Kalshi Published Its 98% Thirty Hours After the Polls Closed. The Seven Answers Two Senators Asked for by Election Day Still Aren't Public. Only One of Those Numbers Tells You Anything.

Editor's note: TrueEdge Wire's owner builds odds tools, including exchange pricing, and earns affiliate commissions. We have argued against criminalising federally licensed exchanges. That is a position on regulation, and this piece touches regulation.

Kalshi's recap went up Wednesday, November 4, under the headline "Kalshi traders called 98% of races." The number underneath is 479 of 487 races that had resolved by 6 a.m. Eastern, in which "the candidates traders were favoring" won. The post does not say what price made someone a favorite, or at what time the snapshot was taken. It does say traders put $2.5 billion through the midterm markets, against $1.3 billion for the whole 2024 cycle.

A week earlier, on October 28, Senators Elizabeth Warren and John Curtis sent Kalshi and Polymarket seven questions about election-market integrity and asked for written answers by November 3. Among them: how many suspected violations each company stopped before a trade executed, versus caught afterward, and how many cases it referred to the CFTC or to police since January 1. We searched Saturday for those answers in any public form and found none.

One of those two sets of numbers tells you whether the markets work. It is not the one that got published.

What 98% measures

A hit rate on favorites counts a race as a success if the side priced above 50 cents won. It gives the same credit for a 97-cent favorite in a safe seat as for a 52-cent favorite in a toss-up. Most of the 487 races in the recap's table were the former. Of 360 House races counted, favorites won 356. Secretary of state races went 23 for 23. Nobody needed a market to call those.

The recap's own arithmetic shows the problem. Its table lists the House at "95%" and the Senate and governors at "94%," but 356 of 360 is 98.9 percent and 32 of 33 is 97 percent. The combined 479 of 487 does work out to 98.4 percent. Whatever produced the category figures, it was not the fractions beside them. We have no explanation and the post offers none.

The honest test for a market that sells probabilities is the one Thomas Kingsley gave Reason: an outcome priced at 55 cents should happen about 55 percent of the time. That is a calibration question, and it is answerable only with every race, resolved and unresolved, at a stated time. The recap does not attempt it. Neither did we on November 4, when we wrote that no race-by-race calibration existed yet. Three days later, it still does not exist in public.

The polls would also score 98%

Kalshi's claim is not only that favorites won. It is that "the markets beat the polls and the models." For that to mean anything, the polls need the same denominator.

Straight Arrow News reported Wednesday on a Reuters-cited research bulletin that found markets and polls "performed about equally" at picking winners, with Democrats repeatedly outperforming the market odds. That matches the Times' finding, as excerpted by Political Wire, that "for most of the midterm race, prediction markets performed no better than election forecasts." Polling averages favored Republicans in the Texas governor's race and the South Carolina Senate race, and Democrats won both. The markets favored Republicans in both too: Kalshi at about 60 percent for Greg Abbott, per SAN, and about 70 percent for Darlene Graham. Same misses, same direction.

The two reports we read even disagree about Texas. SAN has Kalshi at roughly 60 percent for Abbott on Election Day. Yahoo's Chris Williams has Kalshi at 62 percent for Gina Hinojosa. Both may be right at different hours, which is the point: a market's "call" depends on when you take the photograph, and a recap taken at 6 a.m. after the count tells you the closing price, not the forecast.

Where the edge really was, and why that is the problem

The Times' second sentence is the one worth keeping: "as Election Day neared, prediction markets performed better as forecasts and polling became stale." SAN shows the mechanism. Josh Turek's Iowa odds on Kalshi rose nearly eight points the day before the vote, after a final Times/Siena poll and early-vote data. Markets ingest new information in hours. Polling averages take days. In the last 72 hours of a campaign, that speed is a genuine advantage over any model.

It is also exactly the window in which inside information pays. A county clerk who has seen early returns, a campaign staffer with the final internal, a pollster's employee with a result before publication: all of them are most valuable in the final days, and all of them trade in the same stream that gives the market its late edge. That is what Warren and Curtis asked about. Their letter said that identifying isolated instances of improper trading "does not demonstrate that these markets are adequately protected," and cited Kalshi's own record of one candidate trading on his race in 2025 and three more in April 2026.

Kalshi can answer the hit-rate question in 30 hours because the data is a spreadsheet of closing prices. The integrity question needs surveillance logs and referral counts. The company has them. Reason reports Polymarket told NPR it had investigated "dozens" of midterm bets. A number like that, for both exchanges, with the before-and-after split the senators requested, is the only accuracy statistic that bears on whether these markets should exist in their current form. Warren is expected to chair the Senate Banking Committee in January. The questions will get asked again, under different rules.

The case for the 98%, taken seriously

The best argument for the recap is that it describes what customers actually buy. A trader does not need a Brier score. She needs to know whether the side the market liked tends to win, and 98 percent says yes. By the recap's own timeline, Kalshi had Democrats above 99 percent for House control by 8:10 p.m. Eastern, hours before the networks. They favored the Democrat in TX-15 in March, when Sabato's Crystal Ball rated it Likely Republican, and Bobby Pulido won by nearly 25 points. That is information a poll average did not have, and it was there for anyone to read.

All of that is true, and it argues for a different sentence than "beat the polls and the models." It argues for "we were as good as the forecasts for months and faster than them at the end," which is what the Times found and what a subscriber should want from a market. The 98 percent is marketing. The speed is the product. And the speed is what makes the unanswered letter matter, because the late edge and the late insider use the same door.

Here is what would change our reading. Kalshi or Polymarket publishing a full-slate calibration table, at a fixed pre-close timestamp, with every unresolved race included. Or either company publishing the seven answers. The first would tell us how good the markets are. The second would tell us how clean. Until then, 98 percent is a number about the races that were easy, published by the house, about the house.