The presidential election isn't a toss-up
As our model launches, either Biden or Trump could easily win — but the odds are in the ex-president’s favor.
It’s a big day here at Silver Bulletin: the launch of our presidential election forecast. We’ve1 just published three items at the same time — so let me give you the lay of the land.
First, there’s the output of the model itself — rendered in lots of charts, tables and data which you can see on the forecast landing page. You’ll always be able to find the latest numbers on this page, so that’s the one you’ll want to bookmark. In line with what I promised back in April, the polling averages will be available for everyone — but the forecast and probabilistic components of the model are for paid subscribers.
Second, there’s what follows after this introduction: a narrative story to help walk you through the current forecast. (Today’s story is rather long, I have to warn you.) We’ll run these about once per week through Labor Day, and then more frequently. They’ll usually take a paywall at some point — it’s hard to discuss the forecast without citing the probabilities — though today’s piece runs pretty deep before the paywall cuts in.
Finally, there’s a short methodology post. It’s short because the methodology is about 98 percent the same as 2020, so the post only details the handful of changes. In everything but name, this is the same model that I ran from 2008 through 2020 at FiveThirtyEight; I retained the IP to the models when I departed Disney last year.
The original plan was to update the forecast on the landing page only when we also published a narrative story. But now that the model is ready (!) and everything feels less daunting, we’re thinking we’ll probably be able to do better than that. Not right away — but give us a few weeks to automate our backend process, and we’ll see if we can’t run updates on the landing page a few times a week. We’ll also make the various charts prettier and better (especially for mobile users) and probably add more of them. Subscribers can also submit questions about the model for the monthly paid subscriber Q&A that we’ll run this weekend.
We hope all of this will prove to be worth your time — and money — and you can sign up using the link below.
The purpose of a model is to discipline your thinking
At the Manifest conference in Berkeley, California two weeks ago, I was asked by one of my favorite writers, Scott Alexander, about the odds in the presidential race. This conference was literally full of prediction markets nerds: exactly the people who appreciate that there is a meaningful, arbitrage-able difference between, say, a 50 percent chance of an event happening and a 60 percent chance — but who also understand that a 60 percent chance nevertheless implies a high degree of uncertainty.
So I went ahead and said what I thought: it was certainly a close race, but we’d reached the point where it would be dishonest to call it a toss-up. Instead, I guesstimated to Alexander that Trump had a 55 to 60 percent chance of prevailing.2
At the time, I’d barely begun work on getting the presidential forecast ready — the model that we’re officially launching today — not enough work to really have any sense for what it would say. And indeed, I’ve often found myself surprised by the headline numbers once the election model was finally ready for publication. In 2012, for instance — a race that has more parallels to this year than you might think — I remember telling people that I thought that race between Barack Obama and Mitt Romney was a toss-up. But when the model launched, Obama was actually a hair over a 60 percent favorite instead, with his odds steadily increasing as the election approached.
It’s good, though, when something like that happens. Since 2008, I’ve spent literally thousands of hours building election models. If the model always returned the same answer that I could have given you off the cuff, I’d be concerned that those hours had been wasted. Or — worse — that I’d essentially worked backward, tweaking the assumptions in the model until they matched my priors.
True, this is also where some of the “art” of model-building comes in. If that model had spat out something extremely counterintuitive — say, Obama with a 96 percent chance in June 2012 — I’d think it would have been reasonable for me to check for bugs in my code or flaws in my thought process. Indeed, it would have been foolish not to. It helps to have some domain knowledge and some horse sense for when you’ve done something wrong. It’s not always so straightforward as “letting the data speak for itself”.
But what a model ideally provides you with is a structured way of thinking through a challenging problem. Because our intuitions, even if they’re right, often lack any statistical precision.
There were various questions about 2024 that I didn’t really know the answer to until I started work on the model. For instance: what were the chances that Biden could win with exactly 270 electoral votes — the amount he’d win if he lost Arizona, Georgia and Nevada but held on to the other swing states that he won in 2020, including Michigan, Wisconsin and Pennsylvania? My intuition was that this was relatively unlikely — Biden was running too much risk of losing one of the three Rust Belt states, or having his plans foiled by some other state like Virginia or New Hampshire. And that turned out to be correct. Biden finishing with exactly 270 electoral votes and Trump with at exactly 268 is in fact the among the modal outcomes in the model — among the single most likely results if you had to guess any number from 0 to 538 — but there’s still only a 3 to 4 percent chance of it occurring, according to the forecast.
To complicate matters, it’s not as though there’s only one way to build a model or that the right way is obvious. At least not when we’re dealing with something like a presidential election, where there’s no way around the fact that we’re limited by small sample sizes. This is only the 20th presidential election since World War II, and only roughly the 10th or 12th with any sort of robust, state-by-state polling data.
Now, there are wrong ways to build a model. Some models in 2016 gave Hillary Clinton as little as a 1 percent chance of losing, for instance. Those models were wrong, not just “unlucky”. They were wrong because they made obviously wrong assumptions, especially that the outcomes in similar states such as Michigan and Wisconsin were relatively uncorrelated. These models would have been wrong even if Clinton had won, because that is an assumption you can validate empirically — there’s absolutely no doubt that similar states do move in similar directions from election to election.
But there’s not any one right way to build a model. With something like sports forecasting, at least the sample sizes are larger and the data is more robust. Even in Congressional forecasting, there are almost 500 relatively independent races each cycle between the House and the Senate3. So you can get a little closer to saying that Assumption A is definitely better than Assumption B, or that Conclusion P logically follows from Premise Q both in theory and in practice.
With presidential election forecasting, though — forget it. There are a lot of other models out there, which show Trump with anything from a 50 percent chance (538’s new model) to a 71 percent chance (the Economist) of winning. I have disagreements with those models — how couldn’t I, given the amount of time I’ve spent on this particular problem? But I don’t think any of them are unreasonable. It’s possible — in fact, rather likely, since there are several of them — that one of those models will prove to be better calibrated than mine over the long run. Although I do think it’s noteworthy that — self-aggrandizing aside ahead! — the Silver Bulletin model is the only one that has actually reached anything resembling the long run, with a strong out-of-sample track record over 16 years now.
So don’t let anyone tell you that their presidential model reveals some oracular truth, as though handed down on a stone tablet from God or GPT-7. If you get that vibe from someone, they’re either bullshitting you, or — more likely, honestly — bullshitting themselves.
Statistical modeling requires making a lot of choices. And election forecasting is a hard problem. Even choices as seemingly straightforward as how to calculate a “polling average” involve any number of parameters: which polls to include, how to weight them, how much to prioritize the most recent ones, and so on. This is even more of an issue once we start to consider “fundamental” factors such as incumbency or “the economy”. Even if we can agree that it’s a good idea to incorporate these things, there are basically infinite “researcher degrees of freedom” in exactly how you do it.
The human element looms large, in other words, even when it comes to statistical modeling. So let’s talk about it.
One issue is that I’m not only a statistician/analyst/pundit/journalist/degenerate gambler or whatever-the-fuck you want to call me — but also an American citizen who cares about political outcomes for a combination of selfish and altruistic reasons. Now that I’m on my own — not tied to some corporate behemoth like The New York Times or Disney — I feel freer to be transparent about my preferences. Although I have plenty of disagreements with progressives these days, my political values haven’t changed that much. I don’t want Trump to win the election, and I’d never consider voting for him.
It’s not my job to tell you how to vote, and I hope that we have some Trump (and RFK Jr., etc.) voters among the Silver Bulletin readership. Republicans buy sneakers — and sign up for Substack newsletters. But I think it’s important to be up front, because I’ve been rather lucky in one sense in my election forecasting career. I began making election forecasts in 2008, and in literally every presidential year since then, I haven’t really had to deal with a conflict between what I personally wanted to see happen and what my forecast said. This year, I do have that conflict. The candidate who I honest-to-God think has a better chance (Trump) isn’t the candidate I’d rather have win (Biden).
Wouldn’t it be suspicious if, in the first presidential cycle where the Democrat has consistently trailed in polls since 2004, I suddenly started telling you that you should trust vibes rather than polls? Or if I chucked out my heretofore well-performing model for a new one that had Biden favored — or at least had the election as a toss-up?
Yes, of course. It would be a sign that I’d become a hack. I’ve spent years telling people that, although polls are often wrong — indeed, inevitably wrong to some degree — it’s hard to predict the direction of polling error. Biden could easily overachieve his current polls — but it’s roughly as likely that he’ll underachieve them instead. It’s sort of a myth that Democrats outperformed their polls in 2022, but they certainly performed better than the conventional wisdom held. But Trump substantially outperformed his polls in 2016 and 2020. Going by the polls, perhaps along with some reasonable priors about things like the economy, is a lot better than going by the number of yard signs in your neighborhood or by what your friends think — or especially by what you hope will happen.
The other human factor is that the incentives are pretty clearly to tell you that the race is a toss-up. It’s hard to get in trouble that way! Even in 2008, when Obama had an overwhelming lead in the polls against John McCain, pundits were telling you it was a toss-up right up until the eve of the election.4
So this article last month — “Are We Too Bearish on Trump?” — from Sean Trende at RealClearPolitics really resonated with me:
But I started questioning whether I might actually be too bearish on [Trump] when I recently put together my first major presentation on the state of the race. As I got into it, I started asking myself “OK, what’s the good news for Biden here?” As I got further into it, I started asking, “If this were any other candidate pairing, would I really call this a tossup?” As I’ve done a few more presentations and panels, my self-questioning has only become more intense; there was very little pushback from more liberal participants on my analysis, or on the conclusion that this race was a tossup.
I’ve given my share of these presentations too.5 And I’d noticed something similar. The template that I’ve typically been using for these is to tick off six reasons why you might expect Biden to win, and six reasons why you might expect Trump. I’ve given one of these talks every month or two, and so this has been a way to be relatively honest with myself about how my thinking about the election has evolved.
And what I’d noticed over time is that the reasons that Trump would win have gradually become somewhat more compelling than the reasons for Biden. Emphasis on gradually and somewhat. Biden clearly could win in November. He won the same matchup four years ago. Not only would he be within a normal-sized polling error of Trump if the election were held today, but there are still four-and-a-half-months to go.
Still, the items on the “reasons to think Trump might win” checklist have proven to be more robust. There’s Biden’s age, which voters have extremely persistent concerns about. There’s the very high inflation of mid-2021 through mid-2023 — which has considerably abated, but still is reflected in much higher prices than when Biden took office. There’s the fact that the global mood is pessimistic and that incumbents have been getting crushed everywhere around the world. Plus, some of the factors I thought would be an advantage for Biden haven’t proven to be. There’s less of a fundraising gap than I expected, for instance, and I’m not sure that Biden has run the smarter tactical campaign.
What’s happening this year isn’t quite the “unthinkability bias” that Trende wrote about eight years ago after Brexit. Trump was already president for four years, so it’s hardly unthinkable. And I don’t know many Democrats who deny that he could become president again; instead, most are very worried. But you’ll sometimes hear exasperated comments along the lines of how can Biden possibly be losing? — not as a way of questioning the polls so much as questioning the wisdom of the electorate. (Or the media.)
Well, here’s how. Biden has the lowest approval ratings of anyone running for re-election since either George H.W. Bush or Jimmy Carter, depending on how you squint at the numbers. The reasons he might lose are overdetermined. (I haven’t even mentioned things like immigration or the war in Gaza.) Would it really be surprising if he’s become an underdog, even if it’s against a candidate as flawed as Trump?
Biden is losing in the Electoral College now
When the model was finally done on Sunday night, it turned out that Trump was favored by a slightly larger degree than I’d anticipated at Manifest — although Biden retains highly viable paths to victory.
There’s a certain glass-half-full view of the race where the election really is a toss-up for Biden. The case rests upon some combination of these assumptions:
The notion that an incumbent usually wins in a decent economy;
That the polls in Pennsylvania, Michigan and Wisconsin show a tossup, and therefore the Electoral College is a tossup.
Let’s take the second of those claims first. Here’s a chart showing our polling averages nationally and in the key states. (For more detail on these, including how they’ve evolved over time6, please go to the landing page.)
Although these polling averages might not want to admit it, they’re kind of fancy, making adjustments for factors like likely versus registered voters (this helps Biden), and the presence or absence of RFK Jr. in polls (this doesn’t have much effect, somewhat contrary to the conventional wisdom). And they make inferences from state polls to national polls and vice versa, which tend to make them more stable.
The news isn’t all bad for Biden. We do have him trending upward relative to a month ago. Biden probably has been helped by Trump’s conviction on felony charges related to hush money payments to the porn star Stormy Daniels — what sort of world are we living in that I can write a phrase like that about a presidential election?
It would be easy to overstate the case, however. Trump does still lead in our national average — however narrowly. But the bigger problem for Biden though is that elections in the United States aren’t determined by the popular vote. His current popular-vote disadvantage is modest — modest enough that a couple more polls like the recent Fox News national poll could be enough to put him ahead. And the fundamentals part of our model — which in the case of the Silver Bulletin, just means the economy and incumbency — slightly helps Biden, as I’ll cover in the next section.
But I don’t think the glass-half-full view is quite right:
This is the most important complication: if Biden loses Georgia, Arizona and Nevada — and he trails badly in each — he’ll need to win all three of Michigan, Wisconsin and Pennsylvania and not just one of them. Even though these states are pretty heavily correlated, they aren’t perfectly correlated. Winning several different correlated bets is still hard. (Just ask the DraftKings how it makes so much money on same-game parlays.) And these states do have some differences with one another: Wisconsin is more rural, for instance, and Michigan has the largest Arab/Muslim/Palestinian population. In our simulations, Biden wins at least one of these states 54 percent of the time. But he wins all three of them in only 32 percent of simulations. This is the sort of precision that a model can provide that your intuition really can’t.
Biden also has to hold on to states like New Hampshire and Virginia (and New Mexico and Minnesota and Nebraska’s 2nd Congressional District) in the Blue Wall-saves-the-day scenario, and that’s not quite a sure thing given Biden’s mediocre polling elsewhere in the Northeast. This is less of a concern for him, though — conditional on winning Wisconsin and Michigan and Pennsylvania, Biden wins the Electoral College about 97 percent of the time in our simulations.
In those Sunbelt swing states — meaning Georgia, Arizona and Nevada, where he won last time, and North Carolina, Florida and Texas, where he didn’t — Biden trails badly. Not so badly that he can’t win — the model gives him a 22 percent chance of eventually winning Arizona, for instance. But badly enough that he’s borderline out of the “normal range of polling error” scenarios and either needs a big polling error or for something fundamental to change about the race — even though it’s been exceptionally steady so far. Either of those things are possible — just not the outcome you’d bet on at even money. For now, he has nothing in the way of a reliable backup plan if any of the Rust Belt trio falls to Trump. It’s just not good for Biden that he trails by mid-single digits in high-quality poll after high-quality poll of states like Georgia and Nevada.
And there’s also a credible glass-half-empty view of Biden’s Rust Belt polling. In 2020, Michigan was 2.7 points to the right of the national average, Pennsylvania was 3.3 points to the right, and Wisconsin was 3.8 points to the right. So if you buy that Biden is down by a point or something nationally — or even if you think that that the race is tied or Biden is a half-point ahead — you’d expect him to be narrowly but meaningfully behind in these Rust Belt states under a uniform swing.
The model accounts for this possibility — it blends the polling averages in each state with a series of regression-based estimates7 that infer what the polling “should” be there based on polls in other states and the national numbers. Some of these techniques get slightly fancy, but the one that has historically been most reliable is indeed to assume some sort of uniform swing from the previous election. And that’s a nasty assumption for Biden, since he’s doing a net of about 5 points worse than in 2020, when he won the popular vote by four-and-a-half points.
In other words, if the Electoral College/popular vote gap looks anything like it did in 2016 or 2020, you’d expect Biden to be in deep trouble if the popular vote is roughly tied. So if we’re being honest, pundits who obsess over whether Biden is 1 point ahead or behind in national polls are kind of missing the point. Because national polls being tied don’t make for a toss-up race — but instead one where Trump has a material advantage in the Electoral College. In fact, the race looks a lot like 2012 in reverse, when national polls were often close but the swing state polls consistently favored Obama and gave him the far more robust map.
The “fundamentals” help Biden, but only so much
So far, I’ve been taking a “polls only” view of the race. But the version of the model that we’re publishing at Silver Bulletin is what we used to call “polls-plus” at FiveThirtyEight, which also incorporated some assumptions about incumbency and the economy. (We phased out the polls-only version and the dreaded nowcast in 2020.)
This version is somewhat better for Biden than a polls-only forecast would be, but it would be easy to get carried away. One problem is that “fundamentals” models often reflect a large degree of overfitting where the results do not replicate very well out of sample. I could go on for paragraphs about this, but what I’ll say for now is that I think data scientists often talk themselves into thinking they’re avoiding overfitting by using some fancy statistical techniques when they’re actually making it worse.
In practical terms, the upshot is that you probably ought to be cautious about applying too much weight to macro factors when you have so many ways to account for them but such a small sample size of relevant elections. And the Silver Bulletin model is cautious about them — the fundamentals currently account for about 30 percent of the forecast, as compared to 70 percent for the polls, and this will gradually fade to zero by Election Day.
Nor are the fundamentals that great for Biden. To be more precise, they currently project him to win the popular vote by 2 to 3 percentage points. That would still only make for roughly a toss-up race, however, given Biden’s problems in the Electoral College.
But don’t incumbents usually win by more than that when the economy is good? Well, let’s parse that assumption, too.
One problem is that the economy isn’t all that good. Real GDP grew by 2.5 percent last year, following 1.9 percent in 2022. It grew at an annualized rate of 1.3 percent in the first quarter of 2024, though it’s expected to return to trend for the rest of the year. Two-ish percent economic growth is fine — and may reflect the new normal — but it’s below the long-term median growth rate in the US.
The Silver Bulletin economic index does not actually use GDP, which is only updated quarterly. (Well, it doesn’t use actual GDP. It does use forecasted GDP from the Survey of Professional Forecasters to help predict changes in the other indicators.) But the indicators it does use are a mixed bag for Biden:
These z-scores are evaluated from the standpoint of the incumbent party — higher is better for Biden (and for the economy). At this point, the period of very high inflation is enough in the rearview mirror to not really hurt Biden in the model. However, Biden is harmed by sluggish growth in real take-home incomes, historically among the best economic variables at predicting election outcomes
Yes, the labor market is tight and that’s great for many types of workers. But it’s just kind of BS to imply that voter concerns about the economy reflect misinformation — when throughout Biden’s term, the average American has struggled to see enough gains in her paycheck to keep up with inflation. And if anything, the average conceals a lot of differences in how voters have experienced the economy. (If you’re on some sort of fixed income, inflation has made this a rough few years, for instance.) The manufacturing numbers are also just average, and consumer spending has cooled off a bit. Ironically — given that it’s something Democrats are usually reluctant to talk about — Biden is buoyed in the economic index by the stock market. But overall, the economy is just average.
What about the fact that Biden is an incumbent? Indeed, not just an incumbent, but an elected incumbent (as opposed to someone like Gerald Ford who was appointed to the office) — and an incumbent who beat the exact same opponent last time? This is one area where I did double-check some assumptions — there’s now a slightly more complicated formula for evaluating the effects of incumbency that accounts for greater persistence in repeat matchups like this one, and also for the incumbent party’s vote share in the previous election. But it makes essentially no difference. Biden gets 51.4 percent of the projected two-party vote (that is, excluding third parties) in the revised formula instead of 51.2 percent as under the original one.
That’s not much higher than 50/50, which is the point. The high incumbent margins in the post-World War II era — like Ronald Reagan winning re-election by 18 points — now look like as much of an exception as a rule, an artifact of an era of unusually low polarization. In times of high polarization, like today, or in the early 20th century — think of Grover Cleveland winning or losing by a point or two for several elections in a row — the incumbency advantage has been negligible. (The model is actually calibrated based on elections dating back to 1880.) And although the model doesn’t consider the poor performance of incumbents elsewhere around the world, that does provide some further validation to its story that incumbency is only a minor advantage in this sort of political environment.
For all that said, Biden is favored by the fundamentals in the popular vote — just not by all that much. If he’s “supposed” to win by 2.5 points and instead he’s trailing by 0.5 points, that isn’t that big a gap to parse. For instance, it could easily be explained by voters having more lingering concerns about inflation than our economic index implies — or concerns about Biden’s age. If roughly 1 out of every 70 voters who would otherwise vote for Biden won’t do so because of his age, for instance (and instead vote for Trump), that would be enough to fill the gap.
Of course, you could also argue for subjective adjustments that go the other way, like for Trump’s criminal convictions. I just don’t think it’s so obvious that there are strong gravitational forces pulling in Biden’s direction. In a time of extremely high polarization, elections tend toward being 50/50 affairs, and it’s a challenge to win the 50/50 races when you’re at a disadvantage in the Electoral College.
C’mon, Nate. What does the model say? And what do you really think?
OK, then. It’s finally time to step behind the velvet rope of the paywall so that we can have an adult conversation.