In the final days before the 2020 presidential election, polls generally pointed to a clear victory for now-President Joe Biden. But when the votes were counted, it turned out the polls had overestimated him — Biden won, but by the skin of his teeth. That, of course, was similar to what happened in 2016, when former President Donald Trump significantly outperformed the polls in Michigan, Pennsylvania and Wisconsin to win a surprise Electoral College victory despite relatively accurate national polls.
This raises two questions for 2024: First, what would happen if the polls are off again? And second, how likely is it that the polls will be off by as much as they were in 2016 or 2020?
In 2020, polls overestimated Biden's margin over Trump by about 4 percentage points in competitive states. As of Oct. 30 at 11:30 a.m. Eastern, the margin between Vice President Kamala Harris and Trump in 538's polling averages is smaller than 4 points in seven states: the familiar septet of Arizona, Georgia, Michigan, Nevada, North Carolina, Pennsylvania and Wisconsin. That means that, if the polling error from 2020 repeats itself, Trump would win all seven swing states and 312 Electoral College votes.
Of course, if the polls are off, it won't necessarily benefit Trump. The direction of polling error is impossible to predict in advance, and polls have overestimated Republicans plenty of times in the past. In a scenario where the polls overestimate Trump's margin by 4 points in every state, Harris would win all seven swing states and 319 electoral votes.
Note that neither of these are particularly close outcomes, at least in the Electoral College.
Both of these outcomes — and everything in between — are very much on the table next week. But are these scenarios actually likely, or more like outside possibilities? Well, that's where the work we do for our election forecasting model can be helpful. As of Oct. 30 at 11:30 a.m. Eastern, our forecast gives Trump a 52-in-100 chance to win the White House and Harris a near-identical 48-in-100 chance. The model arrives at that probability by calculating how many Electoral College votes each candidate would win given certain amounts of polling error in their favor, and then counting up how many times each candidate wins among these simulations. (More about that in our methodology.)
Based on how much polls have been off in the past, our election model estimates that the average polling error in competitive states this year will be 3.8 points on the margin.* This error is not uniform across states — for example, states with different demographics tend to have different levels of polling error — but, generally speaking, when polls overestimate a candidate, they tend to overestimate them across the board. In other words, the model is expecting a roughly 2020-sized polling error — although not necessarily in the same direction as 2020. (In 50 percent of the model's simulations, Trump beats his polls, and 50 percent of the time, Harris does.)
This point is worth dwelling on. Because our average expectation is for there to be a decently large polling error at least half of the time, there is actually a very low probability that the polls are perfect and the election plays out exactly how the polls suggest. Let's look at this using the largest lead either candidate has in the seven swing states: Trump's current 2-point lead in Arizona. Nationwide, our model expects polling error to be greater than 2 points in either direction 62 percent of the time. In other words, there's only about a 1-in-3 chance that polls miss by less than 2 points (which we would consider a small polling error historically).
Given that all seven key swing states are so close, even small polling errors in the same direction can have a big impact on who wins the election. According to the simulations from our model, there is a 60-in-100 chance either candidate wins over 300 Electoral College votes — which Harris could do by winning five of the seven swing states and Trump six out of the seven. By modern standards, I think it's fair to consider this a blowout win — given how closely divided the country is, it's relatively unlikely for either candidate to win much more than this (even to get to 320 electoral votes, Trump would have to win a state like Minnesota and Harris would have to win a state like Florida).
Of course, the probability of a blowout either way depends heavily on the popular vote outcome. If Harris wins the national popular vote by 3 points, she's much likelier to win the states that will decide the Electoral College than if she loses the popular vote by 3. This is on vivid display in the chart below, which takes all the simulations from our model and buckets them by popular vote outcome.
As you can see, Trump is favored to win the election even if he loses the popular vote by 1-2 points, which is what our national polling average currently suggests. And if the national polls turn out to be underestimating him, with Trump winning the popular vote by 1-2 points, he would be favored to win in a blowout.
Meanwhile, our model reckons Harris needs to win the popular vote by 2.1 points to be favored to win the election because swing states are more Republican-leaning than the nation as a whole. And if she wins the popular vote by 4.5 points (Biden's popular-vote margin in 2020), she is favored to win in a blowout of her own.
So far, we have said little about the actual polls themselves. And there actually is some reason to believe polls may be more accurate this year than in the past. While the share of polls conducted or sponsored by Republican-aligned organizations is up — something we have written about — the overall share of partisan polls is lower than in previous years, and the average 538 pollster rating of 2024's polls is higher than past years'. All else being equal, that should make for better polls than in 2016 and 2020. We have also seen fewer polls from the firms that overestimated Democrats the most in those years.
However, the news is not all good. In particular, pollsters are still reporting difficulty reaching voters at all, and Trump supporters may still be less likely to respond to polls — even high-quality ones. This means that pollsters are as reliant as ever (or maybe more!) on weighting and modeling to get good estimates of public opinion. But the decisions they make matter a lot, and in particular, there appear to be large differences between polls that try to use these techniques to balance their samples by party or past vote and those that do not.
And this is the big, fundamental problem with preelection polling: We don't know what the demographic and political composition of the actual electorate will be, so pollsters are just making the best guesses they can. Those guesses have always, and will always, come with error attached to them.
And that's where election models like 538's become really helpful. The point of creating election forecasting models isn't to provide a hyper-accurate, laser-like predictive picture of the election that removes all error from the polls. Rather, it's to give people a good understanding of how the polls could be wrong and what would happen if they are. By analyzing possible errors and uncertainties in the polls, these models help us approach the election with a clearer sense of how likely each side is to win (and by how much).
As we enter the final week of this election, it is a good time to remember that uncertainty is an inherent part of polling and elections. That is especially true this year, with deadlocked races across the swing states. Given that polls are imperfect, our expectation is for them to be off by some amount in either direction. And if the polls do end up being off, given the closeness of the election, there is a rather wide spread in the range of Electoral College outcomes.
In other words, we can sum up the current state of the race like this: Although Trump and Harris have roughly equal chances of winning the election, the final margin is not necessarily going to be close. In fact, there's a pretty high probability that it won't be.
*We simulate potential polling errors for future elections using a fat-tailed distribution — specifically, a Student's t distribution with five degrees of freedom (a parameter that increases or decreases the likelihood of surprise "tail" events in our simulations). This 3.8-point error is the spread, or sigma, of that distribution — analogous to the standard deviation of a normal distribution. 538's distributions are slightly wider than the ones used by other forecasting models. This is because our model accounts for the fact that polling misses have gotten bigger over the last decade. Therefore, our model anticipates more polling error than it would if we assumed a constant level of error over time, like most other forecasts do.