Skip to content

How unpredictable is the 2020 election?

Our poll aggregation partially pools toward a “fundamentals”-based election forecast. Elliott Morris summarizes:

What I did was create an index of economic growth over the past years using standardized percentages changes in 9 different variables (some of which you and Chris use in your LEI work). This gets combined with approval ratings, trial-heat polls and an interaction for polarization (what share of people tell the ANES they switched parties over the past 4 years) in a regression for the national forecast for Election Day.

Then I disaggregate those with a simple model based on how each state voted in the past few elections, plus the forecast national vote. The formula is literally state vote lean ~ lean in past cycle + lean in cycle 8 years ago + some home state effects + democratic national vote share. All interacted with polarization. Everything going back to 1972 here (1948 nationally).

Then we just random walk toward that prior, which implicitly puts more weight on the polls over time.

We fit our model to the available polls in 2012 and 2016 and it gave reasonable results.

Bob Erikson followed up with some skepticism:

For 2020 you can squeeze out enough info for the best model possible and then you run into a combination of Trump’s caustic history, a pandemic, a pandemic-induced economic catastrophe, BLM, then Trump’s response and the real fear that Trump will do what he can to sabotage an election he sees as likely to lose. Put another way, if the GOP candidate were a generic candidate like Bob Dole in his prime, you could be quite confident in your model’s predicted trajectory. With Trump….? I do not mean to bash forecasting in general or your model. And I do not necessarily expect Trump to pull a rabbit out of a hat; he is more likely to totally crash. But unlike in other elections, believing any model going forward, including even the model in my head, gives me the heeby jeebies.

Elliott replied:

I’m not so sure about that. Trump’s victory in 2016 matched the fundamentals dead-on. Our best prior is to expect the same to be true this time, +/- some error.

Further, polls should account for most of these whacky differences in how voters evaluate him (If indeed they do exist). The real potential problem that keeps me up is that the polls are just wrong again because of weighting, filtering, sampling issues etc that haven’t been fixed or have crept up since 2016.

Bob responded:

I will separate two arguments.

(1) I think models cannot infer the effect of the virus-induced economic disaster. It’s not the same as if it occurred as a natural result from dynamics of the economy (2008, 1929). So if the model were driven solely by the economy (the Economist’s model is not), I would not trust it much.

(2) I am a believer in understanding the forecasting power of trial-heat polls. So that does provide an anchor. I should not have left any impression otherwise. I do fret about an extra amount of variance in the polls that can result from here to election day. But then again campaign effects tend to be overrated and the polls have some remarkable stability.

And I remarked that most, if not all, elections are idiosyncratic in one way or another:

1948: first election without Roosevelt

1960: first Catholic candidate

1964: candidate abandoned by his own party

1968: first serious third-party candidate in a long time

1972: youth vote

1976: first major-party candidate in a long time who was an outsider

1980: Reagan was a scary unknown quantity

1992: volatile 3-candidate election

2000: seemed like a normal election but Gore should’ve won easily and he didn’t

2004: first election in a long time where foreign policy was a key issue

2008: first nonwhite candidate

2012: first Mormon candidate

2016: first female candidate, also Trump was an unusual candidate.

Is 2020 more unusual than all of these past elections? I don’t know.

And Chris Wlezien wrote:

One thing I wonder about is (possible) changes to how people vote and whether and how this matters for turnout and vote-counting, and polling.

Back in the old days when the Monkey Cage was a blog, we could’ve had this discussion there.


  1. Anon says:

    One thing I think Bob is on the right track about are the unforeseen effects on in-person voting and the election in general from the coronavirus.

  2. Justin says:

    “Is 2020 more unusual than all of these past elections? I don’t know.”

    But the unusual things in those past elections were mostly not important drivers of your model’s results. I think Bob’s point (1) is potentially very important.

  3. Carlos Ungil says:

    Kind-of-relevant XKCD:

  4. Ben says:

    > “fundamentals”-based election forecast

    The word fundamentals sorta bothers me.

    The ( page introduces it as something like, “Our model is updated every day and combines state and national polls with economic indicators to predict a range of outcomes.” which I like much better.

    Why call it “fundamentals” rather than the “Elliot/Merlin/Andrew” model, or the “Economist poll/aggregate + economics indicator model”?

    I think it’s fair to point out the wild parts of all the other elections but,

    > 2012: first Mormon candidate

    Was that even an issue? I probably would have gotten that wrong if you’d asked me it like a yes/no. I don’t think it’s fair to stack that up against double digit unemployment + 100k deaths.

    And there’ll definitely be tons of procedural covid19 effects like Anon and Bob said. I applied for an absentee ballot in NY via e-mail in late April. This was under a special provision covid19 provision that allows anyone to vote absentee (even if they aren’t absentee). I got an e-mail back just a few days ago (the primary is 7 days from now) that I’d e-mailed my application to the wrong address and needed to send it to somewhere else. I verified on that I had done the right thing, and just said screw it and early voted in person.

    I think the model is valuable as a reference point (this is what works well enough for 2012/2016) from which we can think about all the other craziness. I guess it’s the same refrain — without the actual model we’d just be making up a reference point. So even if we can’t model all the wild stuff going on now, this is probably a better place to guess from?

    • Andrew says:


      1. “Fundamentals” is not precisely defined, but it generally refers to economic and political conditions that are separate from the particular candidates running.

      2. I agree that in retrospect, and even at the time, the Mormon thing maybe wasn’t such a big deal. But there have been polls asking people, “Would you vote for a candidate who is X,” with different X’s, and when X = Mormon, you do get some chunk of people saying, No, they wouldn’t vote for a Mormon. Some evangelical churches are really anti-Mormon.

      • Ben says:

        > it generally refers to economic and political conditions that are separate from the particular candidates running

        Right, and I think that sentence also makes a lot of sense.

        Like, who decided these were fundamentals? Doesn’t fundamental sorta imply they’re well known basics? If so aren’t all models using fundamentals? If so what makes this model different? Why not “economic and political election model”? Why would particular candidates running not be fundamentals? Surely this model does take into account who is actually running (presumably the polls ask Trump vs. Biden), so is it actually a fundamentals model? Etc.

    • Martha (Smith) says:

      “The word fundamentals sorta bothers me.”

      Reminds me of the time a student wrote on a teaching evaluation, “She spends too much time on the fundamentals, and not enough on the basics,” — or perhaps it was, “She spends too much time on the basics and not enough on the fundamentals”.

  5. John Williams says:

    “1980: Reagan was a scary unknown quantity” He had been governor of California, so we in California had a pretty good idea what was coming. I’ve been following elections since 1952, and this really is an outlier. When before did we have a president assure us that he would leave office if not re-elected?

  6. Joshua says:

    I think that an election prediction that changed as much as yours did over a period of three months has a fundamental problem of being too reactive to short-term influences.

    > (1) I think models cannot infer the effect of the virus-induced economic disaster. It’s not the same as if it occurred as a natural result from dynamics of the economy (2008, 1929). So if the model were driven solely by the economy (the Economist’s model is not), I would not trust it much.

    I think that any election prediction methodology that is based to a large degree on economic indicators has a problem in the context of this particular presidential election. I”d be willing to bet that this year is an outlier as much of the country, and not just Trump diehards, will give Trump a Mulligan on the recent economic downturn. That may turn around a bit by November if the economy doesn’t improve…but I question how much it would turn around. In basically any other year, the incumbent would a huge it for this kind of economic downturn.

  7. somebody says:

    The objection is that idiosyncrasies of the moment break the “usual” relationship between “fundamentals” like economic performance and election results.

    The counterargument is this

    > Further, polls should account for most of these whacky differences in how voters evaluate him (If indeed they do exist).

    The only relationship that matters for these forecasts is the one between polls and election results, and I don’t see evidence of any decoupling there. However unusual the times, elections are still essentially a big poll with more at stake and stricter auditing.

  8. Alex says:

    Andrew, your reply concerning the idiosyncratic nature of previous elections is dead on, and reminds me of David Mayhew’s arguments in Electoral Realignments. We can always come up with a story about why This Election is Different, but typically, that’s all it is: storytelling.

  9. paul alper says:

    In the last presidential election, the following excuses were offered by experts to explain why their forecast went awry:

    a. Even in Russian roulette there is 17 % chance of failure
    b. Well, she did win the popular vote by almost three million
    c. We noted that things were changing near the end
    d. The FBI and KGB did it
    e. Our competitors did even worse
    f. Sam Wang of the Princeton Election Consortium had to eat an insect and we didn’t

    For the upcoming one, let us hope the excuses do not include unanticipated assassination(s), nuclear war, a military coup and/or The Biblical Rapture.

  10. Michael Nelson says:

    Polls are amazing for the same reason statistics in general is amazing: you don’t have to observe a thing to account for it. The election may be idiosyncratic, but then, so will be the opinions of poll respondents. It is generally also helpful to add objective measures of some of those factors (like the economy) that historically have reduced the noise. The worry expressed by Erikson is that, when elections are too weird, maybe those additional factors will go against the historical record and turn out to add more noise than information. The proposed solution is that we drop the fundamentals altogether. But consider that the “argument from idiosyncrasy” implies that any good model of any election needs to have a variable representing “amount of idiosyncrasy” that automatically rebalances the weights on each of the fundamentals to account for their lower contribution of information. And where would we get this variable from? Two options: if we have a direct measure of the weirdness of an election, we can use that; if not, we can ask people how weird this election seems to be and use that.

    In other words, we either add a new fundamental to the model for which we currently have no reliable measure, or we add an opinion that is likely already accounted for in the person’s choice of candidate (and adds noise). These are both bad ideas. Dropping the fundamentals completely is just a special case of the former, where the weight on fundamentals is either 1 or 0, and the threshold for switching from 1 to 0 is determined by the pollster’s subjective (i.e., noisy) perception of the election’s weirdness. So that’s a special case of a bad idea.

    Until we have a systematic, principled way to account for election weirdness, it’s just not feasible to let weirdness influence the model. Except via its unobserved impact on opinion and on fundamentals. Which is amazing.

Leave a Reply