The model exactly predicted the most likely election map

The actual map was the most common one in our 80,000 simulations! Even so, it contained some revealing surprises.

Before we get into the meat of today’s newsletter, just a quick update on our plans. Silver Bulletin, we hope, will have a long future. But I’m thinking about things in a few different stages:

• Short term: We’re treating next week (and the remainder of this week) as “election wind-down mode.” You’ll almost certainly get a few more newsletters covering what happened and why. But frankly wind-down mode also means catching up on sleep, exercise, and seeing friends I haven’t gotten to hang out with much lately … so I’m treating everything as kind of optional. Because we tilted heavily toward paid content toward the end of the race, everything in this period will be free (including today’s Model Talk¹) — except for a bonus Subscriber Questions post we’ll run in a week or two with any outstanding questions you had about the election, the polls or the model now that we’ve seen the results. (You can submit your questions in the SBSQ #14 comments section.)

• Medium-term: Starting the week of Nov. 17, this newsletter will be more like it was before the election, with a greater breadth of topics. So, yes, polls and elections. But also sports, essays about political trends, media analysis and criticism — and then a lighter dose of poker and gambling, economics, markets, AI and tech, plus occasionally something unexpected. The plan is to have one paid post and one free post a week, in additional to SBSQ² (almost always paid) each chosen from a cross-section of these topics. Usually, I wind up adding three or four items over a typical month based on breaking news or sudden bursts of inspiration, but then having some one-post weeks because of vacation or other work. So, let’s call it 9-12 posts a month.

• Long-term:  Meanwhile, I’ll be starting work to reboot and improve some of the sports models. I’ll probably begin with college basketball Elo ratings just because those are pretty straightforward, with a goal of launching by the end of the year, if not sooner. But beyond that, I’ll have to start making some decisions about staffing and long-term plans. And I am not even ready to think about that yet … other than to say Eli will be a part of them so long as he wants to be, we had a really good year, we really appreciate your business, and even though there’s going to be a lot of post-election subscriber churn, we plan to reinvest in Silver Bulletin for the long term. The question is at what scale and at what pace. The answers are probably “modest” and “slow-ish”, but I’ve got to think about that. 

Here’s how to subscribe:

And here’s today’s Model Talk.

Tuesday was a perfect demonstration of correlated polling error

From the admittedly narrow standpoint of what would make the model look smart³, I’m not sure exactly what outcome I was rooting for on Tuesday night. The exact Blue Wall map where Harris won the Electoral College 270-268 via Michigan, Wisconsin and Pennsylvania would probably have been a good one, demonstrating how close the polls had been — although honestly, I think I’d still have taken some crap from Democrats for any Harris win. (I have a mini-rant about this, but it’s somewhat of a diversion so I’m consigning it to the footnotes.⁴)

A map where the model “called” every state correctly would certainly have looked wise, but this year it had an interesting twist. If you treated the model’s forecast in every state as a deterministic forecast — which it very much isn’t — the map you get is this one:

That is, Trump winning 5 of the 7 swing states: Michigan and Wisconsin are highlighted in a lighter shade of blue because they are the two we missed, although not by much (based on current vote counts, Trump leads by 1.5 points in Michigan and 0.9 in Wisconsin, which was actually the leftmost of the Blue Wall states this year). The other 48 of 50 “calls” will be right, assuming that Trump’s leads hold in Arizona and Nevada, which they almost certainly will.

The twist is that, if you’re treat our predictions as deterministic, then we also have to own the fact that Kamala Harris was technically the favorite in our final forecast, even though it was by such a narrow margin — 50.015 percent of simulations — that it was literally more random than a coin flip and within the model’s margin of sampling error. So no matter what, actually, we were going to be wrong about something! Either our Electoral College “call” would be wrong, or we’d been wrong about our “call” of Pennsylvania for Trump.⁵

But the real value-add of the model is not just in calculating who’s ahead in the polling average. Rather, it’s in understanding the uncertainties in the data: how accurate polls are in practice, and how these errors are correlated between the states. The final margins on Tuesday were actually quite close to the polling averages in the swing states, though less so in blue states, as I’ll discuss in a moment. But this was more or less a textbook illustration of the normal-sized polling error that we frequently wrote about. When polls miss low on Trump in one key state, they probably also will in most or all of the others.

In fact, because polling errors are highly correlated between states — and because Trump was ahead in 5 of the 7 swing states anyway — a Trump sweep of the swing states was actually our most common scenario, occurring in 20 percent of simulations. Following the same logic, the second most common outcome, happening 14 percent of the time, was a Harris swing state sweep.⁶

Relatedly, the final Electoral College tally will be 312 electoral votes for Trump and 226 for Harris. And Trump @ 312 was by far the most common outcome in our simulations, occurring 6 percent of the time. In fact, Trump 312/Harris 226 is the huge spike you see in our electoral vote distribution chart:

The difference between 20 percent (the share of times Trump won all 7 swing states) and 6 percent (his getting exactly 312 electoral votes) is because sometimes, Trump winning all the swing states was part of a complete landslide where he penetrated further into blue territory. Conditional on winning all 7 swing states, for instance, Trump had a 22 percent chance of also winning New Mexico, a 21 percent chance at Minnesota, 19 percent in New Hampshire, 16 percent in Maine, 11 percent in Nebraska’s 2nd Congressional District, and 10 percent in Virginia. Trump won more than 312 electoral votes in 16 percent of our simulations.

But on Tuesday, there weren’t any upsets in the other states. So not only did Trump win with exactly 312 electoral votes, he also won with the exact map that occurred most often in our simulations, counting all 50 states, the District of Columbia and the congressional districts in Nebraska and Maine.

There were a handful of other winning maps where Trump got 312 electoral votes but with a different combination of states — see if you can “think like the model” by guessing what those are. I’ll reveal the answers at the end. These occurred only 0.4 percent of the time, however.

So, indeed, Tuesday’s exact map was the modal (most common) one in our final model, coming up in 4,660 of our final 80,000 simulations. The next-most-common map (2446 sims) was Harris sweeping the swing states and then everything else going to form.

Pretty good, I guess? ¯\_(ツ)_/¯. But we can also look at the margins of victory the model projected in the 4,660 simulations where this map came up. In general, they were uncannily close to the actual numbers, missing by an average of only 2.3 points, given results as reported so far. (Many states are still finalizing their vote counts.⁷) Now, that isn’t quite as impressive as it sounds, because we’re giving the model a lot of clues — it’s been told the winner in all 50 states. Still, it’s a sign that the model’s internal logic is sound, and that American elections are quite predictable on a state level once you have a sense of the national landscape.

The exceptions are also interesting, though. Where did Trump overperform the most? (Keep in mind these are conditional forecasts: the margins the model would have expected given that it knows what states Trump won and lost.)

There are three groups here. First, Iowa unto itself, frankly because of the Sezler poll that was, uh, not highly accurate.⁸ Next, states that were affected by migration patterns. Lots of people are fleeing blue states because of the high cost of living. The conservatives who leave tend to wind up in Florida and Texas, so those states are getting redder⁹ — and many polls weren’t able to catch up to this.

But mostly, it’s deep blue states where Harris underperformed. Including her native California! But also New Jersey, where Trump lost by only 5 points! Then Illinois, Hawaii, New York (especially NYC) and Maryland too.

Not great for Harris on the surface. But this is actually one of the few silver linings to come out of Tuesday for Democrats. They had more votes than they needed in exactly these states, so losing voters there, especially while Republicans begin to get more voters than they need in high-population Florida and Texas, will greatly reduce and perhaps even reverse the Electoral College bias.

And where did Trump most underperform?

There are interesting stories here, too. Utah has gradually been getting more purple, between attracting a tech-sector workforce and Mormons not always being so keen on Trump. Is it going to be a swing state in 2028? No. But if you told me it was by 2040, I wouldn’t be shocked.

Washington, Oregon and Colorado have bucked other blue states by becoming even bluer, meanwhile. Though these are mostly vote-by-mail states that have lot of vote yet to count, my guess is that at most of this will hold up because they’re places where liberal blue state emigres tend to wind up. I know about this personally, actually: a whole branch of my family basically decided a couple of years ago to relocate from California to Seattle and various places in Oregon. This is true to some extent for Maine also on the East Coast. In Delaware, meanwhile, there was probably some residual sympathy for Biden.

On a night where Democrats got little good news, however, the fact that New Mexico remains solidly blue was a favorable sign: the state often doesn’t follow the trends we see from other states with lots of Hispanic voters. And under the circumstances, Harris’s 2-point loss in Georgia wasn’t so bad since it means Georgia is now closer to the tipping point.

OK, almost done. I’ll give you one more encouragement to subscribe, and then I’ll reveal the answers to the question I posed above.

Bonus round: other Trump maps with 312 electoral votes

Trump finished with exactly 312 electoral votes — his actual tally — in 5,002 of our 80K simulations. In 4,660 of those, he got the exact map we wound up with, sweeping the swing states but with Harris holding her ground everywhere else. (The inner Blue Wall remains solid even if the outer one doesn’t.)

But what happened in the handful of others? We need maps where Trump loses one of the swing states, but then wins some other state(s) to exactly make up his deficit. And that’s pretty hard to do. But here is the most logical option, according to the model:

In this map, Harris wins Nevada (+6 electoral votes) but loses New Hampshire (-4) and the Maine statewide vote¹⁰ (-2). It’s logical enough: in universes where Trump is overperforming among northern whites to the extent he wins the Blue Wall, Maine and New Hampshire could plausibly follow, and their results are likely to be highly correlated with one another. Whereas Nevada had been one of the bluer swing states in 2020 and marches to a different drummer.

The next-most-common map, the only other one to occur more than 0.1 percent of the time, was this one:

The logic here is that if Trump is really overperforming in the Blue Wall, he’ll also sometimes win its geographic and demographic neighbor, Minnesota. But if he’s overperforming in the Blue Wall and lagging among African-Americans, he could lose Georgia while doing so, which has the largest Black voter share among the battleground states. It’s not a very likely map, coming up in only 83 out of 80,000 simulations, but you may see versions of it more often if Georgia continues to trend blue relative to the rest of the country. And on Tuesday, Democrats mostly held their ground among Black voters while Hispanic, Asian American and “other” voters become true swing demographics.

1. This is very inside baseball, but during the election, we treated the Model Talk brand as concomitant with paid posts, and ran almost everything else other than Model Talks and SBSQs free. Outside of election peaks, we don’t use the Model Talk brand much, and generally the distinction between paid and free posts will be more fluid.

2. Weeks with SBSQs will generally get two paid posts, in other words. I’ll either skip the free post or run 3+ items that week. Once in a blue moon, SBSQ may run free if we’ve overdosed on paid posts, but the ability to ask questions will always be limited to paid subscribers.

3. For the record, I voted for Kamala Harris, so that’s the outcome I was rooting for as a citizen. Specifically, I voted for her on the Working Families Party ballot line in New York, which has fusion voting, meaning that the same candidate can be nominated by multiple parties. I don’t have any particular opinions about the Working Families Party, but it was the only party other than the Democrats to nominate Harris. I support third-party ballot access, and stronger performance helps third parties to qualify for the New York ballot in future elections.

4. Even though our forecast was near 50/50 for almost the whole race, there were certainly periods that were relatively better and worse for Harris and Trump. Our narrative content followed accordingly, with about an even mix of newsletters that presented optimistic cases for Harris and Trump. (That was not true when Joe Biden was running, but that’s because Biden was way behind in the polls.) So I essentially got to perform a randomized control trial on how partisans in both camps reacted to good and bad news.

And there was an asymmetry. Republicans are generally happy when you agree with them partway or half the time. Admittedly, the sorts of Republicans who encounter our work are not a representative sample, probably being on the moderate side — though you can find plenty of Trump supporters in the Silver Bulletin comments section.

Democrats, however — and here, I’m not referring so much to Silver Bulletin subscribers but in the broader universe online — often get angry with you when you only halfway agree with them. And I really think this difference in personality profiles tells you a little something about why Trump won: Trump was happy to take on all comers, whereas with Democrats, disagreement on any hot-button topic (say, COVID school closures or Biden’s age) will have you cast out as a heretic. That’s not a good way to build a majority, and now Democrats no longer have one.

5. Harris could have lost Nevada but won the Electoral College, so long as she retained Pennsylvania.

6. The asymmetry — why Trump was more likely to sweep even though the topline forecast was almost exactly 50/50 — is mostly because of Arizona, which was teetering into “lean Trump” territory with a 2.4-point Trump lead in our final polling average. Harris won all swing states but Arizona in about 4 percent of simulations, which basically gets her to Trump’s 20 percent when added to her clean sweeps.

7. Results are as of Thursday morning.

8. Our forecasts are based on a combination of polling averages and a regression estimate based on national polls and polls of other states that tends to assume a relatively uniform swing from the previous election. In states with lots of polls, the polling averages get nearly all the weight, whereas in states without any polls, the model basically just defaults toward 2020 plus a uniform shift toward Trump.

9. North Dakota, meanwhile, has been getting redder amidst the fracking boom.

10. She’d still split the two Maine congressional district electoral votes.