Developing Prediction Markets Called the Midterms First and Called the Size Wrong. The Big Misses All Pointed the Same Way.

Editor's note: TrueEdge builds odds tools that include prediction-market prices and earns affiliate commissions from licensed sportsbooks. Our standing position is that states should not criminalise federally registered exchanges. This piece argues about how much credit those exchanges earned on Tuesday. Weigh the conflict.

Kalshi priced Darline Graham at 67% to win South Carolina's Senate seat at noon on November 2, and both major exchanges still had the Republican near 70% late on Election Day, according to Joe Wager's reading. Annie Andrews won. In Texas, markets had Greg Abbott around 60% to 65% for another term. Gina Hinojosa beat him.

The exchanges' defenders have the correct first response. A 70% favorite is supposed to lose three times in ten. Barnard economist Rajiv Sethi gave CNBC the clean version: "If you toss two coins, there's a 25% chance that you'll throw two heads in a row, and nobody would be terribly surprised if you did."

That analogy is where we part company, because elections are not independent coins. Tuesday's misses were one correlated error about the size of a national wave. The record so far cannot tell anyone that these prices are well calibrated, however many victory laps are being posted on X.

Right about who, light on how many

Start with what the markets got right, which is the thing that mattered most. Kalshi had Democrats at 94% for the House and between 74% and 77% for the Senate, depending on the hour. Democrats won both. On Kalshi's November 2 sheet, Josh Turek was 52% in Iowa, Adam Hamilton 63% in Kansas, Sherrod Brown 71% in Ohio and Troy Jackson 73% in Maine, and all four won. That is a real performance by any fair standard.

Now the magnitude. CNBC reports that before polls closed, Democrats were narrow favorites to reach 53 Senate seats, with a 54th a toss-up. As of Wednesday evening they had 54, with Mississippi and Alaska outstanding and priced at 90% for Scott Colom and Mary Peltola. If those hold, the count reaches 56, two seats beyond what the market treated as a coin flip. Dustin Gouker of The Event Horizon notes that Kalshi's "megatsunami" contract, requiring 240 House seats and 54 in the Senate, was not favored, and that Kalshi's forecast margins in the Texas Senate race were "blown out of the water." He summarizes it as a market that got the direction right and may have missed the breadth and intensity.

Read the upsets in that light. South Carolina, where Trump won by about 18 points in 2024 per Joe Wager, and Texas governor were not two unlucky draws. They are the same bet: how far would the environment reach into red states? Iowa at 52%, which CNBC describes as 50-50, belongs in that group too. Each one moved in the same direction. When outcomes are this closely linked, a sheet with thirty races contains far fewer than thirty independent tests, and the error shows up everywhere at once.

We do not know how bad the correlated error was in percentage terms. That needs closing prices from every race, taken before any votes were counted, and no one has published that work yet. Joe Wager is right that the South Carolina number was a late-afternoon price, not a formal close. Our claim is narrow: the misses in hand point one way, and that deserves the first look from anyone who runs the calibration.

Thirty-one minutes is a different product

What the exchanges are selling this week is speed. Polymarket said that in 10 key races it called winners ahead of CNN, Fox News and the Associated Press by a median 31 minutes. "It's now routinely hours," CEO Shayne Coplan posted. Kalshi spokesperson Jack Such told CNBC the exchange makes a call when a candidate tops 99% and trading against that candidate has "dried out." Andrews became the Kalshi favorite shortly after 7:30 p.m., 30 minutes after South Carolina's polls closed, and Hinojosa jumped from 39% to 73% on Polymarket within five minutes of initial closings.

Those numbers are impressive, and Polymarket's is self-reported over a sample it chose. But they measure how fast traders digest early returns, and that is a separate skill from forecasting. Networks wait because a call is a public statement with consequences, and a 99% market price made by people with money at stake is not the same object. When an exchange posts "projected winner" on X, it borrows the authority of the second thing using the mechanism of the first. If a market ever flips after a confident post, the exchange will have to answer for that call while having earned fees on the trading around it.

Also keep the scale in view. Gouker cites Bloomberg saying Kalshi's final-week turnover on US election markets was roughly what it was before the 2024 vote, and that Polymarket's combined election trading shrank significantly. That sits awkwardly beside the "first true prediction markets election" line from Rutgers statistician Harry Crane. It also complicates a comparison we reported on Tuesday, drawn from Crypto Briefing, that Kalshi's midterm total beat its 2024 presidential number; a cumulative book and a final week are different measures, and both can be accurate.

And note who has a stake in the story. CNBC, to its credit, discloses at the bottom of its own piece that it has a commercial relationship with Kalshi including a minority investment.

The strongest case against all this

A skeptic should say we are grading on hindsight. Covers notes that the information available before Election Day did not fully anticipate South Carolina, so the market was hardly alone. It got the question that determined power in Washington right, with prices anyone could check. Demanding precision about a 56th Senate seat is moving the goalposts after a winning night. Worse, correlated error is not a flaw unique to markets. Every forecaster who relies on national environment shares it.

All true, and it is why we are not calling this a failure. The market's top-line job was to say which party would take Congress, and it did. Our objection is to the inference being sold on top of it: that a sheet of prices is a trustworthy probability for each individual race, ready for news desks to print as fact. Dartmouth's Eric Zitzewitz told CNBC that first-time users see 55-45 and think the 55 is "way ahead." The Anti-Corruption Data Collective's Michelle Kendler-Kretsch said, "Risks and vulnerabilities didn't change overnight just because the markets were accurate this time around." We may not share her regulatory conclusions, but that sentence is simply correct. CNBC's own piece lists the recent reminders: Lula at about 20% on Kalshi before Brazil's polls closed, and high-profile Democratic primaries in Wisconsin and Michigan over the summer where it says markets whiffed.

Thin markets deserve a further discount. Nebraska's two betting initiatives traded at 69% and 68% to pass on a few thousand dollars and both lost. That number appeared in a national outlet with the appropriate caveat attached. Many others will not get the caveat.

What would change our mind is a published calibration across every state race from Tuesday, using pre-poll-closing prices and a real accounting for correlation. If the 70s won about seven in ten after that adjustment, we will say so here, plainly. Until then, "the markets called it" is true of the night and unproven for the price sheet. Anyone trading these contracts next cycle should size positions as if a wave can move many of them together, because this one did.

Gambling and event-contract trading involve risk. If you or someone you know has a gambling problem, call 1-800-GAMBLER.