Model Update: Harris Is Surging…So Why Is It Still a Tossup?
Hello! How’s it going? It’s been a while!
My last blog post was on July 21st, the day Joe Biden announced he was bowing out of the Presidential race after receiving intense pressure to step aside from the media and the brain trust within his own party. At that time, I had decided to put a pause on front-facing model updates until at least August 5th, to allow more data to come in for the new matchup between Donald Trump and Vice President Kamala Harris. While many were still thinking there could be an open convention, it felt inevitable to me that everyone would realize very quickly there was only one plausible alternative to Biden. Sure enough, it only took a day or two for Harris to become the de facto nominee, and her poll numbers have been taking off like a rocket ship ever since. With the ensuing polling collapse of Robert Kennedy Jr. and surging favorables for Harris and her VP pick, Governor Tim Walz of Minnesota, it’s plausible that momentum could continue.
(Quick aside: Thank God we aren’t having an open convention. See how nasty Veepstakes week got when it looked like Josh Shapiro might have been the pick? Democrats would have torn themselves apart before likely ending up with Harris anyway.)
“But Zach…the forecast says Donald Trump is still a slight favorite to win the election. The polling averages have Harris up 2-3 points, but you’re showing the popular vote as a virtual tie. What the fuck, bro?”
I can explain that! But before I do, I want to talk briefly about how FiveThirtyEight, a website that receives more pageviews in an hour than Solid Purple will this entire election cycle, still hasn’t turned their forecast back on. Many are gossiping, including and especially the site’s former steward Nate Silver, that the reason it is still down is because their model is profoundly flawed and continuing to produce results that buck the conventional wisdom. Maybe that’s true, maybe it’s not, but to be perfectly honest with you, I think continuing to stall is the right call either way.
Solid Purple quietly relaunched a week ago for one reason and one reason only: I am competing for attention. Very few people on Election Twitter or in national political media know about this forecast. Even fewer give a shit. Since I am trying to attract as many eyeballs as possible, the moment the model had enough input data to limit the amount of uncertainty to an acceptable level, we were back in business. In order to make that happen, I decided to include results from surveys taken prior to 7/21, back when the race was purely hypothetical and Trump was consistently well ahead. Some of those polls continue to hold substantial weight in the back-end aggregator - enough to be dragging Harris’ popular vote margin down by about a point and a half compared to polls taken solely after she unofficially officially became the candidate.
I know I keep promising to go long on the full methodology of Solid Purple and how it differs from the approaches FiveThirtyEight, Nate Silver, Decision Desk, etc. take to produce their forecasts. I’ll get around to it, I swear. In the meantime, I will say that a fundamental assumption I made when putting the forecast together is that the national environment would be relatively stable, not just in that the candidates would stay the same, but that a poll from January would be reasonably predictive of what a poll in August would say. The model assumes the past is prologue, and is therefore very small-C conservative is how it interprets a scenario where Trump suddenly goes from being up 6-8 points nationally in a New York Times/Siena College poll to losing by 4 in the same firm’s surveys of Pennsylvania, Michigan, and Wisconsin. Unless such a rapid, sudden shift holds for a sustained period of time, the forecast will hold the line.
Additionally, Solid Purple is heavily reliant on national surveys relative to state polls until about 4-6 weeks before the election, when it becomes safer to expect the data is robust enough to pick up on geographic trends that are distinct from demographic trends nationwide. There are more sophisticated ways to account for this, some of which go way over my head, but depending on how individual forecasters calibrated their models, it may be crucial for there to be a lot of high-quality state-level data for their forecasts to work as intended. While we’ve had a fair amount of national polling the last few weeks, state polling has been pretty dire, especially outside of the six key battlegrounds. We’ve barely heard a peep out of Texas or Virginia. How are things going in Ohio or Minnesota? Are Trump and Harris receiving bumps from their respective VP selections? We don’t really know, and as much as you can rely on how states correlate with each other to fill that gap, extrapolation can very quickly become distortion.
All this is to say: any forecast of the November election is kinda bullshit right now. There isn’t enough data, there hasn’t been enough time, and because of that, my model is essentially reverting to 50/50. It does not have a firm grasp on what’s going to happen because it was not designed for the scenario it is attempting to interpret. If you decided to stop paying attention until a week after Labor Day, I wouldn’t blame you. In fact, that is what I would probably advise you to do if I were not a self-interested party. Instead, I would suggest you click through every page on the site, meticulously comb the maps and tables for insights into each House and Senate contest we are trying to predict, and please, for the love of God, click on the truly awful advertisements Google injects into each page. I get a quarter every time one of you accidentally hits on the one that says hypersexuality is a symptom of ADHD or whatever.
