> It really does matter how it got that way. If 20% of people of color vs. white people have a low income (a non-protected variable), I am perfectly allowed to use that variable to deny them a loan more often, even if the result is that 20% of people of color don't receive a loan. That's not a racist algorithm, however it would be racist if I start ignoring income of certain people as to achieve racial demographic parity in loans.

It's not that simple. The principle of "disparate impact" means that you can't, in general, make the decision in a way that adversely impacts people of color just because you didn't directly consider race in your decision.

You can use income in your decision because considering income "is necessary to achieve one or more substantial, legitimate, non-discriminatory interests." Income is directly related to ability to repay a loan, so you need to consider it. You can't just use any old variable you want to. You can't deny people for a loan based on what genre of music they listen to, because that has no justifiable connection to whether they will repay a loan.