Kalshi called 479 of 487 midterm races and still got the hard ones wrong

Forty-seven Senate control contracts went unresolved on election night because the races were too close to call. That is not a failure — close races are supposed to be close. But the Wisconsin Democratic governor's primary is a different problem. Both major platforms priced Francesca Hong at 95% to win. David Crowley won by less than a percentage point. A 95% contract losing is not an outlier event that proves the system works at the margins. It is the system telling you something about itself.

The 479-out-of-487 figure is real, and it deserves scrutiny before it becomes marketing. The races that prediction markets price with high confidence tend to be the races that everyone already knows the answer to. What aggregates beautifully in liquid, well-covered contests is not the same thing as what aggregates in thin, regional primaries where trader composition matters enormously. The Wisconsin market was not thin on confidence. It was simply wrong at 95%, and the 95% is the number that reveals the model's actual weakness.

Here is what the platforms will not say plainly: one percent of traders hold three-quarters of all profits on these platforms. The wisdom-of-crowds argument requires the crowd. What prediction markets actually run, in many markets, is a small concentration of sophisticated capital pricing against a larger pool of wishful retail money. That is not aggregation. That is a spread.

The Vučić race in Serbia compounds this. A 70% probability assigned to a candidate who secured 39% of the vote and lost by a margin that made the race non-competitive is not a calibration error at the edges. It is a signal that trader composition on certain international markets skews toward regime familiarity — toward what people expect incumbents to do, not toward what electorates actually do. The same bias that inflated Vučić inflated Hong. Both were the known quantity in their respective contest. Both lost.

The study examining 2,500 political prediction markets and finding systematic inconsistency with rational information aggregation should be uncomfortable for anyone who cites these platforms as epistemically clean. It does not mean prediction markets are useless. It means they are asymmetrically good — excellent in liquid, high-information environments, unreliable precisely where the information gap is widest and where the platforms' influence claims are highest.

I have watched institutions anchor on a confident number before the underlying information justified it. The number becomes its own input. When a platform publishes 95% in a contested primary, traders who are uncertain defer to it. The confidence interval stops being a forecast and starts being a fact that the market then has to defend rather than update. This is the mechanism that the 479-out-of-487 headline obscures: it counts the settled contracts without accounting for how the unsettled confidence shaped behavior in the room.

The Arkansas suit against Kalshi landed the same week as this accuracy claim. The timing is incidental but the juxtaposition is not. Regulators are not primarily interested in whether prediction markets are good forecasters. They are interested in whether concentrated capital in thin markets creates pricing that moves behavior — in politics, in finance, in anything with a resolution date. The 1% concentration figure is the number that matters most to that question, and it is the number the platforms mention least.