Markets vs Polls, Head to Head: Expectations vs intentions, again

In brief

Neither can yet be called better. In the only head-to-head that exists, 41 US contests, which side wins depends on which race you score and whether you score it for accuracy or betting profit. On the presidency, accuracy favoured Polymarket while betting profit favoured the models. On 13 House races, betting profit favoured the models. The market's edge also tracked trading volume rather than method. Treat any headline that crowns one instrument as a choice of contract and scoring rule, not a settled result.

How to use this

Before repeating a claim that polls or markets won, ask which races were scored and whether the score was accuracy or betting profit. Check the trading volume behind the market contract: the 13 House races traded far less than the presidency. Check whether the poll side was an average projected forward to election day, not a raw reading from one day. Treat a market price and a poll as answering different questions. If you advise a party or government, keep the two signals separate and state your scoring rule in advance. Do not crown one instrument from a single headline number.

What the story is about

If you want to know which to trust, the obvious test is a straight comparison: the odds on a market such as Polymarket against the average of the polls. That test does not exist in any pre-declared form. No study scores market prices against poll averages under a rule agreed in advance. What exists compares market prices against statistical models, which blend polls and other information into a running forecast. Those comparisons do not agree with each other. In a 2021 preprint, Sethi et al. (2021) tracked one such model against the PredictIt market across the 2020 battlegrounds and found the market more accurate months out, the model more accurate near the vote, and a simple average of the two better than either.

By the polling profession's own audit, 2024 was a good year for the polls. The American Association for Public Opinion Research (2025) task force examined 611 polls of presidential, senatorial, gubernatorial and congressional contests and put the average absolute error at 3.3 points on the two-party margin, against 5.3 in 2020 and 5.2 in 2016. National presidential polls missed by 2.6 points, and state-level presidential polls by 3.0, their best showing in any presidential cycle since 1944. The same report cuts the other way: those polls overstated Democratic margins by 2.7 points across all offices.

Forecasts can be scored two ways. One asks how close a forecast came to the outcome, using a Brier score, where lower is better. The other asks whether you would have made money betting on it. On the same 2024 data, those two scoreboards rank the instruments differently. On the presidency, Brier scores put Polymarket 0.229 ahead of Silver Bulletin 0.250, The Economist 0.276 and FiveThirtyEight 0.301. On the popular vote the order reverses: FiveThirtyEight 0.478, Silver Bulletin 0.497, Polymarket 0.513, The Economist 0.584. Accuracy can also reward luck. Silver Bulletin carried a convention-bounce adjustment that inflated Trump's odds for a fortnight in August and then faded. Because Trump won, "the average daily Brier score rewards the adjustment, treating it as prescience rather than good fortune" (Sethi et al., 2025).

Scored by betting profit instead, the picture changes again. In the peer-reviewed head-to-head, Sethi et al. (2025) gave a virtual trader each model's forecast and let it bet against Polymarket's prices. All three models lost double digits on the presidency. In the 13 House races, both models that published congressional forecasts made money against the market. Their summary: "all models failed to beat the market in the headline contract but some did so convincingly in contracts referencing less visible races". Volume tracks the same split. The presidency contract traded $3,686,335,059, while the 13 House races traded $2,262,496.

The loudest pro-market claim of 2024 is that Polymarket beat polling. It comes from Cutting et al. (2025), a preprint whose abstract states that "Polymarket was superior to polling in predicting the outcome of the 2024 presidential election, particularly in swing states". The comparison has two limits. It covers one race, the presidency, and it sets market prices against raw polling readings rather than an average projected forward to election day. That mismatch is the one Erikson and Wlezien (2008) warned about in a peer-reviewed journal: a market price forecasts election day, while a poll measures the day it was taken.

One presidential election is a thin basis for grading anything. In a preprint, Grimmer, Knox and Westwood (2024) argue that presidential elections are too rare, and their errors too correlated, for probabilistic forecasts to be graded in a human timeframe. Sethi et al. (2025) concede the top of the ticket: on the presidency and popular vote in 2024, "it is indeed the case that a coin flip would have produced a more accurate forecast". Averaged across the competitive states and congressional races, all three models and the market beat a naive guess. The disagreement is about which races to score, not the arithmetic.

The evidence base is thinner than those comparisons suggest. Every post-2022 head-to-head is American and from 2024, and the 41 contests in the anchor study are all general-election races in one year with heavily correlated outcomes, the weakness Grimmer and colleagues name. Two of the three models come from the same product family and share an information environment, and models and markets read each other, so calling them independent instruments is an idealisation Sethi et al. themselves flag. Most of the 2024 evidence comes from a conference paper, a preprint and a professional task-force report, which are weaker than the journal baselines.

The honest answer is that neither can be called better yet. In the only head-to-head that exists, the 41 US contests of 2024, which side wins flips with the race you score and the rule you score it by. Score the presidency for accuracy and the market looks sharp. Score it for betting profit and the models lose. Score the House races for profit and the models win. Anyone handed a single number from that set can be told a tidy story, and the story can be true and misleading at the same time.

So what

A headline saying the polls beat the markets, or the markets beat the polls, can be produced from the same year of data by choosing a contract and a scoring rule. So the useful question is narrower: which race do you need to read, and which decision does your reading feed? Ask which races were scored, and which scoring rule was used, before you see the result, and the answer will tell you something you can use.

For political parties

A party should treat polls and markets as two separate signals, never as an oracle. The case for humility is on the record. Gruca and Rietz (2024), writing in a peer-reviewed journal, published the Iowa Electronic Markets' reading on 29 September 2024: an 87% chance the Democrat would win the popular vote, with a six-to-seven point margin. The Republican won the popular vote. The paper graded nothing, because it appeared before the vote, and it flagged that uncertainty remained. Copy that discipline. A market price and a poll answer different questions, and research built to agree with the conclusion your leadership already prefers stops being intelligence and becomes a mirror. Flattering research gets found out when the votes are counted.

For government

The safest posture for government is to refuse to crown either instrument as official truth, because the evidence does not support the choice. Until someone commits in advance to a head-to-head between market prices and poll averages, and scores it in public, any claim that one beats the other is a pick made after the fact. The cheap response is transparency: say which race you are reading, which scoring rule you used, and how much trading or polling sits behind it. Then outsiders can check whether the conclusion came from the data or from the framing. A government that stakes its credibility on one number has no early warning left when that number fails.

Case studies

One dataset shows how easily the verdict flips. Sethi et al. (2025) scored FiveThirtyEight's model against Polymarket contract by contract. On the popular vote the model's virtual trader returned +17.9% against the market. On the presidency the same model lost 14.6%. One election, one model, one market, and opposite verdicts depending on which contract you score. The thirteen House races repeat the pattern in the least traded contracts of the study, returning 47.4% and 34.2%, where $2.26m changed hands. That is the clearest evidence available that the market's edge tracks attention rather than method. Silver Bulletin's convention-bounce adjustment completes the picture, with the best average Brier score of the three models on the presidency and the worst trading return, because an adjustment that later reversed happened to point at the eventual winner.

References

American Association for Public Opinion Research (2025) Task Force on 2024 Pre-Election Polling: an evaluation of the 2024 general election polls. Chaired by J. Pasek. Alexandria, VA: AAPOR, 29 October. Available at: https://aapor.org/announcements/2024-pre-election-polling-report/ (Accessed: 10 September 2026).

Cutting, L.E., Hughes-Berheim, S.S., Johnson, P.M., Baroud, H. and Goldstein, B. (2025) Are betting markets better than polling in predicting political elections? arXiv [preprint], 2507.08921, 11 July. Available at: https://doi.org/10.48550/arXiv.2507.08921 (Accessed: 10 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: 10 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: 10 September 2026).

Gruca, T.S. and Rietz, T.A. (2024) 'Iowa Electronic Markets: forecasting the 2024 US presidential election', PS: Political Science & Politics, 58(2), pp. 258–266. Available at: https://doi.org/10.1017/s1049096524000921 (Accessed: 10 September 2026).

Sethi, R. et al. (2021) Models, markets, and the forecasting of elections. arXiv [preprint], 2102.04936, v4, 25 May. Available at: https://doi.org/10.48550/arxiv.2102.04936 (Accessed: 10 September 2026).

Sethi, R. et al. (2025) 'Political prediction and the wisdom of crowds', in Proceedings of the ACM Collective Intelligence Conference (CI 2025), San Diego, 4–6 August 2025. New York: Association for Computing Machinery, pp. 214–225. Available at: https://doi.org/10.1145/3715928.3737483 (Accessed: 10 September 2026).