Democratic Socialist Zohran Mamdani upset the New York City Democratic primary on Tuesday when he beat incumbent Andrew Cuomo. The two men had been locked in a dead heat in polls, and had traded places as favorites to win on prediction markets.
Prediction markets offered a fair overview of the race, based on information available. Commercial prediction market platform Kalshi went into the evening with Cuomo at about 60% and Mamdani at about 40%. Crypto prediction market exchange Polymarket showed an 80/20 split between Cuomo and Mamdani with Cuomo as the clear favorite.
While Kalshi has touted itself as the first to call the election, they weren’t necessarily a more accurate predictor compared to polls. But speed wasn’t the initial promise of prediction markets.
Prediction market platforms have touted themselves as superior alternatives to traditional polls since outperforming them in the 2024 presidential election. However, the reality of prediction markets’ role in data gathering is turning out to be more nuanced than marketing slogans often suggest.
Weak in small races, strong in large ones
Prediction markets showed Donald Trump further ahead in the 2024 presidential election than traditional polls, which largely showed a toss-up between Trump and Kamala Harris. But that triumph at the national level obscured mixed performances at the level of House districts.
Columbia economist Rajiv Sethi compared Polymarket’s performance in predicting 13 House races with polls from FiveThirtyEight and The Economist. In those races, Polymarket underperformed FiveThirtyEight and, in some cases, The Economist as well. Sethi concluded that while prediction markets may have a role to play in forecasting, they’re not “a magic bullet” for divining the future.
Kalshi CEO Tarek Mansour tweeted on Tuesday night: “Kalshi HQ official called the election first: Zohran Mamdani. Prediction markets are the news.”
And while they may have “officially” called it ahead of media outlets, the Kalshi market was not prophetic. But there is a good explanation as to why.
NYC Mayoral primary proves a point
What we saw in the New York City mayoral primary is consistent with Sethi’s findings from his selection of 2024 House races. The mayor’s race is a smaller environment where pollsters could isolate the necessary variables to forecast the winner.
Emerson College’s final poll projected Mamdani’s victory over Cuomo by four points the day before the primary. Kalshi’s market followed that poll with Mamdani pulling ahead of Cuomo for the first time Monday morning and then again late Monday evening and into election day. However, in the final hours of voting, Cuomo had regained his lead as the favorite on Kalshi where he remained until polls closed. It was only at that point that Mamdani began soaring in odds.
BREAKING: ZOHRAN’S ODDS GO STRAIGHT VERTICAL AS POLLS CLOSE pic.twitter.com/DBP00gVVQk
— Kalshi (@Kalshi) June 25, 2025
So, prediction markets could underperform compared to polls in both small mayoral races and in House races, making the midterms a potential comeback moment for pollsters.
But prediction markets could potentially outperform polls again in the 2028 presidential election by capturing the popularity of an idea or a candidate in real time instead of relying solely on past data. The real-time reflection of public sentiment and data aggregation can capture changes in voter sentiment much quicker than polls can.
Assessment of prediction markets’ performance then largely depends on the expected utility and role that these aggregation tools play.
What should we expect from prediction markets?
Prediction market platforms have fallen short of the “truth machines” that Kalshi marketing has claimed them to be. But prediction markets haven’t shown themselves to be completely useless, either.
Polls may attempt to show what a likely vote breakdown would be given a snapshot of voter data, but prediction markets show the odds that a candidate will win. A truth machine should at least show the correct outcome as an early betting favorite.
However, a market aggregation should show the state of public information about a race, and Mamdani winning four out of 10 times was a reasonable estimate given how closely he had been polling with Cuomo. Polymarket’s estimate of Mamdani winning two out of 10 races seems much further off, but the cutoff between reasonable and unreasonable estimates will vary among readers.
The New York City mayoral primary was a fair night for Kalshi. The platform showed a reasonable probability breakdown and called the election quickly, but didn’t offer unique insights about the race’s winner before the results came in. The real-time snapshots of the race has its value, but it wasn’t fully up to the platform’s promises, either.
