How accurate were the 2026 midterm markets? A race-by-race scorecard for Kalshi and Polymarket

Two days after the midterms, enough races have settled to grade the prediction markets. We took the price of every settled Senate and governor's race, and the two chamber-control markets, at 6 p.m. ET on Election Day, just before the first polls closed, and compared them with the results. The headline: both Kalshi and Polymarket were well calibrated overall, Polymarket slightly better, and both underestimated the Democratic wave in the same way.

The race-by-race data are in a CSV file.

Key findings

1. The favourites mostly won. Kalshi's favourite won 62 of the 65 markets and Polymarket's 63.
2. Polymarket was slightly more accurate. Its Brier score was 0.0318 against Kalshi's 0.0330; its log loss was 0.126 against 0.130. The gap is small.
3. Every miss went the same way. The three races where Kalshi's favourite lost, South Carolina's and Iowa's Senate races and the Texas governor's race, were all won by Democrats.
4. The markets underpriced Democrats in close races. In the 18 races where Kalshi priced the Democrat between 10% and 90%, the average Democratic price was 61.7% and Democrats won 15 of them, or 83%.

What we measured

• Markets. 31 Senate races, 32 governor's races and the Senate and House control markets that had settled on Kalshi by 5 November and were also listed on Polymarket. Mississippi's and Alaska's Senate races, still uncalled, are excluded.
• Prices. The Democratic price at 6 p.m. ET on 3 November 2026, the last full hour before polls began to close: Kalshi's last trade in the hour, and Polymarket's last recorded price.
• Brier score. The average squared difference between the price and the outcome (1 if the Democrat won, 0 if not). Zero is perfect; always saying 50% scores 0.25.
• Log loss. The average negative log of the price given to the actual winner. It punishes confident misses harder than the Brier score.

The scorecard by race type

Source: our analysis. The 65 markets include the Senate and House control markets.

The overall numbers are flattered by the many safe races, where both platforms priced the winner at 95% or more and were right. Those races make up most of the sample but say little about forecasting skill. The competitive group is the real test. There, a Brier score of about 0.11 is respectable but not remarkable: it reflects three misses in 18 races, all in the same direction.

On the average price given to the eventual winner, the platforms were almost identical: 89.6% on Kalshi and 89.9% on Polymarket.

The misses

Prices at 6 p.m. ET on 3 November 2026.

Iowa was a coin flip, and the two platforms sat on either side of 50%: Kalshi's favourite lost, Polymarket's won. That single race accounts for the difference in the favourite count between the two.

Where the error came from

Most of each platform's Brier score came from a handful of races.

Source: our analysis.

The three upsets alone account for more than half of Kalshi's total error across 65 markets. Polymarket's edge over Kalshi comes mostly from one race: it gave Hinojosa 38.5% in Texas against Kalshi's 34%, and so was "less wrong" when she won.

Did the last day matter?

Prices barely changed in the final 24 hours, and neither did accuracy. Using prices from 4 p.m. ET on 2 November instead of 6 p.m. ET on 3 November gives almost identical scores: 0.0329 for Kalshi and 0.0322 for Polymarket. The markets' view was set before Election Day; the day itself added little.

Senate vs governors

Both platforms did slightly better on governor's races than on Senate races. Governor's races are often less nationalised than Senate races: voters split their tickets more readily for a familiar incumbent governor, as Vermont's Phil Scott showed again. But the gap is small, and the Senate group contains two of the three upsets, which alone explains most of the difference.

What would a perfect forecaster score?

A forecaster who knew every result in advance and priced each winner at 100% would score zero. One who priced every race at 50% would score 0.25. Pricing each race at the platform's own price, as we did, gave about 0.03 overall. That low number is flattered by the 47 markets where one side was priced at 90% or more and won. The competitive group, at about 0.11, is the more demanding test: it is roughly what a forecaster scores if they are right about direction most of the time but miss a few close calls badly.

Calibration: were 70% favourites right 70% of the time?

A well-calibrated forecaster's 70% favourites win about 70% of the time. Here is how the Democratic prices compared with results.

Source: our analysis. Small groups make individual rows noisy.

The pattern is clear even with small numbers. In every band from 30% to 90%, Democrats won more often than their price implied. The ten races priced between 70% and 90% all went Democratic. Polymarket's table shows the same pattern.

This is what a correlated error looks like. Each price on its own was reasonable. But the races were not independent: whatever the polls and the markets missed about the electorate, they missed it everywhere at once. A wave that pushes one close race toward the Democrats pushes the others too.

Why one election does not prove a bias

It would be wrong to conclude from this that prediction markets systematically underrate Democrats. In a single election, a correlated miss in one direction is exactly what you would expect about as often as not. In other recent elections, the misses have gone the other way. What the 2026 results show is narrower: in a wave election, the race prices together understate how lopsided the result can be.

The chamber markets did better on this. The Senate control market priced Democrats at 75% on Kalshi and 74.5% on Polymarket, high enough that a Democratic Senate was the expected outcome, without pretending the races were independent. Our eve-of-election snapshot showed that independent race prices would have implied about 88% for Democratic control; the market's lower figure was its way of allowing for exactly this kind of correlated miss, in either direction.

How this compares

RotoWire, scoring Kalshi's settled markets across all race types on the morning after the election, found the favourite won 479 of 487 races, or 98%, and put the Senate and House control markets' Brier scores at about 0.06 and 0.003 to 0.005. Our numbers for the same two chamber markets are similar.

Our earlier study of prediction market accuracy, across thousands of resolved markets of all kinds, found the markets well calibrated on average. The midterms fit that picture with one caveat: averages hide the correlated misses that matter most on a single night.

Speed

Accuracy is one measure; speed is another. In the competitive races, the winner's price passed 95% about an hour after the state's polls closed, long before the platforms settled. Our study of how fast the markets called the races has the details.

Method and limitations

• Small sample. Only 18 races were genuinely competitive. One more or one fewer upset changes the competitive Brier score noticeably.
• One snapshot. We used prices at 6 p.m. ET. Prices a day or a week earlier would give different scores.
• Matching. Two settled races could not be matched to Polymarket and are excluded; races with notable third-party candidates were scored on the two-party contracts.
• Open races. Mississippi and Alaska are excluded until they settle. On the morning of 5 November both were priced above 90% for the Democrat on Kalshi; if they go that way, they would add two more races in which Democrats won after being priced as underdogs (Mississippi) or modest favourites (Alaska) before the polls closed.

FAQ

Were the prediction markets accurate in the 2026 midterms? Broadly yes. The favourite won 62 of 65 settled markets on Kalshi and 63 on Polymarket, including both chamber-control markets.

Which platform was more accurate? Polymarket, slightly: a Brier score of 0.0318 against 0.0330 for Kalshi.

What did the markets get wrong? They underestimated Democrats in close races. All three misses went to Democrats, and Democrats won 15 of the 18 races priced between 10% and 90%.

Does that mean markets are biased? One election cannot show a lasting bias. It shows that race-by-race prices can understate how one-sided a wave election turns out.

Download the data

The CSV file lists each of the 65 markets with the winner, the Kalshi and Polymarket prices at 6 p.m. ET on 3 November and each platform's Brier score.

Data: Kalshi and Polymarket public APIs; RotoWire (4 November 2026) for its comparison figures. Not investment advice.