With twelve days to go, a fair question is how much of what the polls show now will still be true on election night. History gives a reasonably clear answer.

We used FiveThirtyEight's archive of 11,500 Senate, governor and House polls from the eleven federal elections between 1998 and 2022, each paired with the result that followed. All of them were taken in the final weeks of a campaign.

Polls do get better, slowly

A poll taken 15 to 21 days before the election missed the final margin by 5.8 points on average. One taken 8 to 14 days out missed by 5.5, and one taken in the final week by 4.9.

So the last three weeks shave about a point off the typical error of a single poll. That is real, but it is not a transformation. A final-week poll is still, on average, about five points away from the result, which is why no serious forecast treats any one survey as decisive.

Averages do better than single polls, and improve in the same way. In the 520 races that were polled both two to three weeks out and in the final week, the average of the earlier polls missed by 5.1 points and the later one by 4.5. Those averages moved about three points between the two windows in the typical race, though much of that is the noise of a race having only one or two polls in each window rather than voters changing their minds.

What a lead two weeks out is worth

The more useful question is what a given lead, two to three weeks out, has meant for who actually won.

Average lead, 2–3 weeks out	Races	Leader won
Under 2 points	128	49%
2 to 5 points	205	66%
5 to 10 points	192	85%
More than 10 points	467	98%

A lead under two points at this stage has been, historically, no lead at all: the candidate ahead won about half the time. Between two and five points, the leader won two times in three. It takes a lead of five to ten points before the outcome is reasonably safe, and even then roughly one race in seven went the other way.

What that means this year

Several of the races deciding the Senate and governorships this year sit in the first two rows of that table. The Senate races in Texas, Iowa, Alaska and Kansas, and the Georgia governor's race, all have average leads under two points as of this week, which history says should be read as genuinely open, whichever candidate is nominally ahead.

It also explains why the probabilities on this site are rarely as extreme as the polls might seem to justify. A three-point lead sounds clear in a headline, but in the historical record a lead of two to five points at this stage has lost about a third of the time. A forecast that said otherwise would be claiming more than the polls have ever delivered.

Caveats

These are polls of individual races, not of the national environment, and the errors in a given year tend to run in the same direction across many races at once, which is the part of polling error that decides close chambers. The archive also ends in 2022, and polling methods have changed since. The pattern, though, has been stable across a quarter of a century: the final fortnight sharpens the picture a little, and close races stay close until they are counted.