From a month out to election night
Abstract. How far American polls move before polling day, and what the site does when the count begins. Statewide averages taken 30 to 45 days out missed by about 6.9 points on the margin, against 6.1 two weeks out; the 3.2 points of drift are added to every race's error in proportion to the days left. On election night the site reads 17 states' official results feeds and a live model turns each count into a chance. Replayed on the 2024 Senate races, it was well calibrated with a neutral count order but less accurate than the forecast early in the count when early votes leaned to one side.
Summary
This paper covers the last stretch: how far American polls move between a month out and polling day, how the forecast allows for it, and what the site does on election night, when the polls give way to the count.
Statewide poll averages taken 30 to 45 days before the elections since 2000 missed the result by about 6.9 points on the margin, against 6.1 for averages taken two weeks out. The difference, 3.2 points, is the drift we allow for: it is added to every race's error, and to the national House vote's, in proportion to the days left, so that it is 1.7 points with 19 days to go and falls to zero on polling day. The forecast does not pull the polls back towards the last result; the fundamentals already do that work where polls are thin.
On election night the site reads the official results feeds of 17 states and turns each race's count into a chance with a live model: each county's count is compared with what our forecast expected there, the swing is projected onto the votes still to come, and the chambers are re-simulated with the counted races fixed. Replayed on the 2024 Senate races with a simulated count order, the live model was well calibrated when early and late votes leaned alike, but when early votes leaned Democratic, as they did from 2020 to 2024, it was less accurate than the pre-election forecast until about three quarters of the vote was in. The live model is behind an owner's switch and stays off until the night.
How far polls move before polling day
The evidence
For every Senate, governor and presidential race by state since 2000 we hold the polling average at fixed distances before polling day: 14, 30, 45, 61 and 91 days. At 14 days that is 814 races: 312 for the Senate, 245 for governor and 257 presidential contests by state. At each distance, the margin of the average is set against the result:
| Days before polling day | Races | Root mean square miss | Mean miss |
|---|---|---|---|
| 14 | 814 | 6.09 | +1.47 |
| 30 | 731 | 6.71 | +1.58 |
| 45 | 701 | 7.04 | +1.34 |
| 91 | 527 | 8.44 | +1.44 |
Table: Statewide poll averages at fixed distances before polling day against the result, Senate, governor and presidential races by state, 2000–2025. Points of margin; positive means the average overstated the Democrat.
Two features matter. The miss grows with distance, from 6.1 points at two weeks to 8.4 at three months. And its mean does not change with distance: averages taken three months out overstated the Democrats by about as much as those taken two weeks out. The polls do not, on average, move towards one party as polling day approaches; they move in both directions, and the further out, the more.
The drift
The part of the miss that is movement, rather than the final polls' own error, is the drift. Taking the 30- and 45-day misses together and subtracting, in squares, the 14-day miss,
This is how far a statewide average taken 30 to 45 days out typically moved by polling day, beyond what a final poll misses anyway.
How the forecast uses it
The drift is added to the error of every race average and of the national House vote, in proportion to the days left, \(d\):
At 36 days the full 3.2 points is added; at 19 days, today, 1.7 points; on polling day nothing. A race whose newest poll is old gets a second term of the same kind, \(\left(\delta\,a/45\right)^2\) for a newest poll \(a\) days old (State polls: the pollster-adjusted average and the blend).
The scaling is linear in the days left. With measurements at only five distances, and two of them used, a linear scale is a simple choice; a square-root scale, as a random walk would give, would add more drift close to polling day and less further out. The national House vote's error is 5.2 points today, of which the drift contributes \(1.7^2 = 2.8\) of its variance of 26.7 (The generic ballot: the national vote behind every race).
What is not done
The British forecast pulls the older parties' shares part of the way back towards their last result as polling day approaches (From today to polling day: how far polls move before an election). The US forecast does not. Its counterpart is the blend with the fundamentals: a race's prior already contains the state's lean, the national vote and incumbency, and it carries more weight the fewer and older the race's polls (The fundamentals: partisanship, the national vote and incumbency). The national House vote itself is not pulled back towards the 2024 result.
The national trend move (State polls: the pollster-adjusted average and the blend) works in the other direction from drift: it brings older polls up to date with the national movement since they were taken, rather than widening their error. Both are applied: the poll is moved by the national change already seen, and the average is widened for the change still to come.
Publication and polling day
No American law bans publishing polls before or on polling day; the First Amendment protects it. The US pages therefore have no blackout switch and stay live until the counts begin. Networks by custom hold projections for a state until its polls close.
Polls close state by state over seven hours. The first close at 6:00 pm Eastern in most of Indiana and Kentucky; the last at 1:00 am Eastern in Alaska's Aleutian Islands. The clocks go back on 1 November, so the night runs on standard time. By 8:00 pm Eastern, 26 of the 51 states and districts have closed all their polls; the West Coast closes at 11:00 pm Eastern. The election-night page lists each state's closing time, its time zones and the close races decided in each hour, from each state's election office.
Election night: live results
The feeds
For each state we looked for the election office's own results feed for 3 November: its format, its terms and its robots.txt, and whether it lets a named reader in. A state is read only if all of that is clear; a Cloudflare, Incapsula or Akamai challenge is a refusal and is never got round. The Associated Press's results are paid and are not used; nor are any media feeds.
Seventeen states pass: Alaska, Colorado, Florida, Georgia, Idaho, Iowa, Minnesota, Nebraska, New Mexico, North Carolina, North Dakota, South Carolina, South Dakota, Utah, Virginia, Washington and West Virginia. They use seven kinds of feed: Clarity election-night reporting (Colorado, Iowa, South Carolina, West Virginia), Enhanced Voting (Georgia, Idaho, Utah, Virginia, Washington), the shared "ESERS" sites (Nebraska, New Mexico, North Dakota, South Dakota), and the state formats of Alaska, Florida, Minnesota and North Carolina. Each is read every few minutes, with county breakdowns at most every 15 minutes. Alaska's feed gives first-choice counts only; the ranked-choice rounds are run after the count. The other states block automated readers or publish no data file on the night, and the site links to their own pages.
Results are unofficial until certified. The site shows who leads the count and never calls a race.
The live model
The live model turns a race's count into a chance. It starts from our pre-election forecast margin \(\hat{m}\) and its standard deviation \(\sigma_0\).
Expect each county. Each county \(c\) gets an expected margin, its 2024 presidential margin moved so that the counties add up to our forecast, and an expected vote, its 2022 midterm turnout (or its 2024 presidential turnout scaled down where no 2022 race is held):
Read the swing. Where votes have been counted, each county's counted margin \(y_c\) is compared with \(\mu_c\). Counties count for more the more of their vote is in: with \(n_c\) votes counted out of \(V_c\) expected and \(f_c = \min(1, n_c/V_c)\), the weight is \(u_c = n_c f_c\), and the observed swing is
A half-counted county counts for less because its counted votes are often one kind (mail, early or Election Day ballots), which lean differently from the rest.
Shrink it. The swing is noisy: one county's swing strays from the race's by about \(\tau\) = 6 points, and the order in which votes arrive biases every partly counted county the same way at once, by up to 10 points when nothing is counted, shrinking as the counties fill. The observed swing's variance is
where \(\bar{f}\) is the weighted share counted, and the swing applied is shrunk towards zero by the pre-election uncertainty: \(\hat{\Delta} = \Delta\,\sigma_0^2/(\sigma_0^2 + v)\).
Project the rest. The votes still to come in each county are projected at \(\mu_c + \hat{\Delta}\), and the race's projected final margin is the vote-weighted mean of the counted and projected votes. Its uncertainty has four parts: the uncertainty of \(\hat{\Delta}\) scaled by the share still to count, each county's own swing on its remaining vote, how many votes are still to come (each county's remaining vote taken as uncertain by 35% of itself), and the count order: a partly counted county's counted votes may lean as much as 14 points from its final result, and nobody can see which way on the night. The chance is the normal probability that the projected margin is above zero.
Where a feed gives only statewide totals, or for House districts (our county baselines are statewide), the race's total is treated as one unit with \(\tau\) widened by half. The model says nothing until at least 2% of a race's expected vote and 500 votes are in.
Roll up. For control of each chamber the forecast's own simulation is rerun, 2,000 times, with every race that has enough votes counted drawn from its live estimate. Races with nothing counted keep our forecast, moved by half the vote-weighted swing seen so far in the counted races, a share that grows to its full value as about five races' worth of expected vote come in.
The live model never calls a race. Like the forecast, it gives a chance. It runs only when the owner switches live results on; until the night the switch is off.
The live model's backtest
No official feed we can read keeps its election-night versions, so a real count cannot be replayed. Instead, each county's real 2024 Senate result is split into three batches (early and mail votes, then Election Day votes, then late mail and provisional ballots) arriving at random times, early-reporting counties first, and the model reads the count as it would on the night. The batches add up to the county's real result but lean away from it by a set amount. The model starts from our 2024 backtest forecast (fitted without 2024), with a standard deviation of 5.3 points, the miss of the 2018 and 2022 backtests, and county baselines from the 2020 presidential result and turnout, so that it knows nothing of 2024 before it is counted. Twenty-six Senate races are replayed 25 times each; California (two Senate races in 2024), Connecticut, Indiana, Missouri and New Jersey (no county Senate result held) are left out.
| Share of expected vote counted | Favourite won | Leader in the count won | Brier, live | Brier, forecast alone | Result inside 80% range | Average miss |
|---|---|---|---|---|---|---|
| 10% | 92% | 82% | 0.053 | 0.041 | 79% | 4.6 |
| 25% | 92% | 87% | 0.055 | 0.041 | 66% | 4.7 |
| 50% | 92% | 90% | 0.049 | 0.041 | 80% | 4.1 |
| 75% | 96% | 96% | 0.026 | 0.041 | over 99% | 1.3 |
| 90% | 96% | 97% | 0.020 | 0.041 | over 99% | 0.5 |
Table: The live model on the 2024 Senate races, early and mail votes 6 points more Democratic than each county's final result and Election Day votes more Republican (the pattern of 2020 to 2024). 650 cases at each point. Brier: the mean squared gap between the chance given and what happened; 0 is perfect, always saying 50% scores 0.25.
Three count orders were tested:
| Count order | Brier, live, at 10% / 25% / 50% counted | Brier, forecast alone |
|---|---|---|
| Early votes 6 points more Democratic | 0.053 / 0.055 / 0.049 | 0.041 |
| Every batch votes like the final result | 0.035 / 0.033 / 0.031 | 0.041 |
| Early votes 6 points more Republican | 0.048 / 0.061 / 0.065 | 0.041 |
Table: Brier scores of the live model in the first half of the count, three simulated count orders.
The backtest gives a clear message. When the count order is neutral, the live model improves on the pre-election forecast from the first tenth of the vote. When early votes lean to one side, the model is less accurate than the forecast it started from until about three quarters of the vote is in, because the count-order bias moves every county the same way and the model cannot tell it from a real swing. Its allowance for count order (equation 5 and the 14-point term) keeps most results inside the 80% range, but not all: at a quarter counted in the early-Democratic order, 66%. At half the count in that order, the model favoured the wrong side in Ohio (a 67% Democratic chance; the Republican won by 3.6 points) and Pennsylvania (91%; the Republican won by 0.2).
The results page sets out this backtest beside its description of the live model. Reading the count as a forecast is safest once most counties are fully in.
Limits
Five distances, two used. The drift rests on averages at 14, 30 and 45 days. The 61- and 91-day points are held but not used, and the linear scaling between them is a choice.
Drift is measured on statewide races. The national House vote takes the same drift without a measurement of its own, and the presidential contests in the measurement may move differently from midterm races.
The live model has never run on a real night. Its backtest replays a simulated count order on real results; real feeds report unevenly and change their formats.
County baselines only statewide. House districts are read as single units, with less information than statewide races.
Seventeen states. The live chances for the rest of the country come from the forecast moved by the swing seen in those seventeen, which may not be representative.
References
Brier, G. W. (1950). Verification of forecasts expressed in terms of probability. Monthly Weather Review, 78(1), 1–3.
Figures as of 15 October 2026. The pages they come from are rebuilt with each new poll.
How to cite
Polling Forecast (2026). From a month out to election night. Retrieved 17 October 2026 from https://www.pollingforecast.com/research/us-election-day. Live figures change as new polls arrive; the date above is when the text was last revised.