Prediction markets may not be crystal balls, but a new Tel Aviv University study suggests they can offer a surprisingly disciplined way to think about the future. Researchers examined more than one million forecasts and found that crowd-generated probabilities frequently lined up with what actually happened.
The research team, led by Professor Michael Gilead and joined by Justin Dickerson, Hilla Shinitzky, Maayan Peer, and Mattan S. Ben-Shachar, analyzed 1,014,122 forecasts drawn from Polymarket, Kalshi, Manifold, and Metaculus. The dataset covered a decade, 12 subject areas, and prediction windows ranging from a few hours to several years. The resulting study, Calibration, accuracy, and domain variation across one million collective forecasts from prediction platforms, appears in the Proceedings of the National Academy of Sciences.
The central finding is calibration: when forecasters collectively assigned an event a 70 percent chance of occurring, events in that category happened at roughly that rate. That does not mean any individual 70-percent prediction is destined to come true. It means the probability becomes meaningful across a large population of forecasts, turning market odds into a statistical signal rather than a guarantee.
The study also shows why treating those odds as objective truth would be a mistake. Forecasts about military conflict, international relations, and technology tended to overstate the likelihood of events. Election-related questions showed a different distortion, with markets underestimating some outcomes when their predicted chances were low or moderate. In other words, the crowd appears capable of producing useful probabilities while still carrying predictable blind spots tied to the subject being judged.
Timing matters, too. Forecasts generally become more accurate as an event approaches, but the improvement was driven chiefly by better identification of which events were likely to happen—not by a dramatic correction in the underlying probability estimates. The researchers found that calibration was already visible after roughly a quarter of a forecast’s allotted time, when much of the relevant information was still unavailable.
That distinction gives prediction markets a more practical role in politics, business, security analysis, and technology planning. Their value may lie less in naming the future than in quantifying uncertainty early and revealing where a crowd’s confidence deserves scrutiny. The study’s preprint and supporting research point toward a future in which aggregated judgments become an important decision tool—but only when users remember that a probability is information about risk, not a promise about what comes next.
