Polls and election odds: compare the questions before the percentages

Identify which quantity each number describes

A poll percentage and an election winning chance can disagree because they answer different questions. Before comparing the numbers, identify the poll’s population and question, the forecast’s winning event and horizon, and the contract’s payout rule and quote time.

Support for a candidate, intention to vote for that candidate and an expectation that the candidate will win are different survey questions. A percentage of respondents is not automatically a share of future valid votes. Neither percentage, by itself, gives the probability of winning. [1][3]

For a broader introduction, see how a market probability differs from a poll. Here the task is narrower: make a comparison record that connects the survey, the forecast and the contract without treating their percentages as interchangeable.

Read the poll’s population, fieldwork and error limits

Start with the exact question and target population. A poll of adults, a poll of registered voters and a likely-voter estimate do not have the same denominator. Check the number of respondents behind the reported result, the sampling frame and contact mode, any weighting, and whether undecided respondents or third choices remain in the percentage base.

Next, separate fieldwork from publication. A result released today may describe answers collected earlier. Changes between those dates are not evidence that either the survey or a later price measured the same moment incorrectly. The survey record should make clear which period its answers represent.

A published margin of sampling error is not a measure of every possible error. It belongs to a specified estimator, sampling design and confidence procedure. Coverage gaps, nonresponse, wording, contact mode and assumptions about who will vote can add uncertainty that the margin does not summarize. A 95% confidence level is not a 95% chance that a candidate wins. [1]

Check uncertainty for the actual comparison you want to make. A subgroup, a weighted estimate and a difference between candidates can require different calculations from the overall percentage. Do not automatically add or double every published margin, or transfer a classical probability-sampling margin to an opt-in or quota survey without a justified method. This guide does not supply a replacement margin formula. [1]

A forecast needs more than a poll percentage

A forecast must define what counts as winning and when that event occurs. To turn survey evidence into a winning chance, a model needs assumptions about possible future results and the probability attached to each. Those assumptions can include turnout, how undecided responses are allocated, future changes in preferences and uncertainty in the model itself. A poll percentage alone does not supply that distribution. [2][3]

Keep uncertainty about a survey estimate separate from uncertainty about what happens after the survey. Also check the electoral stage: a national vote-share estimate may not answer a question about a nomination, a particular constituency, seats or a winner chosen through a different electoral rule. A forecast can only be compared with a contract after their events and horizons are matched.

The exercise below isolates one reason for disagreement. It sets out every possible result in two small invented models. It is not a method for converting a real poll into odds, and none of its weights comes from a cited study or an actual market.

Same expected vote share, different chances of winning

Imagine an election with only A and B. All valid votes go to one of them, and A wins exactly when A’s valid-vote share is greater than 50%. The listed cases have no ties. They exclude third candidates, undecided categories, abstentions, invalid votes, seats, an electoral college, ranked-choice transfers and recount or tie-breaking rules.

M1 and M2 are two separate fictional models. Each model’s two cases are mutually exclusive and exhaustive: only one can occur, and their probabilities sum to one. They are not independent events or two simultaneous results. The probabilities are assumptions chosen for this example, not estimates of an actual election.

The expected vote share is 51% for A in both models, but A’s winning chance is 50% in M1 and 75% in M2. The mean averages vote shares; the winning chance adds the probability of the cases above the winning threshold. The gap between the two winning chances is 25 percentage points, not a 25% relative increase.

A’s mean of 51% is not even one of the possible results in either model. A mean is not automatically the most likely outcome. For the general method, read how to separate the mean, a possible result and the chance of a positive result. These invented distributions do not establish a fair contract price, a forecast’s accuracy or a profitable trade.

Match the forecast to the contract and its observation time

Read the contract’s event before interpreting its price. Does it resolve on a named winner, vote share or a result within an interval? Which electoral stage and jurisdiction does it cover? Record the named result source, deadline or expiry, and the treatment of corrections, void outcomes and disputes. A headline about an election does not supply those terms.

In a deliberately simplified binary contract that pays one unit if the specified event occurs and zero otherwise, the price as a fraction of that maximum payout can be read as an implied probability under assumptions. That interpretation does not make the price the true probability or a percentage of voters. It also does not describe every contract: a vote-share payout measures something different from a winner payout. [3]

Use the quote that matches your question. A bid is not an ask, a last trade may be old, and a midpoint is not a promise that an order can be filled at that price. Record the timestamp, currency or payout unit, spread and available depth. The existing guide explains how to calculate implied probability; a conversion alone does not settle whether the quote is comparable.

Market participants self-select. A price does not give each person’s view equal weight: stakes, wealth, risk preferences, position constraints and liquidity can matter. Trading volume is not a representative voter sample, and it does not prove that information is independent or accurate. These are mechanism limits, not a present audit of any platform or proof that a particular market is manipulated.

Record disagreement without choosing a guaranteed winner

Put the three records beside each other before deciding what the disagreement means. Start with the original survey question and methods, then the forecast’s inputs and distribution, then the contract’s rules and quote. If you cannot identify a field, mark it unknown instead of filling it with a familiar headline or another source’s number.

Some apparent disagreements disappear when you identify different questions or dates. Others remain after those checks because the model’s assumptions, information or price conditions differ. This record makes those differences visible; it is not a validated way to improve forecasts or returns, and it does not prove that either number must be right.

Different numbers do not establish a profitable trade

Historical comparisons do not give a permanent ranking of polls and markets. Reade and Vaughan Williams’s 2019 paper considers historical US election evidence from 2008–2012; Erikson and Wlezien’s 2008 paper examines historical Iowa Electronic Markets evidence from 1988–2004. Their questions, models, samples, horizons and evaluation measures matter. Neither scope supports a current global accuracy ranking or a claim that one method always wins. [2][3]

Disagreement does not, by itself, prove poll fraud, market manipulation or immunity to manipulation. This guide does not assess an actual candidate or current race, choose a winner, or provide campaign, persuasion or tactical-voting advice.

Even a correct winning prediction does not guarantee profit. Gross payout is different from net profit after the entry price and applicable costs. Fees, spread, the amount that can actually be filled, contract rules and settlement risks still matter. Real-money activity can lose principal; it is not dependable income or a way to recover previous losses. Choosing not to participate is a valid outcome.

Language, website access and an educational explanation do not confirm personal eligibility. Residence or location, service terms and local law are separate matters. Access, signup, KYC, funding, trading, withdrawal and eventual receipt outside the service are distinct checks; this guide tests none of them. It does not confirm that an election contract is currently offered.

Questions about polls, probabilities and prices

Does 51% support mean a 51% chance of winning?

No. Support is a survey quantity for a specified question and population. A winning chance concerns a defined future outcome. You need a future result distribution, a winning rule, a horizon and model assumptions before making that conversion.

Does a 95% confidence level mean a 95% chance of victory?

No. It describes a sampling procedure’s confidence coverage for a specified estimator and design, not a candidate’s winning chance. It also does not include every possible survey error or all uncertainty about changes before the election. [1]

Can forecasts share a mean vote share but have different win probabilities?

Yes. In the complete fictional M1 and M2 examples, A’s mean valid-vote share is 51% in both models. With A winning only above 50%, the assumed case weights give winning chances of 50% and 75%. The distributions, not the mean alone, determine those chances.

Does disagreement prove the poll or market is wrong?

No. First match the question, population or electorate, dates, future event, model assumptions and contract terms. A remaining difference needs interpretation; it does not establish a universal accuracy ranking or show which number is the true probability.

Does reading a favorable probability guarantee profit or permission to trade?

No. Losing outcomes, the entry price, costs, liquidity and contract rules remain relevant. Account and legal eligibility require separate checks. Reading this guide is neither a recommendation to trade nor confirmation that a service or particular contract is available to you.

Sources and limits

This guide uses Andrew Mercer’s 2016 Pew Research Center explanation, the saved accepted-version prefix of Reade and Vaughan Williams’s 2019 paper, and a saved prefix of Erikson and Wlezien’s 2008 paper. The consultation date refers to saved text, not study or article publication. The paper prefixes are not full methods or appendix reviews. Their historical US samples do not establish current accuracy rankings. M1 and M2 are original invented distributions; no source paper supplies their weights. No current poll, election odds, platform fees or service eligibility was checked.

1. Understanding the margin of error in election polls
2. Polls to probabilities: comparing prediction markets and opinion polls
3. Are Political Markets Really Superior to Polls as Election Predictors?