Research Library
Do Prediction Markets Beat Polls?
For more than thirty years, economists have tested small-stakes election markets against the polls. Here is what they found, why markets often win, and where the evidence is mixed.
Usually, yes, but not always, and not by as much as their fans claim. University-run markets where traders could risk only a few hundred dollars beat individual polls most of the time, especially months before an election. Careful poll models narrow the gap, and markets have had famous misses. The strongest result in the literature is that combining markets and polls beats either one alone, which is how our Race Center forecast is built.
74%
Of 964 national polls across five presidential elections (1988–2004), the Iowa Electronic Markets were closer to the final result 74% of the time. More than 100 days out, the market was significantly better in every election.
Source: Berg, Nelson & Rietz, International Journal of Forecasting (2008)
Iowa Electronic Markets vs. final polls
49 markets, 13 countries
Iowa Electronic Markets vs. final Gallup poll
Four U.S. presidential elections
Source: Wolfers & Zitzewitz (2004)
Average miss on each side's share of the two-party vote, in percentage points. Shorter is better.
The markets that built this research record were deliberately small. Under its CFTC staff letter, the Iowa Electronic Markets let each trader risk no more than $500, and the average stake was a tenth of that. PredictIt, run by Victoria University of Wellington in New Zealand, was limited to $850 per contract and 5,000 traders per market from 2014 until 2025.
Even with stakes that low, the Iowa markets beat the polls. A 2004 study found that a play-money market forecast football games as well as a real-money one, which suggests the discipline of keeping score matters more than the size of the bet. In 2008, 22 economists writing in Science asked regulators to create a safe harbor so more small-stakes markets could run for research.
Money on the line
A poll asks whom you prefer. A market asks who you think will win, and charges you for being wrong. Traders who know something have a reason to act on it, and traders who are guessing lose money and tend to drop out.
Many kinds of information in one number
A price blends polls, fundraising, turnout reports, local knowledge and news as it breaks. Economists since Hayek have argued that prices gather scattered knowledge no single analyst holds.
A forecast, not a snapshot
A poll measures opinion on the day it was taken. A market price is a bet on Election Day, so it already allows for the time left in the race. That is why the market's edge is largest months out.
It updates every minute
Polls are released days apart and take days to field. Market prices move the moment a debate ends or a scandal breaks.
Fair comparisons narrow the gap
Erikson and Wlezien showed that once poll leads are discounted for the time left before the election, poll-based forecasts beat the Iowa markets. Much of the market's edge is over raw polls, not over good poll models. Source
Late prices may just follow the polls
In a second paper, Erikson and Wlezien found that once scientific polls existed, betting prices in the final week added nothing beyond the last polls, and in 1948 they followed the polls in making Dewey the favorite. Source
Longshots are overpriced
Bettors pay too much for unlikely outcomes and too little for favorites (Snowberg & Wolfers, 2010). A 10-cent contract usually means less than a 10% chance.
Markets have missed too
In 2016, betting markets gave Clinton more than 80% in the final days, and in 2022 PredictIt gave Republicans more than 70% to win the Senate. Polls missed in 2016 as well. Source
Thin markets can be pushed
In 2012 one Intrade account placed more than a third of all bets on Romney to win and lost close to $7 million (Rothschild & Sethi, 2016). Big national markets absorb this; small down-ballot markets may not. Source
Well-run forecaster panels can win
In a large forecasting tournament, teams of forecasters scored and weighted by past accuracy beat prediction markets (Atanasov et al., 2017). Source
Polls, read properly, beat the markets
Erikson & Wlezien, Public Opinion Quarterly (2008), with the historical follow-up in Electoral Studies (2012)
Discount each poll lead for the time left in the race and the polls win, because traders act as if campaigns can still move more than they do. Late in the race, prices add little beyond the polls.
How our blend answers it: We never use a raw poll lead. The poll average becomes a win chance with a 6-point allowance for a typical polling miss, and in races with markets polls typically carry about a third of the blend or more, which pulls back a market that overrates how much the race can still move. The allowance does not yet shrink as Election Day nears; that is the refinement this paper argues for, and our public scorecard rates the polls alone against the blend so the claim can be checked on our own record.
A scored panel can beat the market
Atanasov, Tetlock, Mellers, Ungar et al., Management Science 63(3) (2017)
On 261 geopolitical questions with more than 2,400 forecasters, market prices beat a simple average of forecasts, but team forecasts weighted by track record and recency, then recalibrated, beat the markets.
How our blend answers it: This is the case for our member panel, not against our method. Markets are a fast input, not the forecast. Panel results join the blend only after they have a published, scored track record, and their weight is set by that record.
1868–1940
Wall Street curb betting
New York
Open betting on presidential races, reported daily in the newspapers, that faded as scientific polling arrived.
1988–today
Iowa Electronic Markets
University of Iowa, USA
A real-money research market run by the Tippie College of Business. CFTC staff letters in 1992 and 1993 let it operate for academic purposes, with each trader limited to $500 per market.
1999–2013
Intrade (and TradeSports)
Ireland
The best-known election market of the 2000s. The CFTC sued it on November 26, 2012 for offering contracts to Americans; it closed U.S. accounts that December and halted all trading in March 2013.
2004
Australian betting agencies
Australia
Legal election betting, offered by several bookmakers, that Leigh and Wolfers found more stable and more useful than polls in Australia's 2004 federal election.
2014–today
PredictIt
Victoria University of Wellington, New Zealand
A university-run, small-stakes market under a 2014 CFTC no-action letter: $850 per contract and 5,000 traders per market. The CFTC withdrew the letter in August 2022; PredictIt won in court in July 2025 and became a regulated exchange.
2024–today
Kalshi and Polymarket
USA (CFTC-regulated)
Kalshi won the right to list U.S. election contracts in court in 2024; Polymarket returned to U.S. users through a CFTC-licensed exchange in 2025.
- 2004Favors markets
Historical Presidential Betting Markets
Rhode & Strumpf · Journal of Economic Perspectives 18(2): 127–142
Wall Street betting markets on presidential elections from 1868 to 1940 forecast results well, before scientific polling existed, and were fairly efficient despite manipulation attempts.
- 2003Favors markets
Results from a Dozen Years of Election Futures Markets Research
Berg, Forsythe, Nelson & Rietz · Handbook of Experimental Economics Results (Plott & Smith, eds.)
Across 49 Iowa Electronic Markets in 13 countries, election-eve prices missed the vote share by 1.49 points on average versus 1.91 for final polls, with traders capped at $500.
- 2004Favors markets
Prediction Markets
Wolfers & Zitzewitz · Journal of Economic Perspectives 18(2) / NBER w10504
Over four presidential elections the IEM's average error in the final week was about 1.5 points, against 2.1 points for the final Gallup poll. Known manipulation attempts had little lasting effect on prices.
- 2004Mixed / nuance
Prediction Markets: Does Money Matter?
Servan-Schreiber, Wolfers, Pennock & Galebach · Electronic Markets 14(3): 243–251
On the 2003 NFL season, a play-money market (NewsFutures) forecast as well as a real-money one (TradeSports). Both beat most individual experts.
- 2006Favors markets
Competing Approaches to Forecasting Elections
Leigh & Wolfers · Economic Record 82(258): 325–337 / NBER w12053
In Australia's 2004 federal election, polls varied too much across pollsters and over time to be very useful, while betting markets were stabler and accurate across the campaign and in individual electorates.
- 2008Favors markets
Prediction Market Accuracy in the Long Run
Berg, Nelson & Rietz · International Journal of Forecasting 24(2): 283–298
Compared with 964 polls across five presidential elections (1988–2004), the IEM was closer to the result 74% of the time, and significantly better in every election more than 100 days out.
- 2008Favors markets
The Promise of Prediction Markets
Arrow, Forsythe, Hanson, Shiller, Smith, Wolfers et al. · Science 320(5878): 877–878
Twenty-two economists, including Nobel laureates Kenneth Arrow, Thomas Schelling and Vernon Smith, called prediction markets a potent research tool and urged the CFTC to create a safe harbor for small-stakes markets.
- 2008Favors polls
Are Political Markets Really Superior to Polls as Election Predictors?
Erikson & Wlezien · Public Opinion Quarterly 72(2): 190–215
Comparing prices with raw same-day polls is unfair. Once poll leads are discounted for the time left, poll-based forecasts beat both IEM vote-share prices and winner-take-all prices.
- 2009Mixed / nuance
Combining Forecasts for U.S. Presidential Elections: The PollyVote
Graefe, Cuzán, Jones & Armstrong · Foresight
Blending polls, the IEM, models and experts cut error by a further 7–68% versus the best single component, and picked the wrong winner on only 4 of 957 days.
- 2009Favors markets
Forecasting Elections: Comparing Prediction Markets, Polls, and Their Biases
Rothschild · Public Opinion Quarterly 73(5): 895–916
Raw Intrade prices were slightly less accurate than a debiased poll model in 2008. After correcting both for known biases, market forecasts of state races gave more accurate win probabilities early in the campaign and in races that were not yet certain.
- 2010Mixed / nuance
Explaining the Favorite-Longshot Bias
Snowberg & Wolfers · Journal of Political Economy 118(4): 723–746
Bettors systematically overpay for longshots and underpay for favorites, and the cause is misjudged probabilities rather than a taste for risk. Low-priced contracts should be read with that in mind.
- 2012Favors polls
Markets vs. Polls as Election Predictors: An Historical Assessment
Erikson & Wlezien · Electoral Studies 31(3): 532–539
Before scientific polls (1880–1932), betting prices were good predictors. Once polls existed, last-week prices added nothing beyond the final polls, and in 1948 the markets followed the polls in making Dewey the favorite.
- 2014Mixed / nuance
Combining Forecasts: An Application to Elections
Graefe, Armstrong, Jones & Cuzán · International Journal of Forecasting 30(1): 43–54
Over six elections (1992–2012), averaging polls, the Iowa market, experts and models cut error by 16% to 59% against the individual methods' average error, and by 16% against the Iowa market alone.
- 2016Favors markets
Forecasting Elections
Vaughan Williams & Reade · Journal of Forecasting 35(4): 308–328
Comparing polls, betting markets, experts and a statistical model, markets gave the most precise forecasts and improved as the election neared, while polls did worse.
- 2017Mixed / nuance
Distilling the Wisdom of Crowds: Prediction Markets vs. Prediction Polls
Atanasov et al. · Management Science 63(3): 691–706
With 2,400+ forecasters on 261 geopolitical questions, markets beat a simple average of forecasters, but team forecasts with performance weighting and recalibration beat the markets.
- 2025Favors markets
Polymarket vs. polling averages in the 2024 U.S. presidential election
Cutting et al. · arXiv 2507.08921
Polymarket favored Trump from mid-October and in five of seven swing states while the polling average leaned Harris. The authors flag whale positions and an unknown trader base as caveats.
We do not pick a side. The PollyVote project found that blending polls, markets, models and experts cut forecast error by a further 7 to 68 percent compared with the best single source. Our Precision Combined forecast follows that lesson: the Precision Funnel™ weights Kalshi, Polymarket, PredictIt and the public polls by the evidence each carries in that race, and prints the weights and reasons on every race page.
The rule’s trust values (1 for Kalshi, Polymarket and polls, ½ for PredictIt) are a judgment call, kept simple and close to equal on purpose: a review of more than 200 studies found that simple averages are hard to beat (Clemen, 1989). They are not a finding. We will refit them on our own scored history once enough races have resolved, and publish the change.
Findings are summarized from each paper's published abstract or text; follow the links for the full studies. This page is research commentary, not betting or investment advice.