Early vote counts missed results by 11 points across 75 races

Read as a forecast, the party split of ballots cast before Election Day missed the final margin by an average of 11.3 points across 75 presidential, Senate and governor races from 2016 to 2024. In the 58 of those races where our database holds a final three-week poll average, the polls missed by 4.1 points and the early vote by 10.8.

That is the main result of a backtest built to answer one question before the midterms on 3 November: does early-vote data improve a projection that already has polls. Of five ways of using the numbers, four failed. One appeared to work, and the reason it worked is the strongest evidence that the information is not in the early vote.

The data cover more than 130 million ballots in states that register voters by party. For 2020, 2022 and 2024 we used the final pre-election reports from Michael McDonald's US Elections Project, now hosted by the University of Florida Election Lab. The archived pages for 2016 and 2018 could not be reached, so for those years we used election offices in Colorado, Florida, Iowa, Maryland and North Carolina, and an Associated Press account of Nevada's 2018 figures. Each state was paired with the certified result of its top race: president, or in midterms Senate, else governor (Connecticut in 2022, where our database lacks the Senate result). That gives 4 races in 2016, 5 in 2018, 20 in 2020, 20 in 2022 and 26 in 2024.

Read alone, the split often points the wrong way

On average the early vote was 7.5 points more Democratic than the result, but the misses were too erratic to correct with one number, with a standard deviation of 13.4 points. In Pennsylvania in 2022, registered Democrats led the early vote by 47.9 points and John Fetterman won by 4.9. In Pennsylvania in 2020 the lead was 43.1 points and Joe Biden won by 1.2. Louisiana's 2022 early vote split almost evenly, with Republicans ahead by 1.0 point, and John Kennedy won by 43.7. Kentucky's 2020 early vote leaned Democratic by 11.7 points and Donald Trump carried the state by 25.9.

Registration is not a vote: Kentucky and Louisiana still have many registered Democrats who vote Republican. And methods differ by party: in Pennsylvania, Democrats voted by mail while Republicans voted on the day.

In the 58 races with polls, the early-vote leader won 44 times and the poll leader 51 times. In the 26 races decided by less than 10 points, the early-vote leader won 14, close to a coin toss. The poll leader won 18 of the 25 of those races with polls.

Correcting for each state's habits is not enough

A fairer test subtracts each state's gap between early vote and result at an earlier election, so a state whose early ballots always run 20 points more Democratic is read 20 points lower. Using the previous election of the same type, president to president or midterm to midterm, the corrected figure missed by 7.1 points across 29 races. In the 25 of them with polls it missed by 6.6 points against 3.9 for the polls, and beat the polls in 12. Using the previous cycle instead, it missed by 7.7 points across 47 races, and by 6.9 against 3.6 for the polls in the 36 with polls.

The correction fails because the habits change. Between 2020 and 2024 the early vote moved 11.5 points toward Republicans on average across 20 states, while the results moved 4.3 points. Between 2022 and 2024 the early vote moved 9.8 points toward Republicans in 20 states and the average result did not move at all. Most of the movement in the early vote was a change in how people voted.

Comparing changes directly gives the same answer. Across 29 pairs of elections of the same type, the correlation between the change in the early-vote split and the change in the result was 0.18, and the fitted slope was 0.10: a 10-point swing in the early vote went with about 1 point of change in the result. Across 47 consecutive pairs the correlation was 0.03. A model built on early-vote change predicted the change in results no better than assuming every state moved by the average.

The one test where early votes seemed to help

The decisive test moves the poll average part of the way toward the corrected early-vote figure, with a weight from zero, polls alone, to one, early vote alone. We chose the weight using every year except one and scored it on the year left out.

On this test, the best weight was above zero. With same-type corrections, the average error was 3.86 points with polls alone, 3.27 at a weight of 0.3 and 6.60 at a weight of one. Scored on the years left out, the blend averaged 3.29 points against 3.86 for the polls. With previous-cycle corrections the blend averaged 3.37 against 3.58. Taken at face value, a 30 percent weight on the early vote improves the projection.

The gain did not come from anything the early vote said about individual states. We split the early-vote signal into its average for the year and the state-by-state differences around that average. The yearly average alone, used the same way, gave an error of 3.06 points. The state-by-state part alone gave 3.87, against 3.86 for the polls. When we shuffled the early-vote changes at random between states within each year, 55 percent of the shuffles did as well as the real data. With previous-cycle corrections the state-by-state part made the error worse, 4.07 against 3.58, though only 26 percent of shuffles matched the real data. Without 2024, the same-type blend scored 4.44 against 4.46 for the polls. The previous-cycle blend still helped without 2024, 3.34 against 3.90, but moving polls toward the previous result, with no early-vote data, did almost as well at 3.39.

The 2024 gain came from a coincidence of direction. The early vote swung sharply toward Republicans, mostly through a change in how they voted, and the polls also underestimated Trump. The two pointed the same way in 15 of 17 states. A rule that uses no early-vote data at all, shifting every poll by the average poll miss of the other years, did better than the blend: 2.88 points against 3.29 on the same-type races, and 3.24 against 3.37 on the others. With previous-cycle corrections, moving polls toward the previous election result, again with no early-vote data, scored 3.08.

Little sign that early votes reveal the polls' misses

The early-vote signal agreed in sign with the polls' miss in 20 of 25 same-type races and 25 of 36 consecutive ones, but those agreements come from the same yearly pattern. Within a single year, the correlation between the change in the early vote and the polls' miss was 0.04 in the same-type set and 0.25 in the consecutive set. The second figure falls only to 0.22 once the previous result is allowed for: a weak link that 36 races are too few to confirm.

We also tried a measure that allows for party registration: the share of each party's registered voters who had voted about three weeks out, compared with the same day in an earlier election. We have it for seven comparisons in North Carolina, Pennsylvania, Maine and Colorado. The correlation with the change in the result was minus 0.64, the wrong direction, although the two moved the same way in five of the seven. In North Carolina between 2022 and 2024, Republicans' early pace rose 2.4 times against 1.2 times for Democrats, and the margin did not move. Seven cases prove little either way.

What forecasters do with early votes

Nate Silver keeps early votes out of his model, with one minor exception, and has written that in 2016, in 13 states where Michael McDonald found party data, the early-vote margin missed the Clinton-Trump result by 14 points on average. Split Ticket wrote in November 2024 that "persuasion can never be learned from early vote." The Economist wrote in October 2024 that divining insights from early votes about the final result "is difficult."

Known misreads fit the pattern. In Florida in 2020, registered Democrats led the pre-Election Day vote by 1.2 points and Trump won by 3.4. In Nevada in 2024, the raw count, which showed Republicans ahead by 3.8 points, was close to Trump's 3.1-point win, but Jon Ralston of The Nevada Independent, drawing partly on the early vote, forecast a narrow Harris win.

Early votes do have real uses. Decision Desk HQ says it collects absentee and early-vote tallies alongside election-night returns, and that its race calls weigh which types of vote are in and which are still to be reported.

Our forecast will not use early-vote figures for 3 November. A state-by-state signal that survived the shuffle test in future years would change that. We will rerun these tests once results are certified.