Poll, Forecast, Market: Three different questions

In brief

A poll, a forecast and a betting price can disagree without any of them being wrong, because each makes a different kind of claim. A poll counts what people told an interviewer on a given day. A forecast states a probability that nobody can grade on a human timescale. A price records a trade on one venue, and the cheapest Trump and Harris contracts together cost under $1 on nearly every day, so there was never one market probability. Comparing a same-day poll with a market price is invalid.

How to use this

Before comparing a poll, a forecast and a price, name the claim each one makes: a count of stated preferences, a probability, or a trade on one venue. If it is a price, check which exchange and what fees applied, and whether the contracts for both outcomes together cost less than a dollar. Do not grade a forecast on a single election result; the data to do so does not exist yet. When you see an accuracy headline, ask which contracts made up the sample. Use each number for the decision it can actually inform. For regulators, focus on venue rules, not on whether a price was a good measurement.

What the story is about

A poll and a betting price are answers to different questions, even when both are taken on the same afternoon. A poll counts what people say they will do. A price reflects what traders think will happen on election day itself. Erikson and Wlezien (2008), in a peer-reviewed study of trade on the Iowa Electronic Markets, an academic exchange, between 1988 and 2004, argued that comparing the two on the same day is inappropriate: "market prices reflect forecasts of what will happen on Election Day whereas trial-heat polls register preferences on the day of the poll". The same study found that poll-based forecasts beat vote-share market prices once poll leads were discounted properly. The evidence is old and small, from markets with position limits in the hundreds of dollars, and says nothing about a market of today's size.

A forecast can be sound and still nobody can grade it. Grimmer, Knox and Westwood (2024), in a preprint, show that American presidential elections are too rare to score a probability against. State results add little, because a winner needs state winners added up in a weighted way, and the errors within a single year move together. The authors conclude that scientists and voters are "decades to millennia away from assessing whether probabilistic forecasting provides reliable insights into election outcomes". That is a limit on what can be evaluated. It does not say any particular forecast is wrong, only that the data to grade it does not exist yet.

The 2024 markets' headline accuracy comes mostly from contracts that were never in doubt. Clinton and Huang (2026), in a preprint covering the final nine weeks of one US election, report that 86% of actively traded markets beat a coin-flip. That figure mostly reflects which contracts made up the sample. On the eve of election day, 66% of active markets were priced below $0.20 or above $0.80, and 96% of those resolved as priced. Among the 12% priced between $0.40 and $0.60, 55% resolved in favour of the priced favourite. The authors also make a definitional point: a set of contracts can be accurate without being well priced, since contracts that all sit at $0.55 and all resolve true score 100% accurate while their price was still too low. The study covers four exchanges, one cycle, one country.

A market price looks like one number. Here it was several. Clinton and Huang (2026) checked two exchanges, Kalshi and Polymarket, where a contract on a Trump win and a contract on a Harris win could be bought side by side. After each exchange's fees, the pair of cheapest contracts covering those two outcomes could be bought for less than a dollar. Two prices for the same event that do not add up to $1 cannot both be the probability of that event. So the market's probability was not one quantity, and the same event did not carry the same price on two venues. The dollar figures are weaker than they look. The $3.2 billion headline is a number Clinton and Huang attribute to a news report, and Yang and Tsang (2026), in another preprint, show that the raw totals recorded on a public blockchain overstate what a conventional exchange would count as trading, by about two and a half times in the case they picked apart. Read the dollars as a platform's own account of itself.

The popular reading of 2024 is that the markets called it while the polls missed it. Three things in the same preprint undercut that reading (Clinton and Huang, 2026). The accuracy figure rests on a sample dominated by contracts that were already near certain. The band where the race was genuinely close went 55% to the priced favourite. And prices for the same event differed across exchanges instead of converging as information arrived. The comparison itself is also invalid as usually made, for the reason Erikson and Wlezien set out. Clinton and Huang do suggest markets may provide useful signals, and they qualify that immediately.

Several things are still open. The evidence is all from US presidential races, so nobody knows whether markets behave the same way in a quieter contest or in a country with more than two parties. Whether rules that make a venue easier to trade on make its prices match outcomes better, or only bring in more trading, is untested. Whether the gaps between exchanges close as these markets mature is unsettled. Nothing yet establishes how a reader weighs a market price against a poll margin. And if a probability cannot be graded on a human timescale, the field has not agreed how it should be graded instead.

Each number is a different kind of claim, so they can disagree without any of them being wrong. A poll counts what people told an interviewer on a given day. A forecast states a probability about something that has not happened, and in this domain nobody can mark it. A price records a trade between two people at a moment, on one venue, under that venue's rules. The three can point in different directions, and none of them has necessarily failed. A poll and a price from the same week can disagree sharply and both still be doing exactly what they were built to do.

So what

In practice, these numbers arrive side by side, often in the same news story, and the temptation is to rank them. Ranking makes sense only if they are answering the same question, and these three are not. So before comparing them, say which claim each one is making. That single step tells you whether a disagreement between them is news, or just three different questions answered at once.

For political parties

A party reading all three should treat each as what it is. Polls describe preferences today, and their error is measurable, cycle by cycle. Forecasts give a probability that nobody can grade, which is useful for planning under uncertainty and useless as a claim about what will happen. Prices are trades on a particular venue, under that venue's rules and fees, and the same event can carry different prices on different exchanges. A party that leans on whichever number flatters its strategy has paid for a mirror rather than intelligence. The test is simple: does the number serve the decision you are making, or the conclusion you already preferred?

For government

Regulators face a choice about what they are regulating. Treat a price as a measured probability and the job becomes deciding whether the measurement is accurate. Treat it as a trade and the job becomes the venue: which contracts it lists, who may buy them, what it charges, and how those rules shape the numbers on the screen. The 2024 record supports the second reading. Two exchanges quoting different prices for the same event is a fact about venues and their rules, not about the candidates. That is where regulatory attention does the most good, and it is also where a government gets an honest reading of what the market is telling it.

Case studies

The cleanest demonstration is a test anyone can run. From 1 September to 4 November 2024, on Kalshi and Polymarket, buying the cheapest Trump contract and the cheapest Harris contract together cost less than a dollar on nearly every day, after each exchange's fees (Clinton and Huang, 2026). That gap widened over the final two weeks and peaked in the last days, then closed to nothing on election day once results began to be reported. Two contracts covering the only two outcomes, priced together below a dollar, cannot both be carrying the true probability of their outcome. That is why the market's probability was never a single number.

References

Clinton, J.D. and Huang, T. (2026) Prediction markets? The accuracy and efficiency of political prediction markets in the 2024 presidential election. SocArXiv [preprint], v5, 10 June. Available at: https://doi.org/10.31235/osf.io/d5yx2_v5 (Accessed: 9 September 2026).

Erikson, R.S. and Wlezien, C. (2008) 'Are political markets really superior to polls as election predictors?', Public Opinion Quarterly, 72(2), pp. 190–215. Available at: https://doi.org/10.1093/poq/nfn010 (Accessed: 9 September 2026).

Grimmer, J., Knox, D. and Westwood, S.J. (2024) Assessing the reliability of probabilistic US presidential election forecasts may take decades. OSF [preprint], 26 August. Available at: https://doi.org/10.31219/osf.io/6g5zq (Accessed: 9 September 2026).

Yang, Z. and Tsang, K.P. (2026) The anatomy of a blockchain prediction market: Polymarket in the 2024 U.S. presidential election. SSRN [preprint]. Available at: https://doi.org/10.2139/ssrn.6336679 (Accessed: 9 September 2026).