How our poll averages work
We average public polls with a method we wrote down. It is an average, not a forecast. It does not say who will win.
Poll data: VoteHub, CC BY 4.0 (license). Averages are our own calculation from that data.
Where the polls come from
All polls come from the free VoteHub polling data. The data is licensed Creative Commons Attribution 4.0, so every page and chart that uses it says so and links to the license. We keep a dated copy of each response so any chart can be rebuilt. We average 2026 generic congressional ballot, presidential approval: donald trump, favorability: donald trump.
We check each poll before using it. A poll must have both expected answers (for example Approve and Disapprove) exactly once, with percentages between 0 and 100. A handful do not, and we set them aside and list them on each page rather than guessing what was meant.
How each poll is weighted
A poll's weight is the product of five parts. Each part is a simple rule.
- Recency. A poll loses half its weight every 10 to 56 days. The half-life adapts to how many polls there are: it is 280 divided by the number of polls that ended in the last 30 days, kept between 10 and 56 days. When polls are plentiful, old ones fade fast. When they are scarce, we hold on longer.
- Sample size. Weight grows with the square root of the sample size, with the sample capped at 2,000. A poll of 2,000 counts twice as much as a poll of 500, and a poll of 8,000 counts the same as 2,000. If no sample size is given we treat it as 400, which is our own conservative assumption.
- Who was asked. Likely voters count fully, registered voters 0.85, and all adults or unstated 0.7. When one pollster reports the same survey for several groups, we count only the group closest to likely voters.
- Busy pollsters. If a pollster has k polls on the same subject in the last 60 days, each of its polls is multiplied by 1 divided by the square root of k. Four polls from one pollster together count twice as much as one, not four times.
- Sponsors. Polls sponsored by a party-aligned group (VoteHub marks these partisan) count 0.5 as much. Internal campaign polls also count 0.5 as much. A poll that is both counts a quarter.
The average for each answer is the weighted mean of that answer across polls. The gap is the weighted mean of each poll's own gap.
The range
The shaded band and the ranges are plus or minus 1.96 standard errors. The standard error adds two things: how far the polls spread around the average (shrunk by the effective number of polls, which is smaller when a few polls carry most of the weight), and each poll's own sampling error from its sample size. It describes our average only. It does not include errors that many polls share, such as who ends up voting, so the true uncertainty is larger. It is not a forecast.
When there is no average
If the newest poll behind a series ended more than 60 days before the date shown, we show no average and say why. Each point on the history chart uses only polls that had ended by that date, so the past is not rewritten by later polls.
What we do not do
- No forecast, and no probability of winning.
- No pollster quality ratings and no adjustment for a pollster's habitual lean. We have no track record data to support either yet.
- No state or district polls. Only the national series above.