Your model sucks if you are arbitrarily choosing factors. You choose the factors with the most significant correlations, because those correlations are least likely to be "overly simplistic".
Why would we reject that zip codes are more likely to default on a loan? Seems like information I would like to be aware of if I was a home lender.
I certainly look at crime rates of a community before I live there. While a bad crime rate certainly doesn't make potential neighbors "guilty by association", it certainly increases the likelihood that one of my neighbors might be actually guilty.
Directing focus at "people who drink a lot of beer" means considering people who who drink a lot of beer at home as guilty by association, ultimately due to the subjective priorities of whomever pushed for that model.
Obviously in the expected value sense, charging on correlations is lucrative for the company (as is any justification for raising prices on a set of customers if your competitors do it too). But in the exact same way as saying certain zip codes are more likely to default on a loan, which we rightfully reject.