Credit is, at its core, trust. Trust, well-placed, makes everything about humanity better. Seen that way it feels warm and fuzzy. But, "credit," meaning the

Credit is, at its core, trust. Trust, well-placed, makes everything about humanity better. Seen that way it feels warm and fuzzy. But, "credit," meaning the business of telling you whom you can trust (like a bureau), is very very weird indeed. Without pre-aggregated or nearly-instantly-compiled predictions about trust, it can only spread out in a personal way. But impersonal relations are crucial to trade, to fairness, to the open society (that is, you shouldn't be constrained to trade only with your family/tribe/neighbors). Hence, a credit bureau, in the ideal, is a trust-promoting mechanism -- something meant to facilitate the best in mankind. And it rewards faithfulness to promises. Furthermore, in most commerce we can use a default assumption of "trust." Most people will return the rental car, will repay the loan, will queue up to pay for the groceries, etc. Here's the rub: with the assumption of trust, then the target outcome to model (the labels you're looking for in ML terms) are the defaults or "don't trust" outcomes. Which means that the best credit bureaus will hungrily look for all labels of, and all features which can be predictive of, defaults. And it must do so opaquely so as to avoid being gamed by adversaries. So now the institution which is meant to facilitate trust and reward cooperation ... is a secretive digger for dirty laundry. Just a reflection. It's weird. (Background: I spent the last five years starting a fairly successful alternative lender, and a less successful alternative credit bureau.)