Recently I highlighted Andrew Gelman's excellent discussion of his--The Economist's--election forecasting model. Here are some additional forecasting discussions plus one new model.
The first item is Natalie Jackson's piece on the problems of election forecasting, both the known unknowns and unknown unknowns. She is pretty down on the whole business now, believing it's too hard to get right, the public doesn't understand the results and in the end it does more harm than good. No doubt her views are influenced by the HuffPost model she developed in 2016 which had Clinton's probability of winning at 98 percent (!)
The second item is a tweetstorm by Drew Linzer, who was responsible for Daily Kos' model in 2016, which was also quite optimistic, if not as optimistic as Jackson's. It's a good discussion of some of the general issues around election forecasting, touching on some of the topics discussed by Gelman. Too bad it's a thread on Twitter instead of a proper post or article somewhere. Still, it's worth looking at.
Finally, Alan Abramowitz is out with a (very) simplified model to predict the election. He is throwing out the influence of the economy and incumbency because of the peculiarities of this election year and instead just using Presidential approval (net in late June)--which he argues is currently driven heavily by approval of Trump's response to the coronavirus crisis--to predict the election. Yes, it's the proverbial one variable model! I wouldn't be surprised if it turned out to work quite well this year.
Abramowitz' current prediction based on -15 net Trump approval in late June gives Biden a 70 percent chance of winning the election and a predicted electoral vote of 319-219. If Trump's net approval was still at -15 in late October, the prediction would be 90 percent for a Biden victory and a 361-177 electoral vote (shades of Obama '08!). Abramowitz provides a table where you can interpolate predicted results based on other values for net approval.