Every election cycle has polls that come out nearly every day. The polls will typically state their margin of error, but that’s just a statement of the error range of the poll at that moment in time, and is based on the sample size relative to the overall population. While it might reflect the error range if the election was held that day, it doesn’t seem like it would accurately reflect what will happen if the election is months away. The more time there is between the poll and the election, the more opportunity there is for things to happen which affect the election. With that in mind, have studies been done to determine how good the polls are at predicting the election relative to the amount of time before the election?
As an analogy, consider a weather forecast for the next 10 days. The accuracy of the prediction goes down the farther out the forecast goes. The predicted weather for tomorrow is likely something you can count on (90% accurate?), but the predicted weather for the 10th day is practically just a guess (10% accurate?). If they were predicting rain for tomorrow, you would probably want to cancel your outdoor event. But if they predicted rain for day 10, you wouldn’t really count on that enough to make any decisions. There is too much chaos in weather systems to count on a 10-day forecast. Is there anything like that for election polling? Are there guidelines which say that if the election is X months away, then the accuracy of the poll is decreased by some magnitude relative to X to account for all the chaos that can happen before the election?
Polls do not intended to reflect the final outcome of an election, but instead are intended to be a sampling of political sentiment at that point in time to be used by campaigns to adjust or modify their campaign focus and policy statements. Unlike the weather, campaigns actually interpret and respond to polls, so the notion that they are projecting final outcomes is not a sensible interpretation.
I would still expect some kind of time factor could be used to get some predicative aspect of the election. For instance, if the poll has a candidate 20 points down and it’s 1 month until the election, it seems like it’s almost a certainty that the candidate will lose. One month seems like too short a time to make up 20 points. But if the election is a year away, then 20 points doesn’t seem insurmountable. There’s a lot of things that can happen and can be done to make up that 20 points over the course of a year.
In my mind, I feel like an poll that’s 1 month out might have about +/- 2% range that the poll can move based on things that might change in the next 1 month. For an election that’s a year out, I would put that range at maybe +/- 15%. That’s the sort of thing I’m wondering about. If an election is X months out and a candidate is Y points down, what Z% chance do they have of raising their poll numbers by Y points on election day? That may be impossible to say for any specific candidate, but it seems like historical polling data could be used to determine a probable increase a candidate could expect for a given period of time.
This question is complicated (or…simplified?) by the fact that American politics have become much more stable than they were before. In the past, even huge leads were no gimme - in 1988, Dukakis once blew a 17-point summer polling lead over Bush Sr., which became an 8-point Bush win. Twelve years later, Gore blew a 10-point lead over Bush Jr. (although, in fairness, Gore won the popular vote.)
Today, it’s unlikely any double-digit summer lead would be blown, barring some highly unforeseen event.
538 publishes (or published; I haven’t followed them much since when they were an independent website) both a “now-cast” that says what the odds would be if the election were to be held that day, as well as a forecast for the actual election. The forecast for the actual election has some level of uncertain suprise-factors built in, and so always shows a tighter race than the now-cast (because the surprises could come in either direction). As election day nears, the now-cast and forecast tend to converge together, as one would expect.
Nate Silver has written extensively about all of the details that go into his models, and the statistical analyses behind them, including this time-uncertainty factor.
“Polls” vary widely in any predictive capability, and because they are estimating an outcome that can never be definitively determined (i.e. you can’t rerun an election over and over to determine whether a likelihood presented as a percentage of the universe of possible outcomes is ‘true’ or not) they really represent guesswork with some greater or lesser degree of information. Poll watchers like Nate Silver do keep a record of outcomes and weigh the accuracy of different polling organizations accordingly, but even that is somewhat suspect as pollsters will change methodologies or have a shift in biases over time.
In the end, polls are only as good as the questions being asked of the sample population of respondents, and trying to divine some general rule for the ‘accuracy’ of polls with respect to how far they are from the election is like trying to make a rope out of sand. At best, the best polls give a general indiciation of near-current popular sentiment, and even then only to the degree that the various demographics of the population are willing to honestly respond to those polling questions. It is a ‘science’ that is just slightly more advance than alchemy, and produces about as much actual worth.
Looking at one polling house at random July 2020 Emerson had it at Biden 50 to 46. Final was 51 to 47.
I can’t find Emerson same time but eyeballing aggregates July 2016 had Clinton up 3 or 4 ish. Popular vote was Clinton up 2.
Statistically speaking in Trump races the polling margins in July, before the ups and downs of conventions and events, were relatively not far off from the popular vote results in November. Depressingly.