Like all folks modeling presidential elections, we're forced to use small samples (e.g. n=12) to infer what the broader universe (n=∞) looks like. This is hard! But if you overfit models to past data, you'll often make poor inferences about what unobserved data looks like.
So while we spend a lot of time thinking thru different model specifications, it's mostly *not* about trying to maximize fit on past data. It's more about thinking through edge cases and specifications that will be robust under a variety of conditions.
The reason we generally prefer polls>economy is because different modeling choices on the economy can yield radically different answers (i.e. a Biden landslide if you look at GDP or a Trump landslide if you look at income) while different polling averages produce similar answers.
