What a poll actually measures

The prediction-market contract

A binary claim that pays a fixed amount (commonly $1) if the stated outcome occurs — a candidate wins, a party takes a chamber — and nothing otherwise, cash-settled to the certified result. Its price is, directly, the market-implied probability of that outcome.

The poll (a survey)

A measurement of stated opinion in a sample at a moment: whom respondents say they would vote for, whether they approve of someone, or how they feel about an issue. It has no settlement and no payoff — it is not a claim on the outcome, so there is nothing for it to be right or wrong about in the way a contract is. Read as a probability it is a category error: it reports opinion, not odds.

A poll measures support — a snapshot of stated preference in a sample, carrying a margin of error and a set of methodological choices, at one point in a campaign. It is not designed to answer "who wins"; it answers "who is preferred, among those asked, right now."

The Scopes Fair Value for a race is not a poll. It is a forecast model's win probability: the output of a system that takes many polls, weights them by quality and recency, corrects for known house effects, and combines them with turnout, undecided behavior, the electoral structure, and the time left until the vote to produce a single probability of the outcome. The model is the machinery that turns opinion into an outcome probability — the same quantity the market contract prices.

That is why the desk compares the market to a model and never to a raw poll: the model and the contract measure the same thing (the probability of the outcome), while a poll measures a different thing (present opinion). A raw poll placed next to a market sets a snapshot of preference against a probability of victory, and the distance between them is mostly the distance between those two ideas.

Support (vote-choice) polls

The familiar "who would you vote for" poll. It measures present preference among those sampled, before undecideds resolve and before turnout is known. It is an input to a forecast, not a forecast itself: a lead inside the margin of error is not a coin toss, and a lead outside it is still not a probability of winning. Comparable to a market only after a model converts it into an outcome probability.

Expectation (citizen-forecast) polls

The "who do you think will win" poll. Because it asks respondents to forecast the outcome rather than state a preference, it is the poll type that most nearly approximates an outcome probability, and it has often been found more accurate than vote-choice polling. Even so it is aggregated opinion about the outcome, not a market clearing price, and it carries biases of its own.

Approval and favorability polls

Measures sentiment toward a person or an institution — a job-approval number, a favorability rating. It is not attached to any contest and has no outcome to settle against, so it is not comparable to a prediction-market contract at all, however often the two are discussed together.

Issue and policy polls

Measures opinion on a topic — support for a policy, priorities among voters. Like approval polls it describes the state of opinion, not the probability of a future event, and belongs nowhere near a market price on an outcome unless a specific, settleable event is attached to it.

Support share is not a win probability. The map from a candidate's polling share to a probability of winning is steep near a tied race and flat far from it: a 2-point lead can mean very different odds depending on how many voters are undecided, how they are expected to break, and how much time remains. Treating "ahead in the polls" as "likely to win" flattens that curve into a straight line it is not.

A margin of error is not a probability. A 3-point margin of error describes sampling uncertainty around the measured share; it says nothing directly about the chance of winning, and a lead "within the margin of error" is not the same as a 50-50 race. The two numbers answer different questions, and one cannot be read off the other without a model.

Opinion moves; an outcome resolves once. A poll is a snapshot that shifts as a campaign unfolds, while the contract settles a single time on the certified result. A gap between a market and a model built from polls can be the market pricing news the latest polls have not captured, or the model weighting history the market is discounting — a difference of information and method, not a scoring error on either side.

A forecast model is not an instrument. Unlike a call spread against a digital, or a sportsbook line against an event contract, a poll or a model cannot be held against a market position: there is nothing to buy, no other side to take, and no settlement on the model itself. So a gap between a market price and a model's probability is not a discrepancy anyone can act on by trading — it is information about where two methods disagree.

That is also why the comparison has to be built with care. Placing a raw support poll next to a market invents a false gap — the apparent "disagreement" is mostly the distance between a preference snapshot and an outcome probability, not a real difference of view about who wins. The Poll desk avoids that by comparing the market only to a forecast model's win probability, the one number that answers the same question the contract does.

So the honest reading of a gap between a market and a model is modest: it marks where market sentiment and a fundamentals-and-polls model diverge on the probability of the same outcome. It does not say which is right — that is what the calibration record measures, over many resolved races rather than one.