Markets vs. Polls
Election forecasts turn poll leads into winning chances that can be compared with prediction market prices to judge accuracy.
Following an election involves judging both who is ahead and how likely that lead is to hold. In a hypothetical two-candidate race, a poll might put one candidate at 51% support while a prediction market implies a 70% chance of winning. A narrow lead is more likely to survive when polling error and changes in support are expected to be small.
Election forecast models estimate those uncertainties to turn polling evidence into winning probabilities. Their estimates can be compared with market prices for the same result, giving readers a way to assess whether traders and forecasters disagree. Markets have beaten raw polls in several election studies, while comparisons with forecasts built from polls have produced mixed results.
What is the difference between prediction markets and polls?
An election poll estimates how many people in a defined population support a candidate. A prediction market lets traders buy and sell contracts whose value depends on a specified result. For a contract that pays $1 if a candidate wins and zero otherwise, a price of 70¢ corresponds to an implied 70% chance of winning.
| Number | What it describes | How to read it |
|---|---|---|
| Voting-intention poll | Support among the people the poll aims to represent | “51% of likely voters support A.” |
| Polling average | Support estimated by combining several polls | “Across the polls, A's support is around 51%.” |
| Election forecast | A projected result, including uncertainty | “A has a 70% chance of winning.” |
| Winner-contract price | The trading price of a claim on the winner | “A's Yes contract costs 70¢ for a possible $1 settlement.” |
A candidate could win narrowly in many plausible versions of an election while losing in the others. That would produce a high chance of winning without a large expected share of the vote. A poll percentage and a win probability measure different quantities even when both describe the same race.
A wins seven of ten hypothetical results for the same election, while A's vote share averages 51%. Counting the results above 50% gives a 70% chance of winning from a set of mostly narrow victories.
Market prices also depend on the contract. The Iowa Electronic Markets have offered vote-share contracts that settle in proportion to a candidate's share of the two-party popular vote. A vote-share price can be compared with a projected share of votes. A winner contract calls for a comparison with a forecast probability.
How do election polls work?
Election polls ask a sample of people about their voting intentions, then adjust the answers to estimate support in a larger population. The population, recruitment method, weighting, and assumptions about who will vote all affect the published result.
A poll of all adults answers a different question from a poll of registered voters. A poll of likely voters adds an estimate of who will actually cast a ballot. Changing the population can change the headline percentage even when nobody has changed their mind.
Probability samples start with a random selection from a list or other source intended to cover the population. Online opt-in surveys recruit volunteers, sometimes through advertisements or sites offering rewards. In Pew's survey comparisons, opt-in samples had about twice the average error of probability samples. Both methods can collect answers online, so the interview format alone tells you little about how people were selected.
Recruitment also determines how to interpret a poll's uncertainty. An opt-in survey cannot use the usual sampling margin of error. Its model-based interval depends on assumptions about how volunteers represent the wider population.
Why pollsters weight the answers
Weighting gives some respondents more influence than others to make the sample resemble the population on characteristics such as age or education. If younger adults are underrepresented, their answers can receive more weight. This changes the combined estimate while preserving the answers people gave.
The adjustment helps when the characteristic is related both to survey participation and to the answer being measured. Matching age proportions helps less if, within every age group, the people who answer differ politically from those who do not. A larger sample can estimate respondents' views more precisely while leaving that difference unresolved.
A survey invitation sits between a phone displaying a questionnaire and another placed face down. The poll's answers come from the people who choose to respond.
How a likely voter model changes the result
A likely voter model estimates turnout using information such as past voting, interest in the campaign, and stated intention to vote. Pollsters can classify respondents as likely voters or weight their answers by their estimated chance of voting.
Suppose younger respondents prefer one candidate and older respondents prefer the other. Two pollsters who measure those preferences similarly can still publish different leads if one expects more young people to vote.
The two hypothetical pollsters use the same survey answers but reach different leads because they expect different turnout. Younger voters favor A, so raising their expected share of the electorate from 20% to 35% changes the result from a 4-point lead for B to a 2-point lead for A.
Question wording and undecided voters
Approval of a governor's work, a favorable opinion of a candidate, and an intention to vote for that candidate measure different things. Someone can disapprove of a politician and still prefer them to the available opponent. The questionnaire establishes which question the headline number answers.
Excluding undecided respondents also changes the percentage. Suppose a poll gives A 48%, B 46%, and leaves 6% undecided. Among respondents choosing A or B, A has about 51.1% of support, calculated as 48 divided by 94. Using that share as a final vote forecast assumes the undecided voters eventually split in the same proportions or stay home.
In this hypothetical poll, A's 48% support becomes 51.1% when the 6% who are undecided are left out. The second bar divides the same answers by the 94% who chose A or B, so the larger share comes from changing who is counted.
What is the margin of error in polls?
A poll's margin of error measures uncertainty from surveying a sample of the population. The familiar “plus or minus 3 percentage points” usually applies to an individual candidate's support. The uncertainty in the lead between two candidates is larger.
In a close two-candidate race, a sampling error that overstates one candidate's share also understates the other's. A margin of about ±3 points for each share becomes about ±6 points for the difference. A reported five-point lead can therefore be consistent with anything from a one-point deficit to an eleven-point lead, considering sampling error alone.
A hypothetical two-candidate poll puts A ahead by 5 points, with 52.5% support against B's 47.5%. A 3-point error in A's share also moves B's share in the opposite direction, giving a range from a 1-point deficit to an 11-point lead for A.
Several sources of polling error sit outside that printed margin. People can be missed by the recruitment method, decline to participate, interpret a question differently, or vote at a rate the pollster did not expect. Collecting more responses does not automatically correct those problems. A lead can be real at the time of the survey and still leave the result on Election Day uncertain.
How accurate are election polls?
In its 2024 review, the American Association for Public Opinion Research reported an average absolute error of 3.3 percentage points on the two-party margin across 611 polls in the final two weeks. The polls covered presidential, Senate, House, and governor races. The comparable figures were 5.3 points in 2020 and 5.2 points in 2016. These measure how far the estimated gap between candidates was from the final gap, regardless of which candidate the poll favored.
Polls from the final two weeks of US election campaigns missed the two-party margin by an average of 5.2 percentage points in 2016, 5.3 in 2020, and 3.3 in 2024, according to the American Association for Public Opinion Research's review. These absolute errors count misses in either direction, with 2024's average about two points smaller than the earlier two.
A study of 4,221 state-level polls from 1998 to 2014 measured how far polls missed a candidate's final two-party vote share. Across presidential, Senate, and governor races in the final three weeks, the root mean square error was about 3.5 percentage points. That measure gives larger misses extra weight. The actual errors were roughly twice as large as expected from the polls' reported sampling uncertainty. This was an average across many elections, rather than a multiplier for an individual poll's margin. Each poll's additional uncertainty depends on how it recruits respondents and estimates turnout.
What is a polling average?
A polling average combines several polls to estimate where support stands. It reduces the influence of an unusually high or low result from any one sample, making it easier to follow a trend.
Some averages take a simple mean of recent polls. Others give more weight to newer polls or pollsters with stronger records. Some also adjust for a house effect, a pollster's tendency to produce results more favorable to one party than other polls do. An average remains an estimate of support during the period it describes. Turning it into a prediction requires assumptions about changes before Election Day and errors that could remain in the polls.
Averaging helps most with errors that vary across surveys. If one sample happens to contain too many supporters of A and another too few, those errors can partly cancel. If several pollsters miss the same kind of voter, the average can remain biased even though the polls agree.
In these two hypothetical sets of five polls, the candidate's actual support is 50%. Polls scattered around that share average to 50%, while polls that all underestimate support average to 47.5%, leaving a 2.5-point error.
An average can shift even when support is stable if an unusually favorable old poll drops out. Several new surveys showing a change in the same direction give stronger evidence that voter preferences have moved. After major news, an average may also mix responses collected before and after the event. The debate-timing example shows why a poll's publication date can conceal that difference.
How does an election forecast model calculate winning chances?
An election forecast model estimates a range of possible final results, then calculates how often each candidate wins within that range. It needs both an expected result and an estimate of uncertainty. The same poll lead can translate into different winning chances depending on how much room the model allows for error and changes in support.
Consider a hypothetical election between two candidates, A and B. A forecast puts A's expected final vote share at 51%, leaving B with 49%. That is a two-point lead. Assume the possible shares follow a bell-shaped normal distribution around 51%, and A wins by receiving more than 50%.
The calculation counts how much of that distribution falls above 50%. You could also simulate many possible results from the distribution and count the share in which A wins.
| Expected share for A | Assumed standard deviation | Approximate chance A wins |
|---|---|---|
| 51% | 2 percentage points | 69% |
| 51% | 4 percentage points | 60% |
The standard deviation measures how widely possible outcomes are spread around the expected share. In the first case, about two-thirds of the model's possible results lie between 49% and 53%. In the second, that range widens to roughly 47% to 55%. The wider distribution leaves more room for A to finish below 50%, bringing the race closer to even chances.
Both hypothetical forecasts center A's possible vote shares on 51%, with the winning results above 50% shaded. A wider spread of results, with a 4-point standard deviation instead of 2 points, leaves more room for A to lose and lowers the winning chance from about 69% to 60%.
A poll's sampling error is only one source of uncertainty in the final vote share. Polls can share mistakes, turnout can differ from expectations, and preferences can change before voting ends. Two forecasters can agree on the expected share while allowing different amounts of uncertainty around it.
You can also read this calculation backward. Under the same normal model, a 70% chance of winning corresponds to an expected share of about 51% with a two-point standard deviation, or 52.1% with a four-point one. A 70¢ winner contract can therefore be consistent with a close race. Inferring an expected vote share from that price still requires an assumption about uncertainty.
A's 70% chance of winning corresponds to an expected vote share of about 51% or 52.1% in these two hypothetical forecasts. The curves assume 2-point and 4-point standard deviations, respectively, so inferring a vote share from a 70¢ winner contract depends on the uncertainty assumed.
How election models combine state results
For an election decided across states, a model must account for how their results can move together. If polls underestimate the same party in several closely contested states, a single broad error can change the winner in all of them. Treating those states as unrelated would misrepresent the chance of winning the presidency.
B needs all three states to win the presidency in this hypothetical election and wins each state in three of ten simulations in both forecasts. Those wins coincide in none of the ten simulations with unrelated polling errors, compared with three when the same error affects all three states together.
The Economist's 2020 model combined national and state polls with economic and political information and allowed polling errors to be correlated across states. Its simulations could therefore produce an election in which the same party exceeded its polling estimates in several states at once.
Why do Polymarket and Kalshi odds differ from polls?
On Polymarket and Kalshi, a winner contract prices a chance of winning, while a voting-intention poll estimates support. A meaningful disagreement appears when a market and a forecast assign different probabilities to the same outcome at the same time. Trading conditions can also move a price away from what the electoral evidence suggests.
In the 51% vote-share example, changing the uncertainty alone changes the chance of winning. Traders can allow for different uncertainty, expect different turnout, or use information published after the latest polls.
A national poll estimates support across the country. Winning the US presidency requires a majority of 270 votes in the Electoral College, and most states award all their electors to the statewide winner. A candidate can lead nationally and still be less likely to win if their opponent has an advantage in the states that decide the contest.
A's wide win in State 1 helps A reach 52.4% of the national vote in this hypothetical election. Each of the five states has equal turnout and awards all ten electors to its winner, so B's three narrow state wins secure 30 of the 50 electors.
The resolution rules for Polymarket and Kalshi contracts specify whether a market concerns the popular vote, a state result, or a winner declared by a named source before a deadline. The forecast must predict the outcome defined by the settlement rules, including any conditions the title leaves out.
How trading affects prices
Market prices form through traders' buy and sell orders. Whether a purchase offers an expected gain depends on the price available for that trade. Suppose you estimate a candidate's chance at 63%, the displayed market probability is 60%, and the lowest selling offer is 65¢. A contract that settles at $1 for a win and zero otherwise is worth 63¢ under your estimate. Paying the 65¢ asking price gives you an expected loss of 2¢ before fees.
The lowest selling offer is 65¢ in this hypothetical market, while the displayed probability is 60%. At a trader's 63% estimate of winning, a contract that settles at $1 for a win and zero otherwise is worth 63¢, so buying at 65¢ has an expected loss of 2¢ before fees.
Polymarket's displayed probability normally uses the midpoint between the best buying and selling quotes, with a last-trade fallback for wide spreads. The order book shows the available prices and quantities. If few shares are offered near the current price, a purchase can exhaust them and reach higher offers.
Political markets have also sometimes priced unlikely outcomes too high and favorites too low. In Page and Clemen's study of Intrade, published in 2013, political outcomes priced at 20¢ had an estimated occurrence rate of about 11%, while those priced at 80¢ occurred at an estimated rate of about 93%. Both prices were closer to 50% than the observed frequencies, a pattern called favorite-longshot bias.
For political outcomes in Page and Clemen's Intrade study, a 20¢ price corresponded to an estimated occurrence rate of about 11%, and an 80¢ price to about 93%. Prices overstated longshots' chances and understated favorites' chances, the pattern called favorite-longshot bias.
A large trader can also hold a price away from other forecasts. In their study of Intrade's 2012 presidential market, Rothschild and Sethi documented large orders favoring Mitt Romney that kept prices stable while prices on Betfair moved. Removing those orders allowed Intrade's price to move sharply toward Betfair's.
Correcting a price also ties up money. After the 2020 election, an Augur market continued to imply about a 15% chance of the result being overturned for more than a month. Vitalik Buterin's account describes the cost of committing capital as one obstacle to trading against that price. A small expected gain may be less attractive than using the funds elsewhere before settlement.
Are betting markets more accurate than polls?
Prediction markets have beaten raw election polls in several studies. Comparisons with forecasts built from polls have produced mixed results. A poll records preferences during its fieldwork, while a market can already incorporate expectations about how those preferences will change.
| Comparison | Elections and timing | Finding |
|---|---|---|
| Berg, Nelson, and Rietz: Iowa Electronic Markets vote-share prices against 964 national polls, both expressed as two-party vote shares | Five US presidential elections from 1988 to 2004, with each poll matched to the market price on its final fieldwork day | Markets were closer to the final vote share 74% of the time, with a particularly strong advantage more than 100 days before the election. |
| Erikson and Wlezien: prices against polls projected forward to Election Day | The same five elections, with projections compared across campaign dates | Projected polls had an average absolute vote-share error of 2.13 points, against 2.41 for market prices. For the four elections with winner contracts, poll-based probabilities of winning the popular vote were closer to the outcome on 87% of dates. |
| Rothschild: Intrade against poll-based forecasts, with and without adjustments to market prices | 74 presidential state and Senate races in 2008, over the last 130 days | FiveThirtyEight slightly beat raw prices on average. After correction for favorite-longshot bias, market forecasts performed better, especially early in the campaign and in uncertain races. |
| Arnesen and Bergfjord: market predictions against raw and adjusted polls | The 2008 and 2012 presidential elections | Markets were more accurate. Adjusting polls made them, if anything, less accurate relative to markets. |
Erikson and Wlezien found that projecting polls forward reversed the markets' advantage. Arnesen and Bergfjord found no such improvement from adjusting polls in the later elections they studied. The result depends on which elections and campaign dates are tested, what the contracts predict, and how each forecast is adjusted.
Rothschild's correction for favorite-longshot bias moved probabilities away from 50%, so the adjusted research forecast differed from the price a trader actually saw. Erikson and Wlezien also attributed their popular-vote result partly to markets assigning too much probability to longshot candidates. The broader accuracy record includes other venues and kinds of event.
Testing forecast accuracy
A comparison needs the same outcome, races, and time before the result. Prices during vote counting incorporate evidence that an election-eve forecast could not have used. A collection dominated by safe seats is also an easier forecasting task than a collection of close races.
For vote-share estimates, an error measured in percentage points is useful. For win probabilities, the Brier score evaluates the probability assigned to the result, giving a larger penalty to confident mistakes. Counting only which candidate was favored treats a narrow favorite and a near-certainty as identical calls.
Daily prices on the same race all resolve against one result, and state races in the same election can share a national swing. A long series of prices contains fewer independent tests than its number of rows suggests.
A bundle of forecast records sits beside the ballot box for the election they describe. Each forecast can be checked against the final result, but they all share that one outcome.
Do election markets just follow the polls?
In US presidential elections from 1936 to 2008 with both market and polling data, prices near Election Day largely followed the polls. Erikson and Wlezien's historical study found that those prices added no predictive information once polls from the final campaign week were known. This record predates Polymarket and Kalshi.
The same study found that markets from 1880 to 1932, before scientific polling, predicted elections more accurately than markets from 1936 to 2008. When prices in the later period departed from polls, those differences increased forecast errors.
Shared polling inputs also explain why a market and a model can agree. A market may react immediately to a new poll while a published forecast waits for its next update. If both then move toward the same candidate, the agreement can come from that one survey. Traders can also react to reporting or revise their turnout expectations between polls, moving prices before another survey appears.
A hypothetical poll released Tuesday lifts the market's implied chance for A from about 58% to 63% within an hour, while a published forecast moves from 57% to 62% at Wednesday's update. The two end up close because both respond to the same survey.
Was Polymarket more accurate than polls in 2024?
On the eve of the 2024 election, Polymarket gave Donald Trump a 59% chance of winning, and he won. The models from Silver Bulletin, FiveThirtyEight, and The Economist gave him chances near 50%, so Polymarket scored better for that result. Polls are judged by their errors in vote margins, and AAPOR's review found those errors were relatively small in 2024.
Kalshi also favored Trump, at 55% in economist Rajiv Sethi's election-eve comparison. For the presidential result, the Brier score was about 0.17 for Polymarket and 0.20 for Kalshi, compared with 0.25 for a 50% forecast. Lower scores indicate smaller probability errors.
The forecasts on the eve of the 2024 election gave Donald Trump a 59% chance of winning on Polymarket and 55% on Kalshi. His victory gives them Brier scores of about 0.17 and 0.20, lower than the 0.25 for a 50% forecast because they assigned more probability to the outcome that occurred.
Sethi and colleagues' broader study tested Polymarket against those models across presidential, state, and congressional outcomes. It asked whether a trader following each model's probabilities would have made money against market prices. All three models failed to beat the market for the presidency, while some outperformed markets for less visible races. Undark's report on the study identifies congressional races as a particular weakness for Polymarket, which the researchers partly attributed to lower trading volume.
For presidential polls taken in the final two weeks, AAPOR found an average absolute error on the two-party margin of 2.6 percentage points nationally and 3.0 points in state polls. A small margin error can put the wrong candidate ahead in a close race while remaining close to the final vote shares. Counting correct winner calls can therefore make polls look worse than their margin errors suggest.
In the hypothetical close race, a 2.6-point error separates the poll's 1-point lead for A from B's 1.6-point win. In the wider race, the poll has A ahead by 7 points and A wins by 3, so it has the right candidate ahead despite a larger 4-point error.
Trump's victory alone cannot establish whether Polymarket's 59% price was well calibrated. That requires many comparable forecasts near 59%, with the predicted outcome occurring about 59% of the time.
Do polls predict elections better when they ask who will win?
Voter expectations polls ask who will win, while voting-intention polls ask whom people plan to support. In a study of US state-level presidential outcomes from 1952 to 2008, the expectation question identified the winner in 60 of the 77 cases where the two questions pointed to different candidates.
When the two questions pointed to different candidates, asking who would win identified the winner in 60 of 77 cases, compared with 17 for asking whom people planned to vote for. Rothschild and Wolfers's study compared answers from the same respondents across US state presidential elections from 1952 to 2008.
The researchers, David Rothschild and Justin Wolfers, compared answers from the same respondents in American National Election Studies surveys. A respondent could prefer A and expect B to win. Their expectation could draw on reporting, conversations, or their sense of the wider electorate. As in a market forecast, they were judging what other voters would do.
If 70% of respondents name A as the likely winner, the number measures agreement among respondents. Each person might think A is barely favored or almost certain to win. Converting those answers into a win probability takes another model or questions that ask for probabilities directly.
In forecasting tournaments, a “prediction poll” asks participants for numerical probabilities. Combining those estimates can involve giving more weight to recent forecasts and to people with stronger forecasting records.
Why do election betting odds change after a debate?
Election odds can change as soon as traders revise their orders after watching a debate. A new poll takes time to collect responses, process them, and publish results. The two measures can therefore reflect different information even when they appear beside each other in the same news report.
Suppose a survey collects answers from Monday through Wednesday, a debate happens Thursday, and the poll is published Friday. The Friday release contains answers collected before the debate. Its field dates, the dates when people answered, establish that timing. Pollsters disclose them alongside sampling and weighting information under AAPOR's disclosure standards.
Monday through Wednesday's survey answers reach the public on Friday in this hypothetical week. Thursday evening's debate has already moved the market price by then, so the newly published poll still describes views from before the debate.
A market price recorded on Thursday night can already reflect how traders expect the debate to affect the race. For a nomination contract, that could include a greater chance that a candidate withdraws, even if voter preferences have barely changed.
Subsequent polls help establish whether voters' preferences changed, particularly when their fieldwork clearly follows the debate. Pew notes that opinion shifts immediately after major events can be short-lived. Surveys collected several weeks later help show whether the change lasted.
The same timing issue applies to a court ruling, a candidate's withdrawal, or an unexpected economic release. Markets can respond while surveys still reflect the earlier situation, and prices can reverse as reporting develops before any new poll has appeared.
How can you compare election forecasts with market odds?
Compare a forecast probability with the market price for the same outcome at the same time. The polls underneath the forecast help explain its estimate of support, while its methodology explains how uncertainty turns that support into a chance of winning.
Suppose a hypothetical two-candidate state election has A at 51% and B at 49% in polls collected before Thursday's debate. On Friday at noon, a model using those polls gives A a 60% chance of winning. At that moment, A's winner contract trades at 70¢ and settles on the same state result. The ten-point difference between the model's probability and the market's implied probability is the disagreement to explain. Subtracting the 51% poll share from the market's 70% would compare different quantities.
A's 51% poll share measures support in this hypothetical state election, while the model and market estimate A's chance of winning. At noon Friday, the model's 60% estimate and the 70% implied by A's 70¢ winner contract are 10 percentage points apart.
Traders could expect the same narrow lead as the model but allow less room for the final share to differ, giving A a higher chance of winning.
Comparing price moves across election markets helps distinguish news about A from a change in expectations for the whole election. A move confined to A's race might follow news about that candidate, while shifts across several races could follow a national poll or revised turnout expectations.
Over 24 hours in the first hypothetical example, the price for A's party rises 10 probability points in A's race and changes by at most 1 point in each of the other four. In the second, prices rise by 3 to 5 points across all five races, suggesting a broader change in expectations for the party.
A gap between a forecast and a market price has several possible explanations, and each one leaves evidence in a different place.
| Possible reason for a gap | Evidence and where to find it |
|---|---|
| The sources describe different results | The pollster's population and question, and the venue's contract settlement rules |
| The market reflects newer information | Field dates in the pollster's release, the forecaster's update time, and the market's price history |
| Traders expect different turnout or allow different errors | The forecaster's methodology, likely voter assumptions in poll releases, and subgroup estimates |
| A trade used up nearby offers, or the displayed probability differs from the purchase price | Available prices and quantities in the venue's order book, alongside its recent trades |
| Large orders hold a price in place | The venue's order book and transaction history, compared with prices for the same outcome elsewhere |
| Longshot prices overstate winning chances | Historical calibration research on comparable contracts, with similar time left before settlement |
| Committing capital makes a small expected gain unattractive | The contract's settlement timing and the return from holding it over that period |
After Thursday's debate, online praise for A's answers or criticism of B's performance could help explain why traders give A a higher chance than the model does. These reactions come from people who choose to post, so online discussion can differ from voters' views.
One hypothetical post praises A's answer on housing, while the other criticizes B for dodging A's question. Both mention A, and the second receives a negative label because sentiment is assigned to the whole post.
If polls collected after the debate continue to put A near 51%, they weaken the case that the debate produced a lasting rise in support. The 70¢ price could still reflect traders allowing less room for an upset around that narrow lead.
History of Prediction Markets
How prediction markets developed from election trading into online exchanges, and why earlier markets closed or never opened.
Markets vs. Sportsbooks
Sportsbook odds and prediction-market prices become comparable when expressed as the cost of the same return, including exchange fees.