Loading…

Hit 500 loan denials reviewed in one month, and the pattern scared me | Algorithmic Fairness Debate - Free To Speak

Hit 500 loan denials reviewed in one month, and the pattern scared me

I work as a data auditor for a small credit union in Cleveland, and last month I went through 500 auto loan applications that got denied by our scoring model. What surprised me wasn't the number itself, it was that 78 percent of those denials had a zip code in the lower income side of town, even when the applicant had a credit score above 650. I pulled the raw variables and found the model was leaning hard on address history, not just payment behavior. A guy with a 720 score and a stable job got denied because he moved three times in two years, stuff a human loan officer would ignore. Meanwhile similar profiles in wealthier neighborhoods sailed through. I get that algorithms need consistency, but this feels like a redlining workaround dressed up as math. Has anyone else found a hidden proxy variable like that in their own system, and how did you get leadership to actually look at it instead of defending the model?

2 comments

2 Comments

the_olivia1mo ago

Wow, that hits close to home. I found something similar when I was checking our smaller loan approvals last year. We had a rule that killed anyone who used a prepaid phone plan, and it turned out way more lower income folks paid that way. I ran the numbers and showed leadership that the prepaid phone thing was just a stand in for income level, and once I broke it down by actual income brackets, the pattern was undeniable. What got them to listen was presenting it as a fairness risk, not just a math problem. I also had a compliance officer look at it and say we were opening ourselves up to a lawsuit, and that got their attention real fast. Ended up pulling that variable out and our approval rates for lower income areas jumped by a solid chunk, with no real change in default rates.

bennett.harper1mo ago

Hell yeah, that's the kind of fix that actually means something. Once you frame it as "we're gonna get sued" instead of "we should be nicer," people suddenly find the time to listen.