The final polls understated Democrats by 4.2 points

Every final poll, against the result

After 2020, the error turned around

In 2020 the final polls of Senate, governor and House races ran D+5.9 against the results (317 polls). In 2022 they ran R+1.0 (431 polls), and this year R+4.2 (496 polls). That is the direction an over-correction takes. The figures cannot say why: a pollster who adjusted toward Republicans and one who simply reached too few Democrats look the same here.

Final polls of Senate, governor and House races and the national House vote, between a Democrat and a Republican: each pollster's last poll of a race in the final three weeks. Earlier years come from the site's graded history, which leaves out pollsters with fewer than five graded polls in all; 2026 includes every pollster. Earlier years: Historical polling rows from FiveThirtyEight / ABC News plus CTM's governed supplement.

This year no way of polling escaped it. By how the poll was conducted, as the New York Times table records it, and by who was counted:

The ranking

Avg. miss is how far a pollster’s margins were from the results, in points. Leaned is the party its polls favored compared with the results. Vs. field compares each of its polls with the other pollsters’ final polls of the same race, so nobody is marked down for polling races everyone got wrong. Score is that figure pulled toward zero when a pollster has few polls; lower is better, zero is an average pollster. Select a pollster for every poll behind its row.

Who saw the wave, and who missed it

The 28 ranked pollsters, from the one whose final polls ran most Democratic against the results to the one whose polls ran most Republican. 27 of the 28 leaned Republican. The number before each name is its rank.

The polls that looked too Democratic

Some late polls in Texas and Kansas showed Democratic leads large enough to be read as outliers. Across the 29 contests polled by five or more pollsters (349 polls), the most Democratic final poll was closer to the result than the average of the polls in 19, and was the closest poll of all in 15. Sorting each contest’s polls into thirds by how Democratic they were:

Each dot is one pollster’s final poll. The solid line is the result and the dashed line the average of the polls. Red dots were more Republican than the result, blue more Democratic.

Partisan and campaign polls, scored apart

A poll paid for by a campaign, a party or an allied group is usually published because its sponsor likes the number. 120 of the 526 final polls had such a sponsor. They are in every figure on this page and marked on each pollster’s page; here they are separated.

The poll’s leader won in 87% of 392 with no partisan sponsor; 69% of 54 with democratic sponsor; 80% of 25 with republican sponsor; 81% of 37 with other political sponsor, counting called races.

The national polls

Democratic House candidates have 56.2% of the votes in our counts and Republicans 42.3%, a margin of D+13.8 over 117,941,629 votes in all 435 districts. That is our own sum of the official district counts, not another outlet’s tally. In the 415 districts that have been called, about 97% of the expected vote is counted; 20 districts are not called, and 18 had no Democrat or no Republican on the ballot. Each pollster’s last national poll, closest first (30 polls):

On Point Polling, which asked us to score its final national poll, is the shaded row: number 2 of 30, 0.7 off. Its other claims are checked below.

The whole field

Averaging the final polls of a race did not cancel the miss. The simple average of final polls was off by 5.3 points across 128 contests, and in 93 of the 117 between a Democrat and a Republican it had Democrats too low.

Where the averages missed most

Did pollsters huddle?

Polls of the same race should disagree at least as much as sampling error predicts. If they agree more than that, someone is adjusting toward the pack. In the 31 contests polled by five or more pollsters (355 polls), final margins spread 3.6 points around their average where sample sizes alone predict 3.5: a ratio of 1.03, so the field as a whole shows no sign of it. By pollster, for those with 8 or more comparable polls, the same ratio (under 1 is closer to the others than sampling predicts), closest-clustered first:

On Point Polling's accuracy claims, checked

On Point Polling asked to be scored. Its 19 final polls are in the tables above: average miss 4.3, leaned R+1.9, rank 10 of 28. Its forecast is not a poll and is checked separately here, from the files it committed to a public repository at 13:58 UTC on Election Day, nine hours before the first polls closed, and from its post on X half an hour before they closed. Our counts are not final.

Each entry gives a claim the firm made to us or on its own accuracy page, what we found, and where. A claim is marked partly confirmed when the firm’s number rests on a count read on November 4 or on its own grading rule, and ours differs.

1. Final national poll: Democrats +13.1 among likely voters

   Confirmed. The release, posted October 29 at 01:16 UTC after one day in the field (October 28, 1,170 registered voters online), has likely voters at 55.0 to 41.9 once leaners are included. The same release has 52.5 to 40.5 before leaners and 53.1 to 41.6 among registered voters.

   On Point's release

2. The House vote is 13.2 points with 93 percent counted (RealClearPolitics)

   A different yardstick. We do not use another outlet's tally. Adding up the official counts in all 435 districts gives Democrats 56.2, Republicans 42.3: D+13.8 on 117,941,629 votes, with 20 districts not called and about 97 percent of the expected vote counted in the rest. On Point's own accuracy page uses a projected D+13.0, not a count.

   Scorecard, national polls · On Point's accuracy page

3. The closest of 42 final national polls

   Not confirmed. Against our count it is number 2 of the 30 final national polls we have, 0.7 off. AtlasIntel (D+14.1) is 0.3 off. Next is The Honest Poll (D+13.0), 0.8 off. The order at the top depends on where the count ends. We count one final poll per pollster in the last 21 days, which is why we have 30 and not 42. On Point's own page ranks itself first on a score that also uses each party's share, against a projected D+13.0.

   Scorecard, national polls · On Point's accuracy page

4. State polls named 14 of 18 winners

   Confirmed. Its 18 final state and district polls: in the 18 races that are called the poll's leader won 14 and lost 4 (Nebraska Senate, Kansas governor, Texas governor, Nebraska 1st District). Counting leaders, that is 14 of 18. 4 of the 18 were paid for by the Nebraska Democratic Party.

   Our page for On Point Polling

5. The forecast called 476 of 506 federal and governor races

   Partly. The frozen forecast has a favorite in all 506 races. On our counts the favorite won 458 of the 484 races that are called and leads in 12 of the 19 that are not: 470 of 503, or 473 if a seat counts as right when the favored party holds it. 3 districts have no usable count. On Point's 476 was read on November 4 and, as its own review says, counts leaders in unfinished races.

   Frozen forecast, committed Nov. 3, 13:58 UTC · On Point's review

6. All 315 Safe ratings were right

   Partly. On Point defines a safe call as a race it gave a 95 percent chance or more. The frozen file has 324 such races; 321 have a usable count and 315 of those are called. 315 of the 315 went to the favored party. Not called, with the favorite behind: North Carolina 5th District. Called, between two candidates of one party, and the named favorite lost: California 34th District.

   Frozen forecast, committed Nov. 3, 13:58 UTC · On Point's accuracy page

7. The closest House seat forecast, at 247 Democratic seats

   Partly. The number is confirmed as published before the polls closed: On Point's post on X, created at 22:30 UTC on Election Day, 30 minutes before the first polls closed, gives the House as 247 Democrats and 188 Republicans, every race called for its favorite. That matches the frozen file, where the Democrat is ahead in 247 districts. The same file's simulations average 255.3 Democratic seats, and On Point's review cites a median of 254. On our counts Democrats have won 250 districts and lead in 12 more of the 20 not called, 262 in all, so 247 is 15 short. Whether that was the closest forecast is a comparison of forecasters, which this page does not make.

   On Point's post on X, Nov. 3, 22:30 UTC · Frozen forecast, committed Nov. 3, 13:58 UTC · On Point's accuracy page

8. The legislative forecast called 80 of 83 chambers

   Partly. The frozen file gives control odds for 87 chambers. Our counts settle 62 of them, and the party it favored took 61; it missed the Pennsylvania Senate. The other 25 are not settled in our counts. On Point grades itself on its district calls added up, on its own reading of the results, which is how it gets 80 of 83.

   Frozen forecast, committed Nov. 3, 13:58 UTC · On Point's accuracy page

What we could not verify. Internet Archive captures of the forecast pages from November 2 and 3 are listed in its index but could not be opened (rate limit). The firm's own post on X, created 30 minutes before the first polls closed, states the seat counts it was publishing.

How this was scored

What is provisional, and who is missing