Everyone Knows The Forecast Did Not Exist in 2020. What This Article Presupposes Is…What if It Had?

Earlier this week, Election Twitter had a bit of a heart attack when S-tier pollster Ann Selzer and The Des Moines Register released a poll showing Donald Trump leading Joe Biden by 18 points in Iowa. To put that result in some context, Trump defeated Biden by 8 in 2020 (which is around where the Solid Purple forecast currently has the race in 2024). Selzer’s final poll of that race last time, which had been considered a tossup by most prognosticators in the run-up to the election despite Trump defeating Hillary Clinton by more than 9 points in 2016, showed Trump leading Biden by 7.

I distinctly remember where I was when that poll dropped - I was walking around my neighborhood on Halloween night with a pit in my stomach, feeling like this was the canary in the coal mine. And honestly, it kind of was! The election ended up being much closer than anticipated and states like Iowa and Ohio ended up tricking Democrats into thinking they were still key parts of the coalition two cycles in a row.

Obviously, you would rather not be down 18 with any population that has a tendency to correlate with crucial Midwestern swing states like Wisconsin. That said, I noticed that missing from the hot takes on Twitter, where people were melting down trying to square how such a result could be possible when every other poll shows Biden keeping things competitive in the region, was an acknowledgment I think is very important for determining how to feel about Selzer’s findings without dismissing or downplaying them.

At this same point in the 2020 cycle, Selzer had Trump leading Biden by only 1 point. Should 2024 be a repeat of 2020, where the margin shifts six points in favor of the incumbent between now and November while the national polling average stays more or less steady, Trump would be winning Iowa by 12 in a roughly tied national environment. Seems like a pretty normal outcome that would be well within expectations!

Which got me thinking...sure, I did not build a model in 2020. The reason I built one this time around was precisely so I did not have to agonize over any individual poll, though it’s a safe assumption that I will be sweating bullets in the leadup to Selzer’s final survey of Iowa in late October. But what if I took the model that exists today, stripped out all of the training data from the 2020 election, and plugged in the polls from this exact moment in time four years ago?

Well...it would not have been any more wrong than anyone else’s forecast!

As you can see, the hypothetical forecast, which I must note was built with the knowledge of the actual 2020 election outcome, regardless of how much I tried to unravel those assumptions from this version, is bearish on Biden’s odds of winning the Electoral College despite having a very robust lead in the national popular vote (for comparison, the FiveThirtyEight polling average was Biden +8.3% on this day, less than half a point lower than what Solid Purple would have projected the margin to be on Election Day). I looked at the other forecasts that did exist at the time and FiveThirtyEight had Biden with a 75% chance of winning and a projected popular vote margin of 7.1%. The Economist, which had given Biden a 97% chance of winning the election at the end of 2020 and is currently the outlier model declaring Trump a clear favorite in the 2024 race, had Biden with a 84% chance and a popular vote margin almost in the double digits.

Iowa, for what it’s worth, is exactly where Selzer had it at the time.

But let’s look at where the model would have ultimately ended up by Election Day.

At the moment of truth, the Solid Purple model would have given Biden a 77% chance of winning the Electoral College despite being statistically almost certain to win the popular vote. The forecast would have called 47/50 states correctly plus each of the congressional districts in Maine and Nebraska, missing Florida, North Carolina, and, somewhat hilariously, Texas. I’d like to think I would have been skeptical of this at the time - I know I was when other people’s forecasts were saying that it was within the realm of possibility - but in defense of Team Blexas, the state’s partisan lean relative to the national popular margin ended up being R+9, so the model would have at least got that right. It only would have taken 1 in 40 people who voted for Trump to have chosen Biden instead, and the polls, along with this forecast, would have been right on the money in most states.

Not all of them, though.

I mapped out the projected lean of each state (relative to the national popular vote) compared to the actual result and there are only a few states that really stick out. Utah is the biggest and brightest, but that makes sense considering Evan McMullin pulled in a substantial amount of support in 2016 and it had been an open question whether or not the state’s heavily Mormon population that had been notably Trump-skeptical compared to other segments of the conservative coalition had come around to him during his first term. Hard to model under the best of circumstances.

But Florida, man. Fucking FLORIDA. The contrast of Georgia’s left shift and Florida’s sprint to the right truly cannot be discussed enough, and I truly have no idea how to square it with 2024 polling that shows Georgia and Florida reverting somewhat back to the mean. The 2024 model thinks the polls are slightly underrating Biden in Georgia while overrating him in Florida, but if you’re not confident about that, don’t worry, because neither am I!

One more thing! Since I was already getting my hands dirty throwing in some hypothetical data and seeing what the model spits out, I decided to override the state and national polling variables with the actual results of the 2020 election, just to see what odds it would give Trump and Biden.

Despite Biden’s national popular vote margin increasing 3 points over the current actual forecast, his odds don’t improve that much, mostly because the tipping point state is currently about two points to the right of the nation (the modal outcome in the forecast is Trump winning, but Biden has enough paths to victory that he still ends up coming out ahead in 51% of simulations), while the tipping point state in 2020 (Wisconsin) was almost 4 points to the right of the nation. We’ll see if that holds up over time, as the delta between the country and the tipping point state grew a point or two between the hypothetical June and November 2020 model runs you see above.

Next model update will be on July 1st, just after the debate and right before the Independence Day holiday. Not sure if there will be much in the way of movement but there’s only one way to find out!