Executive Summary 

• Prediction markets can complement traditional polling by aggregating dispersed expectations and incorporating new information, yet while evidence from the 2024 campaign and the 2026 primaries suggests they are generally well calibrated, their reliability depends on the information available to traders, market liquidity, and the incentives participants face. 

• The regulatory framework governing prediction markets is still being settled, and rules on who may trade, in what, and on what information bear directly on whether prices can be trusted, as the Commodity Futures Trading Commission’s June 2026 proposed rule and its first insider trading cases indicate. 

• Analysts should therefore evaluate markets and polls on accuracy, calibration, and incremental contribution to a forecast rather than assume either is inherently superior, and although markets appear to have earned a place alongside polls, what they add beyond polling remains an open empirical question. 

Introduction 

Prediction markets have become an increasingly prominent feature of political forecasting. By allowing participants to trade contracts tied to future events, they produce prices that reflect collective expectations, and, unlike a conventional poll, which is necessarily tied to a snapshot in time, a prediction market price can change continuously as new information arrives – in the form of traders’ responses to prospective financial gains or losses. A recent American Action Forum (AAF) primer describes how these contracts work and why they attract both enthusiasm and suspicion; this paper addresses the narrower question of how they perform in forecasting elections. 

Measuring public opinion and predicting an election are different tasks. Polling estimates what voters say they intend to do, whereas markets aggregate expectations about the eventual outcome, and neither observes the future directly. The relevant question is therefore whether markets improve on forecasts built from polling and other available information, not whether they can replace polling outright. Three examples show why a price is not an authoritative forecast: in 2016, errors in state polls obscured the presidential contest even as national polls correctly indicated a Clinton popular-vote win; in 2024, the Iowa Electronic Markets (IEM), a small academic market, gave the Democratic candidate an 85.7 percent probability of winning the popular vote on September 29, yet Donald Trump won it; and in May 2025, leading markets priced Cardinal Robert Francis Prevost at between one and two percent before he was elected Pope Leo XIV. None of these results shows that markets are ineffective, but that their limitations need to be understood. 

The Case for Prediction Markets 

The principal advantage of prediction markets is their ability to aggregate dispersed information. Pollsters must recruit respondents, construct representative samples, estimate turnout, and translate responses into an assessment of voting behavior, and each step introduces error. Prediction markets by contrast, let participants fold polling, campaign reporting, and economic data into a common price. The June 2024 presidential debate shows how a market responds to new information: According to Thomas Gruca and Thomas Rietz’s analysis of the IEM, the probability of a Democratic popular-vote victory fell from its 70–80 percent range to 59 percent after the June 27 debate and did not surpass its pre-debate level until July 31, although not every price change necessarily reflected new or accurate information. 

Markets also express uncertainty in a form that can be evaluated, and three standards matter most in doing so. Outcome accuracy asks whether a forecast identifies the winner; calibration asks whether events assigned a given probability occur at roughly that frequency across enough forecasts; and incremental predictive value asks whether market information improves forecasts that already incorporate polling. A market can be accurate without being well calibrated, or without telling analysts anything they did not already know. Financial incentives do not resolve these problems automatically, since correcting mispricing depends on liquidity, informed participation, and traders’ capacity to act on what they know, and prices can reflect shared assumptions and media narratives (as in financial markets generally) rather than independent evidence about voters. 

2016: Measuring the Popular Vote Is Not Calling the Presidency 

The 2016 election shows that the outcome being forecast matters as much as the method used. Hillary Clinton won the national popular vote by approximately two percentage points, while Donald Trump won the Electoral College after narrowly carrying Michigan, Pennsylvania, and Wisconsin. The American Association for Public Opinion Research’s post-election evaluation, An Evaluation of 2016 Election Polls in the United States, found that national polls performed well overall, while state-level polls, particularly in the Upper Midwest, underestimated Trump’s support, citing late movement toward Trump and the overrepresentation of college graduates in some samples. 

These findings distinguish polling error from forecasting error: Polls estimate voter preferences, whereas a forecast must translate them into the probability of an institutional outcome, and a candidate can lead nationally while remaining vulnerable in the states that decide the result. The distinction also limits what a market’s record can show. The IEM’s popular-vote market gave Clinton a 72.7 percent chance on September 29, 2016, according to Gruca and Rietz, and she won it, but the market was never asked about the Electoral College, and an accurate call on one does not establish an accurate forecast of the other. Analysts must therefore compare forecasts of the same outcome and separate errors in polling inputs from errors in the models that convert them into probabilities. 

2024: Market Prices, Liquidity, and Calibration 

The 2024 election offers a test of the IEM’s popular-vote forecasts, though a narrow one. On September 29, 2024, its winner-takes-all market assigned the Democratic candidate an 85.7 percent probability of winning the popular vote, and a distribution combining it with the vote-share market implied 87 percent and a margin of six to seven points, as Gruca and Rietz report; Trump defeated Kamala Harris by approximately 1.5 points. Two caveats apply. The forecast was a snapshot taken five weeks before Election Day, and the IEM is a small research market, with the winner-takes-all market having traded roughly 42,000 contracts worth about $21,000 in total. It is therefore not a close proxy for the commercial exchanges where most of the money now changes hands, and CNN observed that the two most significant prediction markets, Kalshi and Polymarket, outperformed traditional polls in calling Trump’s win. 

The miss does not establish that markets systematically underperform polling. An 87 percent probability allows for an outcome that occurs 13 percent of the time in a well-calibrated system, and calibration can only be judged across a larger set of forecasts. The IEM study did identify a problem specific to thin markets, in that unusual bids in the lightly traded vote-share market temporarily distorted displayed prices and stale last-trade prices persisted after bids and asks recovered. Depth appears to help: according to Kalshi, a single $1.25 million purchase in the July 2026 Los Angeles mayoral race moved one candidate’s price from $0.29 to $0.45 before it reverted within nine seconds, although that is one episode reported by the exchange itself, and liquidity cannot ensure that the information being traded on is correct. 

Broader evidence from the 2024 campaign is encouraging but must be read with care. In a June 2026 working paper, Joshua D. Clinton and TzuFeng Huang examined nearly 2,200 political markets across the IEM, Kalshi, PredictIt, and Polymarket during the final nine weeks of the campaign, and reported that 86 percent performed better than a coin flip and that markets were generally well calibrated, with modest longshot bias. Of the 66 percent of contracts priced below $0.20 or above $0.80 on the eve of the election, 96 percent resolved in favor of the heavy favorite. A contract such as “Will a Democrat win Wyoming’s U.S. Senate seat?,” trading at about $0.02 on election eve, is typical of this group: The market implied a 2 percent chance, and the contract resolved “No” as expected. Evidence on efficiency was more mixed, and an earlier version covering about 2,500 markets over five weeks found accuracy of 93 percent on PredictIt but only 78 percent on Kalshi and 67 percent on Polymarket, so the conclusions appear sensitive to sample and definitions. A high success rate can also reflect the mix of contracts traded, and beating a coin flip says little about whether a market adds information to an existing forecast. 

Evidence from the 2026 primaries points in the same direction, with the same qualifications. A Washington Post analysis found that primary candidates won about as often as bettors expected, including in Georgia and Kentucky, where favorites lost. A Kalshi study of roughly 2.24 million resolved markets found that calibration improved with volume and the number of traders. Preliminary NBER research found no systematic partisan or demographic bias, alongside the familiar favorite-longshot bias, and analysis by Slow Boring reached a similar view on calibration while noting that thinly traded contracts can miss badly, as in Michigan’s 8th and Colorado’s 3rd districts, where volume was roughly $5,400 and $72,000. The practical lesson is to evaluate probabilities rather than count correct calls, and to test whether prices improve on polling-based forecasts, particularly in close contests. 

The Papal Conclave: Forecasting Without Public Information 

The election of Pope Leo XIV in May 2025 illustrates a different limitation: Markets cannot aggregate information that their participants do not have. Before the conclave, Cardinal Pietro Parolin had been the favorite in the prediction markets but, according to Axios, never sustainably exceeded one in three, trading at about 27 to 28 percent on the morning of the decisive vote, while Cardinal Luis Antonio Tagle, the second favorite, was only a few percentage points behind, and CNN reported Prevost at between one and two percent on Kalshi and Polymarket. When white smoke appeared, Parolin’s price rose to 68 percent on Kalshi while Prevost remained at 1.2 percent, as traders seemed to read the rapid conclusion as a sign of consensus around Parolin, an inference that proved incorrect when Prevost was elected on the conclave’s second day. That sequence suggests that the information reached the market only after it had left the room. 

A conclave is a distinctive forecasting problem because a small group decides (133 cardinal electors in 2025), preferences are not publicly observable, and deliberations are secret, with every elector swearing an oath of secrecy before voting, so the difficulty is not simply thin liquidity or a mispriced contract: If decisive information is unavailable to the public, trading cannot create it. Dartmouth economist Eric Zitzewitz made a related point to CNN, observing that conclave markets are among those least likely to be well calibrated because they receive a new data point only every decade or two. Elections offer more public information, since polling provides repeated measurements and administration eventually produces a countable result, but uncertainty about turnout, late-deciding voters, and the geographic distribution of support persists until the votes are counted. Markets can synthesize available evidence, but their prices should not be mistaken for evidence that does not exist. 

Regulation and Market Integrity 

What prices can tell us is inseparable from who is allowed to trade, in what, and on whose information. As the AAF primer explained, the American debate has been less about forecasting than about whether these contracts are financial products or a form of gambling, and the Commodity Futures Trading Commission (CFTC) oversees them as derivatives. The IEM sits largely outside that framework, operating under CFTC no-action letters, filing none of the reports ordinarily required of regulated venues, and not being formally regulated by the CFTC, which suits a research market but means that the best-known academic evidence comes from a venue unlike the commercial exchanges. 

The commercial exchanges face a more active environment. The CFTC published an advance notice of proposed rulemaking in March 2026 and followed it on June 10 with a notice of proposed rulemaking that would amend existing rules deciding when event contracts are contrary to the public interest; comments closed on July 27. The agency has also, together with the Department of Justice, sued New York to defend its exclusive jurisdiction. Enforcement now addresses the information problem directly: On April 23, 2026, the CFTC filed what it described as its first insider trading complaint involving event contracts, against an active-duty Army service member who allegedly traded Polymarket contracts using classified information. Moreover, Kalshi disciplined three political candidates who wagered on their own races the previous day and a second complaint followed on May 27 against a Google software engineer who allegedly traded on confidential information about the company’s year-end search rankings. Such actions should, over time, strengthen the credibility of the markets. 

A tension remains, however. Markets struggle to forecast decisions that turn on private information, and the traders best placed to price such outcomes are those whose access regulators increasingly treat as a violation; insider trading rules and thin payoffs also weaken the incentive to fund the private polling that would give markets new information. Two practical implications follow. Analysts who cite prices should report volume and liquidity and specify the contract quoted, since a probability drawn from a few thousand dollars of trading is not the same object as one drawn from tens of millions. Further, regulators should preserve independent evaluation, since timestamped forecast records comparable across venues are the raw material for the calibration studies recommended below. 

Markets and Polls: What Does Each Add? 

Markets and polling should be treated as complementary sources of information, with their relative value assessed empirically. Polling remains essential for measuring voter preferences and the composition of electoral coalitions, while markets provide a continuously updated view of how participants interpret that information alongside campaign developments, and neither should automatically override the other. The first requirement is to compare like with like, since national popular-vote polling cannot be compared with an Electoral College market and vote share is not a probability of victory. 

Disagreement between the two can be informative, but it does not by itself show that the market knows more, because a market moving ahead of polls may be anticipating a real shift, incorporating information not yet in surveys, or reacting to a narrative that ultimately proves wrong, and the appropriate step is to investigate the divergence before choosing a signal. The early evidence suggests the added value may be modest, as Slow Boring’s review of the 2026 primaries found that markets contributed little beyond polling, but that finding is preliminary and should be tested by recording forecasts before Election Day and comparing calibration, error, and performance in close contests across many races, including whether combining markets with polling beats either alone. Without such tests, advocates risk treating successful predictions as proof of superior information, while critics risk the reverse. 

Conclusion 

Only 25 percent of Americans trust political polling. Similarly, prediction markets can improve election forecasting, but they are not oracles, and the lessons of 2016, 2024, and the Vatican converge: National totals are not state outcomes, thin markets can sound more certain than the evidence warrants, and no price can reflect what its traders cannot know. The evidence on calibration is nonetheless encouraging, since markets have so far been reasonably candid about how often they are likely to be wrong. The harder question is whether they tell analysts anything that polling has not already said, and answering it requires comparable forecast records together with a regulatory framework that polices the trading without foreclosing the evidence. The CFTC’s proposed rule and enforcement actions speak to the second requirement but not the first. Until the data accumulate, claims that markets have made pollsters obsolete remain premature; markets have earned a seat at the table, but not yet the head of it.