Today 538 published our official forecast for the 2024 presidential election. The model builds on our general election polling averages by asking not just what our best guess is about who is leading the presidential race today, but what range of outcomes are possible for the actual election in November. At least once per day, we’ll rerun our simulations of the election with the latest data, so bookmark our interactive and check back often.
At launch, our forecast shows President Joe Biden locked in a practically tied race with former President Donald Trump, both in the Electoral College and national popular vote. Specifically, our model reckons Biden has a 53-in-100 chance of winning the election, meaning he wins in slightly more than half of our model’s simulations of how the election could unfold. However, Trump still has a 47-in-100 chance, so this election could still very much go either way. The range of realistic* Electoral College outcomes generated by our forecasting model stretches from 132 to 445 electoral votes for Biden — a testament to how much things could change by November (and how off the polls could be).
Our model is brand new this year, with tons of bells and whistles and modern statistical tools that you can read all about in our methodology post. Here, I’ll give you the non-wonky version of how the forecast works, offer a few tips on how to read it and explain why we think forecasts are valuable in the first place.
How we forecast
To forecast the election, we rely primarily on polls asking voters whom they support. However, our forecast also incorporates various economic and political indicators that aren’t related to polling but can be used to make rough predictions for the election. For example, we have calculated an index of economic growth and optimism on every day since 1944, gathered historical approval ratings for every president since Franklin D. Roosevelt and derived a formula for predicting state election outcomes using these and other local factors. We also tested whether incumbent presidents do better when they run for reelection (they do) and whether all of these factors are less predictive of voters’ choices when political polarization is high (they are).
Right now, Trump leads Biden in most polls of the swing states that will decide the election, but the “fundamentals” favor Biden. The combined polls-plus-fundamentals forecast splits the difference between these two viewpoints and arrives at an essentially deadlocked race. Here’s what it looks like on the state level:
At this point in the race, our margin of error for these state forecasts is huge. There are two reasons for this: First, it is early. As pollsters are bound to remind you many times between now and November, polls are snapshots of public opinion as it stands today, not predictions of vote share in the eventual election. To the extent they are predictions at all, they predict how people would vote if an election were held today — which, of course, it will not be.
In part, this oft-repeated caveat is a convenient way for pollsters to avoid catching flak for inaccurate numbers closer to the election. But there is an important truth to it: If a voter has not yet cast their ballot, there is the possibility they may change their mind. We also don’t know exactly who is going to turn out in this election yet. All this means polls earlier in the election cycle are worse at approximating the final margin.
This is where forecasting models really become useful. Above everything else, 538 makes forecasts to quantify the uncertainty inherent in the election. Our study of historical presidential election polls finds that the margin between the two candidates shifts by an average of 9 percentage points between June and November. In practical terms, that means today’s polls have a true margin of error of close to 20 points. And while recent elections have not had as much volatility, we can’t assume 2024 will be the same way; it’s possible that this year will be closer to the historical norm.
The second major source of error is the chance that polls systematically underestimate one of the candidates, as happened in the 2016 and 2020 presidential elections. We estimate that, even on Election Day, state-level polling averages of presidential general elections have an expected error of 4 points on the margin — meaning if the candidates are tied in the polling average, then on average we’d expect one to win by 4, and in rare cases they could win by 8!
Why forecast, anyway?
Having such wide margins of error is not our way of absolving ourselves of responsibility if the election result is surprising. It’s our way of giving you, the reader, a more informed understanding of the range of potential election outcomes than you’d get from a single poll (or even a polling average).
Over the last decade, it has become common to view election forecasting — and even polling — as purely making predictions of “what will happen” in the election. But we think forecasting models serve a greater journalistic purpose than a focus on prediction gives them credit for. For us here at 538, forecasting is an exercise in quantifying the reliability of various indicators of public opinion. Yes, that involves making predictions, but the real value of our work is the statistical analysis of the reliability of the numbers you are bound to see plastered all over print news media, social media and television over the next five months.
We think this is a different goal from making predictions for prediction’s sake, or making a model that can “call” every state correctly. If you want someone to give you a prediction of who will win the election with absolute certainty, then look elsewhere. (And buyer beware.)
Instead, we think we offer a unique product that can help you be smarter about the way you think about the range of outcomes for the election. As the stakes of our politics increase, a carefully calibrated sense of what could realistically happen in November — in our case, from a forecast that properly distinguishes between normal and tail risk — becomes increasingly valuable.
How to read the forecast
On that note, I’ll end with a few tips on how to read our forecast responsibly:
Watch the distributions. Our model simulates thousands of possible Electoral College outcomes based on the historical predictive error of the indicators we rely on. The top of our forecast page has a histogram of a random subset of these simulations, showing you which outcomes are likelier than others. We hope you get the impression that there is a wide potential range of outcomes, given all the error we’re talking about.
Unlikely does not mean impossible. In 2020, polls performed worse than in any election since 1980. The average state-level poll conducted in the last three weeks of that election overestimated Biden’s vote margin by 4.6 points — about 1.5 times the average 3-point bias for presidential elections since 1972. In a backtest of our current model, we would have assigned about a 20 percent chance to Biden winning 306 electoral votes (the number he actually won) or fewer in 2020. We think a similar miss this year would be statistically surprising, but a possibility people should mentally prepare for.
Changes in public opinion take time. We have done our best to make a model that reacts the appropriate amount to new polling data. “Appropriate” here means that the model will be conservative early on or when polls are bouncing generally around the same level, but also that it will be aggressive when polls appear to be moving uniformly across states — especially late in the campaign. However, as a properly Bayesian statistical model, the program that runs our forecast generates some amount of uncertainty about the parameters, resulting in unavoidable random error across our simulations. This means polling averages can change by a few decimal places day to day — and probabilities may jitter by around a point, which cascades down into uncertainty in our model. Don’t sweat these small changes; instead, pay attention to bigger changes in the model over longer stretches of time.
Use all the information you (reasonably) can. Polls are reasonably good predictors of election outcomes. In fact, asking people how they are going to vote is about the best single source of information you can get if your goal is to figure out how people might vote. But polls are not the only source of information available to us. 538’s forecast incorporates demographics, polls and the “fundamentals” all the way up to Election Day; our research has found this decreases the chance for uniform bias in our forecast.
Our forecast assumes normal election rules still apply. This is an important disclaimer about what our model is intended to do and what it is not. Because our model is trained on historical polling and election results, it is not intended to account for violations of normal political and election rules. We assume, for example, that if a voter legally casts a ballot, it will be counted accurately and fairly; that the electors a state elects to vote for a certain candidate in the Electoral College get to do so; that their votes are ultimately recognized by Congress; and that, as an extreme example, the election is administered on time, where officials say it will be administered and generally that people who show up to vote will be able to.
That is not to say that we dismiss the possibility of rule-breaking. From an editorial perspective, we stand ready to cover any attempts to undermine a free and fair election. But as a quantitative matter, our forecast is intended to explain variance in election outcomes based on the polls and other indicators, to serve as a supplement to polling averages and to put other political journalism in its proper context.
Footnotes
*Within the 95 percent confidence interval.
