Well, some models say that but not "the" models. There are many, many election forecasting models so it is not so surprising that you can find a few models that predict a Trump victory. This is particularly the case if you look at so-called fundamentals models that focus on economics and disregard any survey measurements. These are the kinds of models cited by Steven Rattner in his recent Times opinion piece.
The models cited by Rattner say that, if the 2020 election was going to be decided solely on the basis of incumbency and selected aggregate economic measures, Trump would be in very good shape. But I think we kinda knew that already. Put another way, if it weren't for the fact that Trump is so unpopular, he'd be pretty popular.
But wait, aren't these models pretty accurate, despite their seemingly obvious problems?. Nope. Economist Ray Fair's model, for example, is not particularly accurate and has been revised by Fair numerous times to make it come out right (or at least less wrong).
As Nate Silver accurately remarks about these models:
"A lot of smart people don't seem to realize these economic models pretty much suck at predicting elections. They do well when backtested but they're overfit/p-hacked and empirically have done a terrible job of actually predicting elections out of sample.
The ones that use economic data as a prior, and blend it with polling data, do OK and seem to add some value. But the ones based on economy/"fundamentals" alone range from mediocre for the best ones to bordering on junk science for the worst ones."
As an example, Fair's model in 2016 predicted Donald Trump would get 54 percent of the national two-party vote. Whoops.