market Probabilities Well
Treat a liquid market's price as a strong baseline, not a verdict. If you disagree with it, ask what you know that the traders do not; most of the time, the honest answer is nothing. That discipline is the single best protection against losing money on prediction markets, and our guide to making money on prediction markets starts from it.
Check depth and cross-venue agreement before trusting any price. A number that holds on both Polymarket and Kalshi, on a market with real volume, deserves much more weight than a single print on a quiet book. Polymarket vs Kalshi explains why the two venues drift apart when they do.
Discount long shots and very long-dated prices.
A contract at 3¢ on an event two years away is the least reliable price a prediction market produces: it combines the favourite–longshot bias with the long-horizon noise described above. Treat it as a rough indication that the outcome is possible, not as a measured probability.
If you trade, track your own calibration: record the probability you believed when you traded, then compare with outcomes after thirty or more trades. That is the only reliable way to learn whether you are better than the market. A spreadsheet is all it takes, and our roundup of free prediction market tools describes a simple journal format that makes the calibration check straightforward. The exercise is humbling for most people, which is exactly why it is worth doing before scaling up. Most traders discover that their most confident calls are overconfident, and knowing that is worth more than any single winning trade. For the mechanics underneath all of this, see how prediction markets work, and for choosing a venue, our ranking of the best prediction market apps.
Frequently Asked Questions
How accurate are prediction markets?
Accurate on average. Polymarket's markets have posted a Brier score around 0.09, and the Iowa Electronic Markets beat 74% of the polls they were compared with over 1988–2004. Individual markets still miss, and long shots are systematically overpriced.
Are prediction markets more accurate than polls?
Often, but not always. Markets incorporate polls along with other information and usually react faster. They missed in 2016 and 2022 and did better than most polling averages in 2024.
What is a Brier score?
The average squared difference between forecast probabilities and outcomes, where zero is perfect and 0.25 matches a forecaster who says 50% on everything. Lower is better.
What is the favourite-longshot bias?
The tendency for unlikely outcomes to be overpriced across betting and prediction markets. A contract at 3¢ tends to win less than 3% of the time. See our expected value guide.
Were prediction markets right about the 2024 election?
They favoured Donald Trump while most polling averages showed a near toss-up, and he won. One election is not enough to judge a market's calibration.
Can prediction markets be manipulated?
Prices can be pushed by large orders, especially on thin markets, and markets can be targeted by people with inside information. Arbitrage and deep books limit how long a false price survives. See our analysis of insider trading in prediction markets.
Why do prediction markets sometimes get elections wrong?
Because they aggregate the same information as polls and models, and when that information is systematically wrong — as in 2016 — the market is too. A 70% favourite is also expected to lose three times in ten.
How should I use prediction market odds?
As a baseline probability from a liquid market, checked against a second venue. Discount long shots and long-dated prices, and track your own forecasts to see whether you can beat the market.
