Prediction Markets Are Not Polls, and the Midterms Proved It

A prediction-market price answers one question: how likely do traders think a single outcome is? It does not say how many people support a candidate, how big a win will be, or who is "ahead" in the sense a poll means it. Our position is that much of the public talk about these markets, including the platforms' own election hubs, blurs that line, and that the 2026 midterms are the clearest case yet of what the blurring costs. Reading a 93% price as a measure of support produces wrong conclusions in both directions: it makes accurate markets look like failures and makes their real errors harder to see.

What does a market price measure?

At 6:00 p.m. Eastern on November 3, 2026, before the first polls closed, Kalshi priced a Democratic House at 93.3% and a Democratic Senate at 75.0%, RotoWire reported. Polymarket had 94.5% and 75.5%. Each is the price of a contract that pays $1 if Democrats win the chamber and nothing otherwise. A 93.3-cent price means traders, on balance, put the chance of that outcome at about 93%.

Nothing in that number describes the size of the win. Democrats were on track to win the national House vote by about 13 points, against roughly 9 in pre-election polls, Straight Arrow News reported on November 5. A 93% House price fits a 3-point Democratic win, a 9-point one or a 13-point one. Any of those would have paid the contract in full.

Ashley Koning of the Rutgers Eagleton Center put the same idea in weather terms in SAN's piece: "The weather forecast tells us there's a 70% chance of rain. It doesn't mean that 70% of the area is going to rain."

How did the midterms get misread?

By our count, the midterms were misread in three ways.

The margin complaint. SAN's headline, "Why polls and markets got the midterm winners right but the margins wrong," groups two different instruments. Polls produce margins. The headline markets on House and Senate control produce a probability for one yes-or-no question, so they had no margin to get wrong. The markets did make a real error, and our midterm scorecard describes it: every known upset broke toward Democrats. A research bulletin cited by Reuters found that Democrats repeatedly beat their market odds, SAN reported. That is a mistake about correlation across races. Calling it a margin miss points readers at the wrong problem.

The upset complaint. Kalshi gave Texas Gov. Greg Abbott about a 60% chance on Election Day, and markets gave South Carolina Republican Darline Graham about 70%, according to SAN. Both lost. If you read 60% as "Abbott is up comfortably," his loss looks like a market failure. Read as a probability, it is an event the market said would happen about four times in ten. One loss tells you very little; whether 60% favorites lose about 40% of the time across many races tells you a lot. The Reuters Election Day preview made this distinction: "While calibration measures whether odds were reliable, accuracy measures whether the market picked the winner" (Reuters).

The forecast framing. Researchers at the University of Washington's Center for an Informed Public wrote on October 26 that platform election hubs "misleadingly imply that prediction market odds are equivalent to election forecasting," and that markets should not be read as "measures of political preference" (UW CIP). Their example: Polymarket's hub said Democrats will control the Senate without saying that the claim came from user bets.

Why does the confusion matter?

Because voters respond to market odds as if they were polls. The Brennan Center surveyed 1,994 registered voters on September 21, 2026, showing them pairs of hypothetical congressional candidates with randomly assigned NBC News poll numbers and Kalshi odds (Brennan Center). A candidate at 67% Kalshi odds was 10 points more likely to be picked as the likely winner than one at 50%; at 84%, the gain was 20 points. The effect was similar to that of the polls.

Two findings stand out. Only 19% of respondents called prediction markets accurate, against 37% for polls, yet the odds moved them anyway. And warning respondents about the markets had no effect on how much they relied on the odds. If people treat a price as a measure of public support even when they distrust it, the least the people publishing prices can do is say plainly what the number is.

What does this look like for 2028?

The gap between polls and markets is widest right now, a week after the midterms, with no declared candidates. A Financial Times/Focaldata survey fielded October 2 to 5 had Kamala Harris leading Democrats at 30%, Gavin Newsom at 15% and Georgia Sen. Jon Ossoff at 8% (Newsweek). On October 5, by Polymarket's price history, the nomination market had Harris at 6.6% and Ossoff at 15.4%.

Both numbers can be right. The poll measures whom Democrats like today. The market prices who will be nominated in 2028, which folds in whether each person runs, how primaries play out and what happens in the next two years. As of November 11, Polymarket has Ossoff at 26.7%, the highest in the field, though he said at an October debate that he has "zero interest in running for president" (Fox News). The market can rank him first because it is pricing the chance he changes his mind and wins. Our 2028 market feature has the full field.

What is the strongest case against us?

It goes like this. Markets are good at the thing most readers care about, which is who wins. A 2025 Vanderbilt study by Joshua Clinton of nearly 2,500 markets from the 2024 election found them "well calibrated overall" (Reuters). In 2026, Kalshi's favorites won 479 of 487 races resolved by the next morning, RotoWire reported, and both chambers went the way they were priced. Rutgers statistician Harry Crane told SAN that markets "don't just follow the polls," taking in registration and early-vote data as well. So when a casual reader sees 93% and thinks "Democrats are well ahead," the reader reaches the right practical conclusion. Insisting on the difference between a probability and a margin, on this view, is pedantry.

We take the point about winners. Our scorecard found that in 10 competitive Senate races, favorites won 9 against about 7.7 expected from their prices. Markets are good forecasts of who wins. That is the reason to read them as forecasts of who wins and nothing else.

The shortcut breaks exactly where it matters most. "Well ahead" invites readers to treat each race as its own sure thing, when the 2026 misses all broke one way. Crane's line about Abbott, that "It's not like [Abbott] lost his race independently of all the other races," describes a risk a single price hides. Clinton's own caveat was that broad measures "do not necessarily show how well the markets perform in close contests." And the 2028 markets show the shortcut failing outright: "Ossoff is ahead" is a false sentence about Democratic voters and a fair one about traders.

What would change our mind

We would drop this argument if we saw three things. First, a follow-up to the Brennan experiment showing that ordinary readers who see market odds correctly separate probability from support, for instance by estimating vote share no differently from people who never saw the odds. Second, platforms labeling election prices on their hubs as trader probabilities with the same consistency pollsters report margins of error, so the framing problem the UW researchers described goes away. Third, deep, well-traded markets on vote share and margins that sit next to the winner markets and beat the polls on those numbers over more than one cycle. If markets start pricing margins well and readers stop confusing the two kinds of number, the distinction we are drawing stops being useful. As of November 11, 2026, none of those three has happened.