2020 Prediction Time

Like most of this post, from my Election Modeler.  More about how it works here (though I've added a fair number of bells and whistles since then).


It's Election Eve, and my models have been hard at work predicting how things are going to play out. And I'm not going to bury the lede: barring incredibly unlikely polling errors or hitherto unknown success by Trump in the courts, Biden is going to win, and likely by a large margin. If you like, you can look at the state-by-state predictions and probabilities [here] (though I definitely stand by the overall averages, not the particular outcome of one state or another).  Here are the main stats:

  • 97.4% - The odds that my model gives to Biden carrying the electoral college
  • 358 - The median electoral votes I expect him to get
  • 8.9% - The expected margin in the popular vote
  • 7.1% - The expected margin in tipping point state (PA) -- which is essentially equivalent to how much the polls need to be off by in order for Biden to lose. They were off by less than 3 in 2016, by the way.
  • Dems are probably going to get to 51 in the Senate (but you can look at my modeler if you want more on that)
This last point, the margin in PA, probably deserves a little comment.  My model is, for the most part, very simple.  I make a weak prior assumption to model unpolled states, but basically I use public, high-quality, non-partisan pollsters rated by 538 with a B or better. But not everyone does that.  538 themselves through everything (even hyper partisan garbage -- which is down weighted, but not by enough), which is why their average for PA is 4.8.  RCP has PA at a nail-biting 2.9.  It matters how you average them, but since there is not a dearth of high quality polls, I use them.

I then smooth over the previous month.  Each state is assumed to have quirks amounting to 3 points of randomness in addition to polling error, and I also put in the possibility that all polls everywhere could be off by 3 points (standard deviation, so occasionally it'll be more).  All of those numbers are justified by historical election statistics.

There's an important point here, found in most models.  It is, indeed, possible that the polls are all off by a certain amount.  But they'd need to be off by 5 or 6 (roughly double typical) in order for this to even be a 50-50 race.  And even if a single state is a crazy outlier, there are lots of backups:

From the tracker.  All of the states that are within 10 points. Well, technically Missouri got cut off.


Biden's starting point, with only states that he leads in by more than 10, is 217. (I was too lazy to treat ME-02 and NE-02 separately, so for now, I'm assuming they vote with their state.  As it stands, Biden will probably win both, bumping him up 1 EV. That happens not to matter in most scenarios).  Including Missouri, there are 17 states in this category, Biden leads in most of them, and he needs to win 4 or so (depending on the 4) in order to win. Not to beat a dead horse, but Biden's lead in PA is slightly larger than Trump's in Alaska.  And not to beat another dead horse, but it's equally likely that Biden outperforms his polls than underperforms.

It is, of course, possible that he loses PA, but then Biden needs to win any of AZ, FL, NC, OH, GA (all of which he's leading in), or TX, where he's behind.  Put another way, even if Trump wins a long shot, he still needs to win 6 coin flips in a row (roughly 3% odds).  It's possible, but tat this stage, far more likely that he's cheated.

For the sake of completeness, though I don't put any stock in it, here's my "winner take all" map -- assuming each candidate wins the states that they're currently ahead in:

A fairly typical prediction, but I can absolutely imagine one or more of these are wrong, and in either direction. I'm personally kind of bearish on Ohio.

We were burned by polls in 2016, so it's worth wondering if we really know how anything is different this time.  There have been lots of analyses pointing out that Biden has consistently gotten over 50% in polling, and that the number of undecideds and third party voters are far smaller this time.  

From my polling trends widget

Or that Biden's lead is much more stable, larger, and (unlike in 2016) shows no sign of collapsing near the end:

Basically, Biden's lead has been between 8 and 11 since June, and Trump has never, at any point, been within 4 points.

Even beyond polling, we do know something.  Prof. Michael McDonald of the University of Florida has been compiling early voting statistics in his US Elections Project.  In 2016, 47 million people voted early.  This time, it's 98 million as of the last update (likely to top 100 once the dust has settled).  And those early ballots are dominated by Dems.  Wherever states have party ID, Dems have returned 45.2% of the ballots.  Republicans, 30.5% (the rest are minor party or no affiliation).  We're on track to likely get between 150 million and 160 million votes in total (compared to 137 million 4 years ago) and it is hard to envision a scenario where that is good for Trump.

One could argue that the early vote advantage is only a consequence of Dems being afraid to vote in person, but that's only partially true.  In states like Colorado, where there is all mail-in, 70.1% of registered Dems have returned ballots, compared to only 63.6%.  And it's not just one state.  Jon Ralston of the Nevada Independent has become known as an expert in looking at votes coming in, and in his latest dispatch from the early voting blog he basically spells out that Biden has essentially already won Nevada.

I am nervous, but honestly, it's hard to imagine a scenario where the Dems could be in a much stronger position going in.


Comments