Prediction markets missed Brazil. U.S. midterms are the big test.
Prediction markets spent three weeks heavily favoring Flávio Bolsonaro to be Brazil’s next president. That collapsed Sunday night. Tuesday, the same markets take on the U.S. Senate.
Kalshi and Polymarket put Democratic control of the Senate at 70% and 73% as of Friday morning. Forecaster Nate Silver’s model put it near 75% as of Oct. 29.
The Cook Political Report rates seven Senate races as toss-ups but said Friday it sees Democrats as “the slight favorites,” citing what Senate editor Jessica Taylor called a “dismal national political environment for Republicans.”
In September, the forecasters were further apart: a paper published Oct. 15 in “PS: Political Science & Politics” found the Iowa Electronic Markets and Kalshi pointing to opposite parties taking the chamber.
Kalshi is a federally regulated exchange, designated by the Commodity Futures Trading Commission, the agency that oversees futures and derivatives trading. Traders buy shares in an outcome. Each share pays a fixed $1 if the outcome happens and nothing if it doesn’t, so the price of a share reads as a probability.
Democratic control of the Senate was trading at 70 cents on Kalshi on Friday morning, which is meant to say the event is about 70% likely. Kalshi alone has taken more than $114 million in trading volume on the question.
The Iowa Electronic Markets works the same way at a fraction of the size. Run by University of Iowa faculty since 1988 as a teaching and research tool, it caps what any trader can put in at $500.
A price on Kalshi or Polymarket looks like a probability, but the researchers who study these markets say that is not quite what it is.
Thomas Gruca, the George Daly Professor in Marketing at the University of Iowa and director of the Iowa Electronic Markets, has been comparing his own market with Kalshi. In a July working paper with Thomas Rietz, a finance professor at Iowa, the two found that Kalshi contracts priced to win one time in five actually won less than 1.5% of the time.
Across 44 Iowa election markets, prices tracked outcomes closely regardless of how big or active the market was. Gruca said the data stops where the paper’s collection stopped, and he cannot say what Kalshi’s prices have done since.
Gruca said the difference is in what moves the price. Momentum, he said, means traders buying because a price has been rising, “not because there’s a change in the underlying fundamentals of what’s going to happen on election day.”
Looking at Kalshi’s daily price charts, he told The Center Square, “these things are kind of jumping all over the place. They’re vibrating all the time. There’s really not that much happening day to day.”
Thomas Rietz, Gruca’s co-author, said the Senate is hard for a reason that has nothing to do with how the markets are built. He compares it to a weather forecast. A 25% chance of rain in Fairfax, Washington, Baltimore and Annapolis on the same day does not mean it rains in one city and stays dry in the other three. A band of showers goes through and hits all of them or none.
Senate races work the same way, Rietz said. Most of the competitive ones are inside the margin of error, and the candidates and the voters resemble each other from state to state. If Republicans beat expectations in one place, they are likely to beat them across the board, or fall short across the board.
“That uncertainty hasn’t gone away,” he told The Center Square.
Others go further. Christopher Wlezien, the Hogg Professor of Government at the University of Texas at Austin, has argued for years that prediction markets add less than their defenders claim.
In a 2012 paper with Robert Erikson of Columbia, he compared election-eve betting prices with final polls across 30 presidential elections going back to 1880. The polls were the better predictor, and when the two were tested together, the market prices added nothing the polls had not already captured.
Wlezien said he has not updated that analysis and cannot say whether poll-based predictions still beat market prices today.
“I’m not anti-market,” Wlezien told The Center Square. “I’m just pro data.”
Thomas Miller, faculty director of the data science master’s program at Northwestern University, agrees that a market price is not a probability and draws the opposite conclusion from it.
His NowCasting model takes prices from Kalshi and Polymarket, adjusts them so the contracts in each race sum to 100%, and runs them through a statistical model. On Oct. 27, it put Republican chances of holding the Senate at 23.4%, against about 36% on Kalshi the same day.
“Prediction markets provide event contract prices, not probabilities,” Miller said in written answers. “Traders are guessing. We are not.”
Kalshi’s own post-election analysis says the market had Flávio Bolsonaro near 80% when polls closed Oct. 25, and that the final Brazilian polls leaned the same way. Lula won 51.79% to 48.21%. By Kalshi’s accounting, the market missed the margin by about 7 points and the polls missed by about 5.6.
Kalshi’s published analysis offers a theory. Three weeks earlier, in the first round, polls had underestimated Bolsonaro. Traders adjusted for that going into the runoff, pushing the market further toward him than the polls were.
In the runoff, the polls were wrong in the other direction, and Lula won. The company noted that an 80% chance should be wrong about one time in five, and that Bolsonaro’s loss was one of those times.
In Wisconsin’s Democratic primary for governor on Aug. 11, Kalshi had state Rep. Francesca Hong at 95.49% on the day of the vote, where she had sat for most of the preceding week. Milwaukee County Executive David Crowley won by 3,673 votes out of 792,290 cast, less than half a percentage point, according to official results certified by the Wisconsin Elections Commission. Traders moved more than $16 million on the race.
Kalshi did not respond to questions from The Center Square about the Brazil pricing or the Wisconsin primary.
Wlezien said that is the part he keeps coming back to. A market price is supposed to be a probability already. If someone has to adjust it before it means anything, something has gone wrong with the number itself.
“You can just go look online and say, hey, 65,” he said. “That means like a two out of three chance, right?”
Cloudy
