Reason Foundation study finds polls and prediction markets roughly even in the 2026 midterms
A Reason Foundation analysis of 371 midterm polls finds Kalshi and Polymarket about level with pollsters, ahead only where the polls showed a tie.

Pollsters, Kalshi and Polymarket called the winners of the 2026 midterms at about the same rate, in an analysis the Reason Foundation published on Nov. 18, 2026. The commentary, by Reason Foundation managing director Jacob James Rich, sets the last poll each pollster released in a Senate or governor’s race against the Kalshi and Polymarket prices at the time that poll closed, then checks which of the three named the winner.
The study covers 371 polls from 97 pollsters, with fieldwork ending between Sept. 8 and Nov. 2, in races called as of Nov. 16.
The headline numbers
Setting aside 22 polls that showed a tie leaves 349. In those, the poll leader won 286 times (81.9%). The Kalshi favorite at the same hour won 279 times (79.9%), and the Polymarket favorite 281 times (80.5%).
A tied poll makes no call. In those 22 cases, the favorite on both Kalshi and Polymarket won 19 times (86.4%). Counting each tie as no call, the score across all 371 polls is 286 for the polls (77.1%), 298 for Kalshi (80.3%) and 300 for Polymarket (80.9%).
Rich calls the result “roughly even.” The markets’ lead over the full 371 comes from the tied polls; on the polls that made a call, the pollsters finished about two points higher.
How the comparison was built
A 23-page data appendix lists every poll with its margin, each market’s win probability for the Democratic candidate (the independent, Osborn, in Nebraska’s Senate race) and the final margin. Market prices are normally taken at 11 p.m. ET on the poll’s last field day, and a market favorite is whichever side stood above 50%. Reason’s data note says the prices were gathered with Claude, an AI model, from internet archives of the markets.
Ten Alaska Senate polls are left out because that race had not been called as of Nov. 16. The Associated Press called it for Democrat Mary Peltola late on Nov. 17, and Kalshi and Polymarket have since settled their Alaska Senate markets on her.
Where the markets missed
Five Senate and governor’s races in the comparison went against the Kalshi favorite as of 6 p.m. ET on Nov. 3: the Iowa, Mississippi and South Carolina Senate races and the Kansas and Texas governor’s races. Polymarket’s favorite missed the same races except Iowa, where it had the Democrat at 51% to Kalshi’s 49%. Democrat Josh Turek won that seat.
Rich uses Kalshi trading volume to test a common argument for prediction markets: that more trading brings more accurate prices. Three of the misses, the Texas governor’s race and the Iowa and South Carolina Senate races, ranked 4th, 6th and 10th of 31 races by Kalshi contracts traded before 6 p.m. ET on Nov. 3. Mississippi’s Senate race ranked 20th and Kansas’s governor’s race 26th. The Texas Senate race led all 31 at 95.4 million contracts. Rich reads the pattern as a sign that close races draw the most trading, and writes that more misses from future elections are needed to judge whether volume tracks accuracy.
Races without listed polls
The study’s larger point is breadth. Across 550 Kalshi races settled as of Nov. 16, covering House, Senate, governor, attorney general, lieutenant governor and secretary of state, the 6 p.m. Nov. 3 favorite won 523 (95.1%). In the 343 races where Wikipedia lists no general-election poll of the matchup, the favorite won 340 (99.1%). In the 207 polled races, it won 183 (88.4%). Rich notes that most seats are safe even in a blowout year.
That count goes further than Kalshi’s own post-election scorecard, which put its favorites at 479 of 487 races resolved by 6 a.m. ET on Nov. 4.
The limits Reason lists
The study scores only who won. It does not weigh margins or the strength of a forecast: a poll showing a candidate up 1 point counts as correct even if that candidate won by 7, and the study likewise did not factor in the strength of the market prices.
Opinion can also shift between a poll and Election Day, so a poll that missed was not necessarily flawed, while traders may price in events a poll cannot capture. Rich links the polls’ unusual accuracy to landslide results, which made the comparison hard to judge, and argues that a clear test needs an election in which the polls are badly wrong.
The same questions sit on Kalshi’s own list. Its research grant program names how market prices and polling averages moved against each other, and what markets showed in races with little or no polling, among the topics it wants studied, with proposals due by 11:59 p.m. ET on Nov. 18.
On winners alone, Reason’s count puts pollsters and the two markets within about two points of each other where both made a call, with Kalshi pricing hundreds of races that no poll on Wikipedia’s lists covered.
Our AI news desk researched and wrote this article, and software checked every factual sentence against the records it links to, following the WriterOps methodology. Report an error.