There’s no “normal” in this election
Why Harris is leading in our polling averages — but not in our forecast. And why it’s hard to know where we’re headed next.
Beginning this week, we’ll run two Model Talk columns per week for paid subscribers instead of one. So I want to come out of Labor Day weekend strong with a Model Talk — but today’s newsletter is really trying to accomplish a few different things, so let me give you a quick table of contents:
First, I’ll reiterate the difference between our polling averages or snapshot of the current state of race and our forecast of the November result, which have begun to differ since the Democratic convention but will eventually come back into line;
Second, I’ll explain why it’s hard to know the “real” state of the race — what default conditions are when there isn’t a bunch of crazy news happening;
And third, I’ll work through what will happen if Kamala Harris holds her current position in polls. Hint: she’ll eventually become the favorite again. This last part is paywalled.
Before any of that, though, a quick note on how we sometimes see our work being cited in the media. The polling averages are free for everyone, and we’re always happy to see them referenced in the press. That includes readers who take screenshots and post them on Twitter or other platforms. We’d just ask for two common courtesies: please don’t cut out the Silver Bulletin watermark in the bottom-right corner of the chart, and please try to link back to the model landing page or otherwise provide appropriate attribution.
By contrast, the Silver Bulletin forecast is paywalled. We’re not inclined to police occasional “leaks” of this content on free platforms like Twitter — and indeed, we sometimes provide teases there or in the newsletter ourselves. But these should fall within the spirit of fair use guidelines; you shouldn’t be sharing every paywalled chart every day, for instance.
Our concern is less about cannibalizing content that could induce people to become paid subscribers and more that these charts can get taken out of context. For instance, we’ve seen cases of people using this chart—
—to imply that we have Donald Trump forecasted to win the election by 10 or 11 points. The biggest landslide since Reagan in 1984! But that’s not at all what this chart is saying. Rather, it displays the probability that Trump will win the Electoral College — which is about 55 percent in our forecast, compared to about 45 percent for Harris.
Rather than a landslide, that implies an extremely uncertain and probably ultimately very close race — well in the range of what we’d consider a “toss-up”. In fact, the chance of a landslide — either Harris or Trump winning the popular vote by 10 points or more — is only about 5 percent in our model.
Why Harris is ahead in our polling averages but not in our forecast
As you can see from the forecast chart, Harris’s odds have declined slightly over the past two weeks, as she’s gone from roughly a 55/45 favorite to a 45/55 underdog. It’s not a huge change. Probability calculations can be highly sensitive just to either side of the 50/50 mark. If the New York Knicks make a buzzer-beater just before halftime to go from trailing the Boston Celtics 61-60 to leading them 62-61, they might tick over from “underdog” to “favorite” in a win probability model. But it isn’t as though the game has been fundamentally transformed. Still, the decline in Harris’s forecast reflects three factors:
(1) Harris is slightly underperforming the model’s benchmark for a convention bounce. Harris is, in fact, polling a bit better now than before the DNC — but only a bit better, with a 3.5-point lead in our national polling average as of Sunday versus 2.3 points before the convention. The model’s baseline expectation was a bounce of more like 2 points. By the model’s logic, she’s gone from a lead of 2.3 points to a convention-bounce adjusted lead of 1.5 points. That’s not a game-changing difference, but it’s enough to show up in the bottom line.
(2) Kennedy dropping out of the race. We initially expected this to hurt Harris by 0.5 points or less, given that RFK Jr. drew more Trump voters than Harris voters but only slightly more. However, it’s plausible that the impact is larger with RFK having not just dropped out but endorsed Trump.
Given the timing of Kennedy’s announcement, this factor is all but impossible to disentangle from the convention bounce or lack thereof. Our model run on Friday, August 23 — the day just after Harris’s acceptance speech and the day that Kennedy dropped out, but before we switched over to the RFK-less version of the model — showed Harris ahead by 4.3 points in our national average. That suggested she was on her way to a typical convention bump of 2 or 2.5 points — or possibly more, given that the impact of the convention probably hadn’t yet been fully realized in the polling.
Now, our polling averages are designed to be very aggressive after big events like conventions, and maybe 4.3 points was an overestimate since it was drawn from relatively few polls. Occam’s Razor, though, is that Harris — who gave an effective speech — was on her way to a typical but not extraordinary convention bounce, and then Kennedy’s dropout/endorsement ate into those gains. I somewhat regret the framing of my story from Aug. 24, which warned that the model could be running a “little hot” on Harris because the impact of RFK hadn’t really been factored in yet, but had a headline that emphasized how there hadn’t been much change yet. If I had to do it over again, I’d instead headline the story with something that underscored the need for a wait-and-see approach.
(3) Comparatively poor polling for Harris in Pennsylvania, which is disproportionately important given Pennsylvania’s likelihood of being the pivotal state. As a result, the Electoral College forecast has swung more than the popular vote forecast.
The race has never been in a “steady state” since the Biden debate
What’s made this race uniquely challenging to forecast is that there hasn’t really been a slow news cycle since the debate on June 28. In rapid succession, we had: the debate, an incredible pressure campaign by Democrats to get Biden to drop out, the assassination attempt against Trump, Trump naming JD Vance as his running mate, the Republican convention, Biden dropping out, Harris securing the nomination overnight, Harris naming Tim Walz as her running mate, the Democratic convention, and then Kennedy dropping out.
In principle, you could say something like: let’s look at the polls from back when things were “normal”, ignore what they say immediately after one of these “crazy” events, and then wait for them to settle down again. (In fact, as I’ll explain below, the model sort of attempts to do this.) But it’s hard to know what counts as “normal” in this election:
You could say that now is “normal”, which would be a good answer for Democrats since normal means Harris leading. And that might be reasonable. The news cycle was quite slow last week heading into the holiday weekend, and there’s no longer much press coverage about Harris’s favorable momentum and so forth. Maybe we’re already out of the convention bounce window.
You could say that the period just before the DNC began was “normal” — Harris had been the presumptive nominee for several weeks at that point. This is a decent answer for Democrats, as Harris was slightly ahead in our forecast. However, she may also have been benefiting from some bounce-type dynamics then. The vibes had been extremely good for Harris, and Trump had been caught flat-footed by her entry into the race. This may not have been sustainable.
You could say that “normal” is about when we relaunched the model on July 30, when Harris had been the presumptive nominee for a week but before her bounce/buzz had really built up. But this is not a great answer for Democrats. At that point, national polls were roughly tied, and Harris was a slight underdog in the Electoral College.
You could say that “normal” was before the Biden-Trump debate, at which point Biden was about a 2:1 underdog — and Harris’s polling wasn’t any better than Biden’s. Indeed, Harris has been unpopular for most of her tenure as vice president until just recently. If her favorability numbers revert to their long-term average, that probably means Trump back in the White House.
Or you could ignore the polls entirely and calculate “normal” based on other factors — what we call “the fundamentals” around here. Our model’s answer, based on the economy being about average and there being no true incumbent in the race, is that “normal” means a tie in the popular vote but Democrats being underdogs in the Electoral College.
Finally, although this isn’t what our model does, you could calculate “normal” based on the fact that Democrats have had an edge in the popular vote in most recent elections. On average between 2016 and 2020, Democrats won the popular vote by 3.3 points; on average since 2000, they’ve won it by 2.3 points. That might support the notion that something roughly like Harris’s 3.5-point lead is sustainable, or maybe we’d expect a decline of a point or so.
Any and all of these answers are defensible, I think. And if you take a mental average of them, they work out to Harris being a favorite to win the popular vote but perhaps a slight underdog to win the White House because of the GOP advantage in the Electoral College. And that’s basically what our model shows right now.
But look, if you have a different theory of the case, I’m not inclined to get into too much of a huff about it. We’ll know more in a week or so — although then we’ll have yet another disruptive event, the first (and possibly only) Harris-Trump debate on Tuesday.
What if Harris holds her current numbers?
But let’s say that something roughly like the first version of “normal” is correct and Harris is able to maintain the status quo in the polls. In other words, every poll from now through November exactly matches our current polling averages: every national poll has Harris up 3.5 points, every Michigan poll has her up 1.9, every North Carolina poll has her down 0.3 points, and so forth:
What would happen then?