Key takeaways
- A poll estimates a distribution of opinion. A market prices a single event once. Both output a percentage, and the two are not interchangeable.
- The clearest win for markets came in June 2024, when the Biden contract repriced within half an hour of the debate while the major forecasting models had to be suspended entirely.
- Dr. Rajiv Sethi calls what those models produced in the following month zombie forecasts, because they were pricing a race that was not going to happen.
- The mechanism's edge sits in events without precedent, where statistical models have no history to calibrate against, rather than in routine polling questions.
- Andrew Gelman of Columbia allows that market prices can be converted into a forecast while arguing prediction markets are unlikely to replace polling, because they answer a different question.
- A 2026 study built a Signal Credibility Index from market microstructure and found the most socially visible markets produced the least accurate forecasts.
Why this comparison keeps being framed wrong
Since 2024, prediction market prices have been quoted in news segments and embedded next to stock quotes on Google Finance, usually with the implication that they have beaten the pollsters. The record is more specific than that, and more interesting. This piece sets out what each instrument actually measures, where markets demonstrably outperformed, and where the comparison breaks down, drawing on the forecasting chapter of the Hodl Up Prediction Markets Report 2026 and our interview with economist Dr. Rajiv Sethi.
The two instruments answer different questions
A poll samples a population and estimates how opinion is distributed, with a margin of error attached. A prediction market prices one binary outcome, resolves once, and produces a number that reflects the balance of capital rather than the balance of opinion. Sethi's caution on interpretation is the one most coverage skips, because the market does not hold a belief and traders disagree with each other. What makes the number behave like a probability is arbitrage, not consensus, since tight spreads and returns on collateral let traders pull a set of linked contracts back into line when they start to violate the laws of probability.
| Opinion poll | Prediction market | |
|---|---|---|
| What it estimates | Distribution of opinion in a population | Price of one binary outcome |
| Update frequency | Per fielded wave | Continuous |
| Who supplies the input | A sampled respondent, unpaid | A trader with capital at risk |
| Main failure mode | Sampling and response bias | Thin liquidity, concentration, reflexivity |
| Handles an unprecedented event | Poorly, since models need history | This is where its edge sits |
Andrew Gelman, talking to Undark, allowed that market prices can be turned into a forecast while making clear that prediction markets are unlikely to replace polling. A single market resolves one event once. A poll estimates a distribution of opinion. Both produce a percentage, and that is where the resemblance ends.
The June 2024 debate is the strongest case markets have
Within the first half hour of the debate, the price of the Biden contract on Polymarket fell sharply. The Trump contract did not rise by a corresponding amount, because traders were simultaneously pricing the possibility that Biden would be replaced. The price of Newsom rose, then Harris. That was a month before Biden withdrew, while he and his surrogates were insisting he would not.
The three models Sethi was tracking at the time, the Economist's, 538 and Nate Silver's Silver Bulletin, all had to be suspended when Biden dropped out, because they were built on Biden-versus-Trump polling data.
"They were generating essentially zombie forecasts for the month between the debate and the decision of Biden to withdraw. They were forecasting a race that was not going to happen."
Sethi's generalisation from that case is more useful than the case itself. The mechanism's comparative advantage lies in uncharted waters, in events with no precedent that models cannot calibrate against. He lists an unprecedented pandemic, a major-party candidate withdrawing and an assassination attempt on a nominee as the class of events where, in his words, the models simply cannot handle it. For a non-political comparison, see the recent study on prediction markets and weather forecasts.
Where markets carry a risk polls do not
Elections are reflexive events, which means belief about the outcome affects the outcome. Campaigns already understand this, which is why they suppress internal polls that look bad and publicise the ones that look good. A campaign thought to be flagging can see fundraising dry up, volunteer effort fall off, and turnout decline once early voting has begun.
The difference is credibility. Sethi put it directly to us.
"When you're talking on media, people are skeptical. They know that if a representative of a campaign is saying you should ignore a poll, they're saying it because that's in their interests, and we can discount that. But with markets, there's a temptation to see it as somehow objective and truthful."
Rothschild and Malhotra asked in 2014 whether published polls function as self-fulfilling prophecies, and the mechanism carries straight over. A market is the stronger version, because it carries the authority of real money and updates in real time.
The 2026 Signal Credibility Index work sharpens the warning. Applying a microstructure-based diagnostic to three political shocks from 2024, the authors found that the most socially visible market produced the least accurate forecast, which they describe as a systematic decoupling of social authority from epistemic quality. The market everyone was watching and quoting was the one carrying the worst information. For how the headline numbers behind those markets are measured, see Nobody Knows How Big Prediction Markets Are, and for the case for using markets to make decisions rather than just forecast, our conversation with Robin Hanson on futarchy.
The full forecasting chapter, including the Federal Reserve evidence and the calibration literature, is in the Prediction Markets Report 2026. Sethi's complete interview is here, and the report summary covers the headline findings.
None of this is investment advice.