How Accurate Are Prediction Markets? 2026 Evidence

Key takeaways

1. There is no single accuracy number, because there are two measurements. Directional accuracy asks whether the favourite won. Calibration asks whether things priced at 70 cents happen 70% of the time. A venue can look excellent on one and poor on the other.
2. On directional accuracy over the final five weeks of the 2024 presidential race, Clinton and Huang found PredictIt called 93% of outcomes better than chance, Kalshi 78% and Polymarket 67%, across 2,500+ markets. Volume did not help: bigger markets were not more accurate.
3. On calibration, Kalshi's own August 2026 research reports reliability curves close to the perfect line, with Brier scores falling from roughly 0.08 to 0.09 three months out to about 0.02 at close. It is the exchange analysing its own prices.
4. Independent work on 353 million trades finds politics the worst calibrated of six domains on Kalshi: 0.117 expected calibration error against 0.008 for sports and 0.007 for crypto, five to fifteen times any other category.
5. The political error runs toward the middle, not the edges. Prices compress toward 50 cents, so a 70 cent political contract a week from resolution has historically resolved Yes nearer 83% of the time. That is the opposite of the longshot bias every other explainer on this query describes.

On this page

1. Accurate at what? The question that settles most of the argument
2. How accurate are prediction markets on directional calls?
3. Calibration: the measure the exchanges prefer
4. How accurate are prediction markets by category?
5. Why politics scores the best Brier and the worst calibration
6. Which way the error runs: political prices lean toward 50 cents
7. Markets against polls: what 2024 actually settled
8. Reading tonight's midterm prices with the error bars on
9. How to check a market's accuracy yourself
10. Frequently asked questions
11. Sources and when we last checked

This is a plain list rather than jump links, because guide pages on this site render section text as paragraphs. If you read two sections, read the category list and the one on which way the error runs: together they are the only part of this that changes what you would pay for a contract.

Accurate at what? The question that settles most of the argument

1. Directional accuracy: did the favourite win? Pass or fail, discards the probability
2. Calibration: do 70 cent contracts resolve Yes about 70% of the time? Keeps the probability, ignores individual calls
3. Brier score: mean squared error of the probability. 0 perfect, 0.25 a coin flip on everything, superforecasters near 0.1 to 0.15

Almost every argument about prediction market accuracy is two people measuring different quantities and assuming the other is lying. Three measurements are in circulation and they answer three different questions.

Directional accuracy asks whether the thing priced above 50 cents happened. It is pass or fail per market and it throws away the number, so a market that said 51% and one that said 99% both score a point for the same call. Calibration asks whether the contracts that traded at 70 cents resolved Yes about 70% of the time. It keeps the number and ignores whether any single call was right. A Brier score combines both, where 0 is perfect, 0.25 is what you get by writing 50% on everything, and good human forecasters land near 0.1 to 0.15.

Only calibration tells you whether a price is a fair price. Directional accuracy cannot: a venue that priced every eventual winner at 51 cents would score 100% directionally and be useless to trade on. If you are deciding whether 68.5 cents is cheap, calibration is the only one of the three that bears on it.

How accurate are prediction markets on directional calls?

1. Study: Clinton and Huang, Vanderbilt University, circulated December 2025
2. Scope: 2,500+ political markets, final five weeks of the 2024 presidential campaign
3. Venues: Iowa Electronic Markets, PredictIt, Kalshi, Polymarket
4. Directional accuracy better than chance: PredictIt 93%, Kalshi 78%, Polymarket 67%
5. Cross-venue divergence on identical contracts, arbitrage peaking in the final fortnight
6. Volume did not improve accuracy once event type was controlled for

The most serious attempt at this is Joshua Clinton and TzuFeng Huang's Vanderbilt paper, covering more than 2,500 political markets over the final five weeks of the 2024 presidential campaign across four venues and more than two billion dollars of transactions. Their headline is a ranking, and not the one the industry would have written.

PredictIt called 93% of outcomes better than chance. Kalshi managed 78%. Polymarket, by far the largest by volume, came last at 67%. They also found prices for identical contracts diverging across exchanges, the resulting arbitrage peaking in the final fortnight, and 58% of Polymarket's national presidential markets showing negative serial correlation, meaning a price spike was typically reversed the next day. Their reading is that the platforms reward visibility and hype rather than news.

The finding nobody quotes is that volume did not predict accuracy. Controlling for event type, markets with more trading activity were not more accurate. That cuts against the usual defence of a prediction market price, which is that a lot of money is behind it.

Kalshi disputed the whole thing. Its media relations lead, Jack Such, said the methodology "completely misunderstands prediction markets, and shouldn't be taken seriously", and argued accuracy should be measured by calibration instead. That is a real methodological point rather than only a defence of his employer, and the next section takes it seriously.

Calibration: the measure the exchanges prefer

1. Source: Kalshi Research, "Calibration in Prediction Markets: Theory and Evidence", 20 August 2026
2. Scope: over 2.2 million data points, including previously unreleased internal data
3. Finding: reliability curves close to the perfect line, no significant distortion either way
4. Brier by horizon: about 0.08 to 0.09 at three months, about 0.02 at close
5. Caveat the paper states itself: the headline chart excludes sports, mentions and intraday markets

Kalshi Research published its own calibration study on 20 August 2026, built on over 2.2 million data points including internal data it had not released before. Its conclusion is that prices track observed frequencies closely, with what the write-up calls no evidence of significant distortion in any direction, and Brier scores improving as a market approaches resolution: roughly 0.08 to 0.09 at three months, falling to about 0.02 at close. It also reports that markets with as little as $10,000 of volume, or as few as 20 traders, still produce a useful forecast.

Two things are worth saying plainly. It is the exchange grading its own prices, which is not a reason to dismiss it but is a reason to want replication. And the write-up makes a presentational choice openly: it shows a second calibration chart excluding sports, mentions and intraday markets, describing the remainder as a higher quality signal. Excluding your weaker categories is a defensible analytical move, and it is why a reader needs the per-category breakdown below.

Kalshi's summary names elections, politics and finance as its strongest forecasts. Hold that claim in mind, because the largest independent study of the same prices concludes the opposite about politics.

How accurate are prediction markets by category?

1. Crypto: calibration error 0.007, Brier 0.174
2. Sports: calibration error 0.008, Brier 0.185
3. Finance: calibration error 0.016, Brier 0.156
4. Weather: calibration error 0.016, Brier 0.172
5. Entertainment: calibration error 0.022, Brier 0.160
6. Politics: calibration error 0.117, Brier 0.119 (worst calibrated, best Brier)

Nam Anh Le's August 2026 paper is the largest independent calibration study of these two venues, and nothing else ranking on this query cites it. It covers roughly 353 million trades across about 429,000 binary contracts: 64.7 million Kalshi trades over 210,608 contracts from July 2021, and 288.7 million Polymarket trades over about 218,000 resolved contracts from March 2023, cut off on 31 December 2025.

Its headline is that there is no such thing as the accuracy of a prediction market. Accuracy varies by domain, by distance from resolution and by trade size, and the spread is not small. Expected calibration error is the average gap between a price and the frequency it resolved at, so 0.117 means a typical political price sat about 12 points from its realised frequency.

Read that list twice, because it inverts the usual intuition. The markets people trust least, sports and crypto, are the best calibrated on the board. The markets written about as proof that prediction markets work carry five to fifteen times the calibration error of anything else. And politics simultaneously posts the lowest Brier score in the table, which is the contradiction the next section resolves.

Why politics scores the best Brier and the worst calibration

1. Brier = reliability (calibration error, lower better) minus resolution (decisiveness, higher better) plus uncertainty
2. Politics on Kalshi: reliability 0.024 (worst of six), resolution 0.147 (best of six)
3. Sports on Kalshi: reliability 0.000, resolution 0.063
4. A good Brier can mean confident and roughly right, not fairly priced

This is what lets Kalshi and its academic critics both be telling the truth.

A Brier score decomposes into three pieces: reliability, which is pure calibration error and where lower is better, minus resolution, which measures how far a forecaster will move away from the base rate and where higher is better, plus an uncertainty term fixed by the events themselves. In Le's Kalshi table, politics has the worst reliability of the six domains at 0.024 and the best resolution by a wide margin at 0.147. Sports has perfect reliability to three decimals and resolution of only 0.063.

Political markets therefore earn their good Brier by being decisive, not by being well calibrated. They commit to 85 cents and 12 cents instead of hovering near 50, those commitments mostly land on the right side, and the resolution term swamps the reliability penalty. Sports markets are beautifully calibrated and tell you comparatively little, because most of their prices sit near the middle.

So if someone quotes a Brier score as proof a venue's prices are fair, they have quoted the wrong statistic. Brier rewards confidence that pays off. Fairness of a price is the reliability term alone, and on politics that is the weakest number in the table.

Which way the error runs: political prices lean toward 50 cents

1. Calibration slope: 1.0 is perfect, above 1.0 means the price sat too close to 50 cents
2. Kalshi political slopes: 0.93 to 1.83 depending on horizon
3. Polymarket mean political slope: 1.45, so the effect replicates across venues
4. Illustration from the paper: 70 cents, one week out, historically resolved nearer 83%
5. Kalshi trades over 100 contracts: slope 1.74, against 1.19 for single-contract trades

Knowing political prices are off by about 12 points on average is half useful. The question a trader has is which direction, and Le's most robust finding answers it in reverse of what every other page on this query says.

Political markets are persistently underconfident: prices compress toward 50 cents, so favourites trade cheaper than they should and longshots dearer. The measure is a calibration slope, where 1.0 is perfect and above 1.0 means the price understated how extreme the truth was. Political slopes run from 0.93 to 1.83 across horizons on Kalshi, and the pattern replicates on Polymarket with a mean political slope of 1.45. As a slope-only illustration, the paper maps a 70 cent political contract one week from resolution to roughly 83%.

A second twist cuts against the crowd-wisdom story harder than anything in the Vanderbilt paper. On Kalshi, large political trades compress more, not less: trades above 100 contracts show a slope of 1.74 against 1.19 for single-contract trades. The size money leans hardest toward the middle. That does not hold up as robustly on Polymarket, so treat it as a Kalshi finding rather than a law.

Four caveats, because this is the finding most likely to be misread as a trading tip. The paper calls itself descriptive, not causal. Its decomposition explains 87.3% of in-sample slope variance on Kalshi but 71.5% out of sample, so the pattern is strong rather than mechanical. The data stop on 31 December 2025, putting every 2026 market outside the sample. And a tendency measured across hundreds of thousands of contracts tells you nothing reliable about any one race.

Markets against polls: what 2024 actually settled

1. Polling averages near 50-50 in late October 2024; Polymarket priced Trump a clear favourite
2. Theo commissioned YouGov neighbour polls in Pennsylvania, Michigan and Wisconsin
3. Position raised from about $30 million to roughly $80 million on that private data
4. Profit estimates: $48 million (Bloomberg, initial), $78.7 million net and above $85 million (Chainalysis)
5. Polymarket's own investigation found no attempt at manipulation

The case for prediction markets rests heavily on one night. Before the 2024 US presidential election, polling averages sat near a coin flip while Polymarket priced Donald Trump a clear favourite, and the market was right. It is a good result and it deserves its reputation. It is also the most over-read data point in the field, for a reason usually told as a triumph rather than a problem.

A large part of that price was one trader. A Frenchman known as Theo commissioned his own YouGov polling in Pennsylvania, Michigan and Wisconsin, including neighbour polls that ask respondents who the people in their social circle will vote for rather than who they will. Those showed materially more Trump support than conventional polls, and he raised his position from about $30 million to roughly $80 million. Profit estimates differ: Bloomberg initially reported $48 million across four accounts, Chainalysis later put it above $85 million, and a separate Chainalysis-derived figure was $78.7 million net. Polymarket investigated and concluded he had not attempted to manipulate the market.

Take that seriously and the story changes shape. The market beat the polls because one well-capitalised participant bought better private polling and pushed the price. That is a real and valuable function of a market, and it is not the wisdom of crowds. It also does not generalise: it needed one person with a correct contrarian thesis and eight figures to express it. One election is one observation, and anyone quoting a clean percentage for how often markets beat polls chose the denominator.

Reading tonight's midterm prices with the error bars on

1. Polymarket, Democrats control the House: 92.5 cents (event volume about $20.3 million)
2. Polymarket, Democrats control the Senate: 68.5 cents (event volume about $7.9 million)
3. Kalshi combination book: Democrats sweep 70.0, D-House with R-Senate 23.0, Republicans sweep 8.5, R-House with D-Senate 0.6
4. Kalshi implied Democrats House: 70.0 + 23.0 = 93.0 cents, against Polymarket's 92.5
5. Kalshi implied Democrats Senate: 70.0 + 0.6 = 70.6 cents, against Polymarket's 68.5, a 2.1 cent gap
6. Kalshi's four combination contracts sum to 102.1 cents, a 2.1 cent overround on the set

The 2026 midterms are on 3 November, twelve days from this writing, and they are the most heavily traded political markets on either venue. Here is what both order books said at 21:05 UTC on 22 October 2026, read from Polymarket's Gamma API and Kalshi's public markets API rather than from a screenshot.

Two things to notice. The venues disagree by 2.1 cents on Senate control, which is the cross-venue divergence the Vanderbilt paper measured in 2024 reappearing in 2026 on the most liquid political question on the board. And Kalshi's four combination contracts sum to 102.1 cents rather than 100, so buying the set costs $1.02 for a guaranteed dollar. That overround is the normal cost of a four-way book, and it is why the cross-venue gap is not free money once two taker fees are added. Our midterms guide works that arithmetic through in full.

Applying the calibration findings honestly means resisting the obvious move. The historical pattern says political prices in this range understated the eventual frequency, which would make the Senate at 68.5 cents look cheap. But that sample ends in December 2025, it is descriptive rather than causal, and a tendency across 218,000 contracts says nothing dependable about one Senate race twelve days out. The durable conclusion is narrower: treat a political price as a figure with roughly ten points of calibration error around it rather than as a forecast, and treat a 2 cent gap between two venues as evidence that neither knows the answer to the precision it is quoting.

How to check a market's accuracy yourself

1. Pick one category and one venue and keep them fixed. Mixing sports with politics hides exactly the difference this page is about
2. Log the price at a consistent horizon, for example one week before close, with the date and the market
3. Wait for resolution and log the binary outcome, with no exceptions for markets you decided were silly
4. Bucket the prices into tens: 0 to 9 cents, 10 to 19, and so on up to 90 to 99
5. For each bucket compare the average price against the share that resolved Yes. A 70 to 79 bucket that resolved 83% Yes is the underconfidence above
6. Watch the direction of the gaps, not their size. Gaps that all point one way are bias you can price; gaps that alternate are sample noise

None of this requires taking anyone's word for it, mine included. Calibration is the one forecasting statistic an ordinary person can compute, because it needs no model: only prices you recorded before the event and outcomes you can look up after. The discipline that makes it work is logging the price before you know the answer, because a record assembled from memory remembers the confident calls that came in. Twenty entries shows whether your own reads are biased; fifty before a venue's bucket frequencies mean much.

Frequently asked questions

Sources and when we last checked

1. Joshua Clinton and TzuFeng Huang, "Prediction Markets? The Accuracy and Efficiency of $2.4 Billion in the 2024 Presidential Election", SocArXiv preprint d5yx2 (circulated December 2025): the 93%, 78% and 67% directional figures, 2,500+ markets, cross-venue divergence and the volume result
2. DL News, 5 December 2025: Kalshi's response, the 58% negative serial correlation figure and Jack Such's quoted objection
3. Kalshi Research, "Calibration in Prediction Markets: Theory and Evidence", 20 August 2026, and Kalshi's own summary of it: over 2.2 million data points, Brier by horizon, the reliability charts and the excluded categories
4. Nam Anh Le, "Decomposing Crowd Wisdom: Domain-Specific Calibration Dynamics in Prediction Markets", arXiv 2602.19520v2, 4 August 2026: the 353 million trades, the per-domain calibration and Brier table, the political slopes and the trade-size result
5. Bloomberg (October 2024), NBC News and Chainalysis via CBS News: the 2024 neighbour-poll trade, the position sizes and the conflicting profit estimates
6. Polymarket Gamma API and Kalshi public markets API, both read at 21:05 UTC on 22 October 2026: every live price in the midterms section

One transparency note. Kalshi's calibration PDF returned HTTP 429 to an automated fetch while this was written, so its figures come from Kalshi's own published summary of the study rather than the paper's tables, and should be read as the exchange's reported numbers. Everything attributed to Clinton and Huang and to Le came from the papers' own text. Where the two disagree about politics, and they do directly, this guide reports both rather than reconciling them.

Both 2026 studies stop collecting before 2026 began, so neither has measured a market in the current cycle. If you are about to act on a political price rather than read about one, carry away this rather than any number above: a political market price is a figure with real error bars, the error has historically pointed toward the middle, and nothing here is financial advice.

Related guides