I'd be surprised if base models don't cover that. You'd need to incorporate far more information for every single voter if you're posing this as a classification task with each voter's information being the input. For instance, you team up with someone like Cambridge Analytica and add in their social media interactions from Facebook, you team with ad companies who will give you sites they've browsed. If ethics are not a bar, you can get places they've visited over the last six months, youtube videos they've watched, articles they've spent most time reading, time at which they paid taxes, amount they spend on guns, interests and voting patterns of their friends and relatives, their likes and retweets on Twitter and Instagram, keywords from their recent posts, heartrate from their tracking devices, sentiment analysis of their recent texts to check their emotions, engagement with political ads, and so on and so forth. Even if you had all this information about every single voter, your models will not be accurate enough because you can't capture all of the dynamic variables. People who can build these models are a dime a dozen like me, but this data is what will be expensive, noisy and inadequate.