The generic ballot: the national vote behind every race

Summary

The generic ballot asks voters which party they want to control Congress, without naming candidates. It is the only national measure of the House vote, and in our forecast it sets the national environment that every House district, Senate race and governor's race moves with. This paper describes how we average it, why we correct it, how uncertain the result is, and how it reaches each race.

On 15 October 2026 the average of nine generic-ballot polls by seven pollsters, each adjusted for its pollster's lean and weighted by recency, sample and record, gives Democrats 54.0% and Republicans 45.4% of the decided vote, a margin of D+8.7. Our national House vote is D+6.6, because the final generic-ballot polls have overstated the Democrats at 11 of the 13 House elections since 2000 and taking off half of that past miss passed the leave-one-out test at 11 of them. The national House vote carries a standard deviation of 5.2 points: the miss of past final generic ballots after correction, plus the drift of the 19 days left.

What the generic ballot measures

The question differs between pollsters ("If the elections for Congress were being held today, would you vote for the Democratic candidate or the Republican candidate in your district?" is a common form), and so do the people asked: likely voters, registered voters or all adults. It is a national question about 435 local contests, and three features of the real House vote make it an imperfect target:

• some districts have no candidate from one party, so the national House vote counts no votes for that party there;
• candidates matter: an incumbent runs ahead of the generic party;
• the national vote is a sum over districts with very different turnouts.

We nevertheless treat the generic-ballot margin as the national House vote margin, and test that choice the only way we can: by how far past final generic ballots fell from the national House vote, and whether the forecast built on them called districts right (Correlated error and the backtests of 2018, 2022 and 2024).

Which polls count

Every poll of the 2026 generic ballot held in the database, from the pollsters' own releases where they could be read (YouGov/The Economist, Echelon Insights, Emerson, Big Data Poll, Quinnipiac, Cygnal, Fox News, Marist, Siena, Harvard CAPS/Harris and Marquette among them) and otherwise from Wikipedia's tables. The database holds 223 generic-ballot polls since January 2025, 150 of them since 1 January 2026, by 24 pollsters.

A poll enters today's average if its fieldwork ended within 21 days of the newest generic-ballot poll. One version per pollster and day is kept: likely voters if published, then registered voters, then all adults. Each poll's shares are put on the decided vote: a poll with Democrats 46, Republicans 41 and 13 undecided becomes 52.9 to 47.1.

The weighted average

Since 13 October 2026 the generic-ballot average is weighted and adjusted the same way as the race polls (an owner's setting, "Weight and adjust the generic ballot like the race polls"). For poll \(j\) by pollster \(f\), with margin \(y_j\) on the decided vote, the adjusted margin is

where \(h_f\) is the pollster's house effect from its past final polls (Rating and adjusting American pollsters) and \(\pi_j\) is 2.52 points towards the sponsor's side for a poll paid for by a party, campaign or allied group, zero otherwise. Its weight is

where \(a_j\) is the poll's age in days against the newest poll, \(n_j\) its sample (800 if not published), \(q_f\) the pollster's record weight (its typical final-poll miss against the typical pollster's, between 0.6 and 1.4), \(\kappa_j\) one half for a partisan poll and one otherwise, and \(k_f\) the number of the pollster's polls in the window, so that a pollster's several polls share one pollster's weight. The 14-day recency scale is shorter than the 30 days used for race polls because the window is short and national polls are frequent.

Today's average:

Table: The generic-ballot polls in today's average. Margins on the decided vote; house effect positive where the pollster has leaned to the Democrats. Weights by equation (2).

The weighted mean of the adjusted margins is D+8.70. Without the adjustments, the same weights give D+8.31: Rasmussen's and Cygnal's Republican lean, and Cygnal's Republican sponsorship, are taken off, which moves the average 0.4 points to the Democrats. The shares are reported as D 54.05 and R 45.35: the weighted mean of each party's decided share, each moved by half the weighted adjustment.

Rasmussen's three polls share one pollster's weight, so the average rests on seven pollsters. The older rule, still available as a setting, takes each pollster's latest poll in the window and counts them equally; on today's polls it gives D+8.6.

What the tests say about weighting

The owner switched weighting on after a test of 13 October 2026, and the evidence is mixed. Run today, the tests in the code give:

Table: Average miss on the margin, points. The first two use house effects fitted without the election tested; the generic-ballot test uses only the eight elections with two or more final polls.

Weighting helped clearly on the national presidential polls, where the record is longest, and was slightly worse on the other two. A third variant, which also measures each pollster's lean in this cycle, predicted the next 2026 poll best (2.38 points) but is not switched on, because the past-election test showed no clear gain from it (Rating and adjusting American pollsters). The two averages differ today by 0.1 points.

Why the generic ballot is corrected

The rule is the one used everywhere (Forecasting the 2026 midterms: an overview): take half the average past miss off each party's share, and keep the correction only if it beats the polls as published when each past election is predicted from the others.

We hold final generic-ballot polls at all 13 House elections from 2000 to 2024:

Table: Final generic-ballot polls against the national House vote. Margins in points; the polls' margin on the decided vote. ORC: Opinion Research Corporation, which polled for CNN.

The polls overstated the Democrats at 11 of the 13 elections, by 4.2 points on average. Five of the elections rest on one poll, and the three largest misses (2002, 2014 and 2020) on one to four polls, so the record is thin; but the direction is the most consistent of any polling record on the site.

The correction is computed on each party's share, as everywhere. The polls' decided shares are compared with each party's share of the national House vote; averaged over the 13 elections the polls had the Democrats 2.43 points too high and the Republicans 1.77 points too low. Half of each is taken off today's shares, with a small share for others, and the shares are rescaled to their total:

Our national House vote is therefore D+6.6, against D+8.7 in the polls.

The test:

Table: Leave-one-out test of the generic-ballot correction. Average miss per party on the D and R shares, points. Each election is corrected with the mean miss of the other twelve.

Two cautions belong with this result. First, part of the measured miss is a difference of basis: the polls' shares are of the decided vote, which in most polls adds Democrats and Republicans to nearly 100, while the official House vote gives minor parties and independents two to four points and gives a party nothing in districts it did not contest. The correction learns that gap along with the polls' lean. Second, the two misses that went the other way, 2008 and 2018, were Democratic waves; a correction learned mostly from close and Republican years did worse in them. The test passes on the record we hold, and it is rerun every time the forecast is built.

How uncertain the national vote is

The standard deviation of the national House vote is

with three parts:

• \(\sigma_{\text{GB}}\), the miss of the final generic ballot after correction: the root mean square, over the 13 elections, of twice the corrected average miss per party (twice, because with two parties a miss per party of \(x\) is roughly a miss of \(2x\) on the margin). It is 4.78 points.
• \(\delta \cdot d/36\), drift to polling day: \(\delta = 3.2\) points is how far statewide polls taken 30 to 45 days out moved by polling day, beyond the final polls' own miss (From a month out to election night), scaled by the days left, \(d = 19\). It is 1.69 points today and falls to zero on polling day.
• one point for everything else.

Today \(\sigma_E = \sqrt{4.78^2 + 1.69^2 + 1} = 5.17\) points. An 80% range for the national House vote is therefore about D+6.6 ± 6.6, from roughly even to D+13.

How the national vote reaches each race

House districts move one for one: in the district equation of From the generic ballot to 435 districts the national House vote enters with a coefficient of 1, and the district's lean, incumbency and intercept are added to it. A district that the fundamentals put at R+1 on today's national vote of D+6.6 would be D+0 if the national vote were D+7.6.

Senate and governor priors move by less. The fundamentals fitted on every contested statewide race since 2000 give the national House vote a coefficient of 0.73 for Senate races and 0.54 for governors (The fundamentals: partisanship, the national vote and incumbency). The weight of the prior falls as polls come in, so a well-polled Senate race moves only about a tenth as much with the national vote as an unpolled one.

Race polls move with the national trend. A separate national trend line is drawn through every generic-ballot poll since January 2025, each pollster's lean and partisan lean taken off, smoothed with a 14-day Gaussian kernel and fitted locally as a line where enough polls lie around the day, so that its newest end is not dragged back towards older polls. Each race poll is moved by how far that line has moved between the poll's fieldwork and today (State polls: the pollster-adjusted average and the blend). The line stood at D+7.4 on 1 August, D+8.1 on 1 September, D+9.0 on 22 September and D+9.0 today; the older the poll, the more it is moved, and on average the polls in today's race averages have been moved 1.1 points towards the Democrats.

The simulation draws the national error once per run and applies it to every race, scaled by the part of each race's uncertainty that comes from the national vote (Correlated error and the backtests of 2018, 2022 and 2024).

How much the chambers depend on the national vote is visible in the readings of past misses. Moving every race by 1.1 points towards the Republicans (the 2022 reading) takes the Democrats from 232 House seats to 222, and from 53 Senate seats to 51.

Approval

The president's approval is collected (158 polls since January 2025) and shown on the polls page. It is not used in the midterm forecast: the national vote comes from the generic ballot alone, and no test has shown that approval adds to a generic ballot taken this close to polling day. Approval is central to the 2028 forecast, where no generic presidential ballot is reliable two years out (The 2028 presidential forecast: a party contest with candidates on top).

Limits

A short and uneven record. Thirteen elections, five with one final poll and only four pollsters before 2010. Most of the record comes from Gallup, CNN's pollsters (ORC, then SSRS), ABC News/The Washington Post, Fox News and Pew; other pollsters' final generic ballots have not been collected.

Decided vote against official shares. The correction learns the polls' lean together with the gap between a two-party decided vote and an official vote that includes minor parties and uncontested seats.

The House vote is not the sum of district contests. Uncontested districts and incumbents' personal votes mean that a national margin can be reached in different ways; the district model handles incumbency and uncontested seats separately, but the national target is the official total.

Weighting is a choice with mixed evidence. The weighted average is in use; its tests favour it on presidential polls and not on generic-ballot polls. The difference today is 0.1 points.

References

Clerk of the House of Representatives. Statistics of the Presidential and Congressional Election, 2000 and 2024; Federal Election Commission, Federal Elections 2002–2022.

Figures as of 15 October 2026. The pages they come from are rebuilt with each new poll.