Dem
51
GOP
49
Although Nate Silver is no longer at FiveThirtyEight, the new Nate, G. Elliott Morris, is also a statistician and modeling guy. The group has a complex statistical model that takes into account everything but the kitchen sink. It is a probabilistic model—that is, it assigns a probability to a large number of variables (say, Biden has a 51% chance of winning Nevada) and then generates random numbers for all the variables and sees what happens. For example, on one run Biden could win Arizona, lose Wisconsin, win Pennsylvania, etc. On the next run, he might win Arizona, win Wisconsin, and lose Pennsylvania. By running the model 1,000 times, the results can be averaged to a daily prediction. Currently it predicts a 53% chance of Biden winning. If you are interested in simulation models, they have a long and detailed post discussing the model. It would be nice if people could download the code and tweak it, but it is not available.
Some of the factors are interesting. They allow correlations between states. As Wisconsin goes, so goes Pennsylvania, or something along those lines. They do have a point. Factors that influence Wisconsin do influence Pennsylvania, but probably not so much Georgia. To determine similarity, they looked at presidential elections since 1948 to see who voted with whom. If you would like that data, we have it from 1900 to 2020 in Excel format and also in .csv format (or look at our Data galore page).
The model also includes multiple economic indicators. For example, have incumbent presidents done well when inflation is at 3%, when unemployment is below 4%, when the stock market is up that year, and a bunch of other things? There is plenty of data on that and the model can be tuned to predict past elections well.
If how the model works is not your thing, just bookmark this page, which gives the results of last night's 1,000 runs. They will be updating daily.
If we may throw in our $0.02, here are some comments. First, every model has a weighting factor for each input. Should the polling average for a state count for 20%? 25%, 30%? Should the unemployment rate count for 3%? 8%? something else? If there are 100 variables in the model, there have to be 100 weights. These are all somewhat arbitrary. In some years, the economy dominated the election, but in others it was not a factor. How do you get them all right?
Second, Donald Trump is not your garden-variety Republican. Much of his base would walk a mile barefoot over broken glass to vote for him. That certainly was not true for Bob Dole or Mitt Romney. How do you factor this in? Similarly, Joe Biden does not have the love Jack Kennedy or Barack Obama had. Many of his voters aren't really voting for him, but against Trump. The ups and downs of the economy aren't going to play a role for people thinking along these lines and history may not be a good guide.
Third, the election is 5 months away and there are known unknowns and unknown unknowns that could play a big role. For example, one of the known unknowns is "What will Judge Tanya Chutkan do when the Supreme Court rules on presidential immunity?" Suppose the Court rules that presidents are immune to prosecution for official acts (e.g., if a president impounds funds for some program, he can't be later prosecuted for failing to faithfully execute the laws). But it could simultaneously rule that presidents are criminally liable for private acts. With a decision like that, Chutkan could hold hearings starting July 15 on whether Trump's actions on Jan. 6, 2021, were official or private. Those hearings could be explosive. A great (or awful) debate performance is also a known unknown. Unknown unknowns could include a terrorist act, a massive natural catastrophe, a candidate having a health incident, a stock market crash (or unexpected surge), or a lot of other things we can't think of now. How do you model these things? Well, you can't. Still, the model is another way of looking at things besides polling averages.
Fourth, with half a dozen states on knife's edge, the decisions of a few thousand voters in each state to vote or not vote could be determinant. What if there's an early November snowstorm in, say, Milwaukee? How do you model that? Maybe long-range weather forecasts for Nov. 5 are included in the model. We don't know. (V)