1 · Raw gap
How many times more often the lender denied Black applicants than white ones. Part of this can be explained by income, debts and loan size.
2 · After finances
A model that sees only finances (no race, no neighborhood) predicts each applicant’s chance of denial. We divide each group’s real denials by its predicted ones. 1.0× means finances explain the gap.
3 · Look closer
The bar is the 95% range. Look closer means the whole range sits above 1.0×, even after correcting for testing dozens of lenders at once.
What this can and can’t tell you
Public mortgage data has no credit scores. A lender whose Black applicants have weaker credit would show a gap here even if it treats everyone the same. So a flag means “pull the files and check,” which is what a fair-lending analyst would do next.
Lenders are anonymised and numbered by size (Lender 1 = most applications). We only include lenders with at least 50 Black and 50 white applications and 5 denials in each group.
An analyst can run the same check on any one lender with python src/audit_lender.py --lei <lender ID>. Method and code: github.com/kabirshah4/race-blind-isnt-fair.