A city auditor told me my loan model was 'just fancy redlining,' and she wasn't wrong
She pulled up a map showing my approval rates dropped hard in two zip codes here in Baltimore where most folks are Black, even though I never fed race into the model. I switched from using zip code as a feature to testing the model blind against each neighborhood, and the gap shrank from 14 points to 3. Anyone else had to strip out a feature they thought was neutral?