Prediction Markets Trail Election Forecasters Before Midterms

Prediction markets struggled to outperform established election forecasting models during much of the 2026 US midterm campaign, according to a New York Times analysis. The platforms became more accurate as Election Day approached on November 3, when updated election forecasts and polling data began losing their predictive advantage.

The analysis found that prediction markets generally performed no better than conventional forecasting models throughout most of the three months preceding the election. In several periods, established forecasters delivered more accurate assessments of congressional races.

Researchers evaluated forecast performance using the Brier score, a statistical measurement that considers both accuracy and confidence. Predictions that express strong confidence in an incorrect outcome receive worse scores than predictions showing greater uncertainty.

The analysis compared six forecasting sources, including prediction markets and established election forecasting organizations. Silver Bulletin and FiftyPlusOne frequently ranked ahead of the prediction platforms between August and October.

South Carolina's Senate contest illustrated the difficulty of anticipating election outcomes. Darline Graham, who sought a full term following the death of her older brother, Senator Lindsey Graham, received President Trump's endorsement before the election.

However, Democratic challenger Annie Andrews won the race by approximately 90,000 votes, equivalent to 4.4 percentage points, according to results available Thursday. The outcome came amid a broader national shift toward Democratic candidates.

The New York Times reported that prediction markets had initially favored Graham, but assessments changed rapidly as election results emerged Tuesday evening.

One prediction-market company celebrated its performance in a social media statement on Wednesday.

"The platform was right," the company declared.

The statement followed the company's efforts to promote its forecasting performance throughout the election season. Its supporters argued that collective participation could produce reliable predictions about future events.

Tarek Mansour, chief executive of one prediction-market platform, also highlighted public participation in a Wednesday social media post.

"Over the last few months, the people showed up on our platform to tell the world what they were going to do," Mansour said. "Last night, the megaphone was passed to them."

The New York Times analysis nevertheless found that these platforms did not consistently outperform established forecasting organizations. Their strongest performance emerged closer to Election Day, rather than across the entire campaign period.

Kalshi Research head Nicole Kagan defended the platform's forecasting approach, arguing that its performance remained competitive even when participation was relatively limited.

"The claim that we're making, very simply put, is just that the election markets we have on our platform are well calibrated, up to say three months out, even with low participation," Kagan said.

In a statement, the company argued that producing competitive forecasts months before Election Day demonstrated its predictive capabilities. It also highlighted improvements as voting approached.

Election forecasters noted that conventional models depend heavily on polling, which can become outdated by Election Day. Prediction markets, by contrast, can respond more quickly when new information changes expectations.

The New York Times compared the rankings of six forecasting sources across 11 assessment dates between September 4 and November 3.

The rankings show that FiftyPlusOne led from September 10 through October 22, while Silver Bulletin maintained second place across much of that period. The two prediction-market platforms reached the top positions only near the end of the campaign.