Here's something which surprised me a bit:
In 2012, (obviously the last presidential election with an incumbent), a simple uniform swing better predicted the differences between the results in different states than the major polling aggregators.
https://twitter.com/davidshor/status/1291819403512885252Although, nerd that I am, I took the numbers in this table, played with simple linear regression a bit, and actually a blend of the 538 state estimates and the uniform swing numbers (more heavily weighted towards the latter) would have done even better.