How accurate were prediction markets on the 2026 midterms? A scorecard of 504 races

Short answer: They got the winners right and the size wrong. When the first polls closed on Nov 3, Democrats cost 94.5¢ to win the House and 75.5¢ to win the Senate on Polymarket, and they won both. The favorite won 439 of the 447 races Polymarket has settled. But all 8 misses went to Democrats, and the seat markets implied about 192 Republican House seats. NBC News projects 176.

These are early numbers, as of Nov 5. Some House races and the Alaska and Mississippi Senate races are still open, and we'll update this page when they're called.

The scorecard

Prices are Polymarket's; Kalshi's were within about 1¢. Seat numbers are the median of the seat buckets on both venues, and the vote margin is Polymarket's. We published the Oct 29 column before the election.

Race by race, the favorite almost always won

We took the price of every House, Senate and governor race on Polymarket at 6 PM ET on Nov 3, when the first polls closed. That's 504 races, and 447 have settled.

That's 98.2% overall. Kalshi reports the same kind of number: its favorite won 479 of 487 races. It's a low bar, though, because most races aren't close. 338 of these 447 cost 95¢ or more.

The better test is the other 109. There the favorite cost 83.7¢ on average and won 92.7% of the time, so favorites were too cheap as a group.

All 8 upsets went one way

Split those 109 contested races by party and the pattern is plain.

Every favorite that lost was a Republican:

More may follow. 54 House races hadn't settled, and Republicans were the favorite in 44 of them. If NBC's projection holds, Democrats need 24 of the 54, and their prices in those races added up to 18.9.

So this wasn't random noise. A national swing moves every race the same way, which makes these 109 races closer to one big trade than to 109 separate ones. One election can't tell you whether the market is calibrated at 70¢.

The size is where they missed

Winner markets only need the sign. Seat markets need the size, and that's where the price was furthest off.

Add up the Democratic price in all 435 district markets and you get 247 seats. The seat buckets said about 243. So the two kinds of market roughly agreed with each other, and both were 12 to 16 seats short of NBC's 259.

A big miss wasn't priced as impossible. "Below 190" Republican seats cost 42.6¢ on Polymarket at 6 PM ET, and "below 193" cost 50.5¢ on Kalshi. But the size that actually arrived was far out in the tail: the 175–179 bracket cost 3.4¢, on a thin market. In the Senate, 44 Republican seats cost 6.5¢.

Depth may be part of it. Polymarket's market on who wins the House traded $63.3M. Its House seat buckets traded $0.7M. The question most money was on was the easier one.

Markets vs the models

The models missed the same way.

The markets were a little closer than these models, but two seats between them and Decision Desk HQ isn't a real gap, and it's one election. Silver Bulletin's final numbers are paywalled, so we show the last ones we could verify.

The polls were further off. The average of generic ballot polls was about D+8.6. The market was already at D+10.1 when polls closed, and it now prices about D+13.4 while votes are counted.

Election night: three hours

The move came fast once votes were counted. All times are Eastern.

Polymarket's numbers, with its lower brackets once the median fell below 190. By 9 PM most of the move had happened. That's why we score at 6 PM: a price from later in the night is reading the count, not forecasting it.

Rizon's own books

Rizon's books on these numbers (HOUSE-R-2026, SENATE-R-2026 and HOUSE-VOTE-2026) are quoted by bots that read these same markets, so they aren't an independent forecast. At 6 PM ET they read 191.5 House seats, 47 Senate seats and D+10.

We made one error of our own. On Nov 4, Polymarket's main Senate event closed, and our maker read Kalshi alone. Kalshi's open-ended "below 45" bucket was spread down to a floor of 35 seats, so our Senate book drifted to 40.5 while Polymarket's brackets priced 44. We fixed it on Nov 5.

How we scored this

• Prices: Polymarket's public hourly price history at 23:00 UTC on Nov 3 (6 PM ET), and Kalshi's hourly candles.
• Races: every Polymarket winner market tagged midterms: 435 House, 35 Senate and 34 governor. Races count once settled on Polymarket. Download the 504 races as a CSV.
• Seats: normalize each bucket's price, find where the running total crosses one half, and interpolate inside that bucket, with each seat covering half a seat either side.
• Results: NBC News for the House projection (259–176, as of 8:07 PM ET on Nov 4) and the Senate count.
• Brier score, lower is better: 0.016 across all 447 settled races, and 0.062 across the 109 contested ones.

FAQ

Were prediction markets accurate on the 2026 midterms?

On who wins, yes: Polymarket and Kalshi both priced Democrats to win both chambers, and the favorite won about 98% of races. On how big the win would be, no: they implied about 192 Republican House seats when NBC projects 176.

Did Polymarket or Kalshi do better?

There's no real gap. Their prices were within about 1¢ on both chambers at poll close. Rotowire's Brier scores put Polymarket slightly ahead on both, but the gap is tiny.

Did prediction markets beat the polls?

On the national vote, so far yes. The polling average was about D+8.6, the market was at D+10.1 at poll close, and it now prices about D+13.4. The count isn't finished, so this can still move.

Which races did the markets get most wrong?

Indiana's 5th, where Victoria Spartz cost 78.5¢ and lost, and Texas's 9th at 73.5¢. In the Senate it was South Carolina at 70.5¢, and for governor it was Texas at 65.8¢.

Does this prove prediction markets are biased against Democrats?

No. It's one election, and errors move together in a national swing. It shows one large miss in one direction, which is what any forecast looks like when the swing is bigger than expected.

Originally published at rizon.markets. Rizon is a paper-money exchange for horizon contracts.