Election polls are accurate but can only reveal voter intentions on the day they were taken. They don’t predict the future
First, election polling comprises only a very small fraction of U.S. survey research.
[...]
And here there is excellent news: Despite low response rates, scientific surveys have “still got it” as the errors found in nonelection surveys are typically very low and not generally correlated with response.
Today's polls on Biden versus Trump are not wrong, per se, they just tell us what voters are feeling today. And a lot can happen between now and Election Day.
Second, polls have real margins of error, which are reported too modestly. True ranges of error in election polls average around 5 percent, and many elections are won and lost with much smaller margins.
Third, election polls have a unique characteristic not encountered in any other survey: they survey a population that does not (yet) exist. Election pollsters must predict who will actually vote. Their likely-voter models tend to be 80 percent accurate, leaving quite a bit of imprecision in deciding who among those polled should actually count in a “horse race” estimate.
Finally, surveys can be designed and fielded with very high rigor, but that is costly. Many election polls, however, are conducted with the lowest possible budget, and journalists and the public are often not equipped or otherwise don’t consider the “build quality” that is put into the polls they report and/or consume.
www.scientificamerican.com/...