Imagine you are following one race, Pennsylvania in the 2024 US presidential election, and you want to know who is ahead. The sensible-sounding move is to average the polls, and to trust the average more the more polls you put in it. Below are the 36 real polls of that state published in the last fortnight of the campaign, in the order their fieldwork finished. You decide how many of them go into your average.
The scale below runs from a Trump lead on the left to a Harris lead on the right. The black dot is your average. The gold band around it shows how much those polls disagree with each other: the range that their own disagreement justifies. The red line is what actually happened when the votes were counted, and it never moves.
your averagehow much the polls disagreewhat actually happened
Two things worth saying plainly. The gold band is built from how much these polls disagree with each other, not from any single poll's own printed margin of error, which is several points wider. And this lean is not a law you can subtract: polling errors in the 2018 and 2022 US midterms were small and ran in both directions. Three presidential cycles leaning one way is a pattern nobody could see in advance.
Data: 36 non-partisan polls of Pennsylvania with fieldwork ending 22 October to 5 November 2024, from the Internet Archive's 29 November 2024 crawl of FiveThirtyEight's poll file, against the certified result from the Federal Election Commission. Polls are added in fieldwork-end order, which is one real order out of many.