You Can't Ask Their Race. Regulators Expect You to Know Anyway.
When self-reported race data is unavailable, most compliance tools rely on a widely-used proxy method that systematically understates racial disparities — creating exam and litigation exposure institutions don't see coming until it's tested. CREST replaces that guess with an auditable, reproducible measurement that accounts for proxy error.
Available now: the aqEthnicity package for Python and R.
pip install "aqethnicity[local]"
remotes::install_github("tleitch/aqEthnicity", subdir = "r")Example notebooks on GitHub (open in Colab). For a free STRATA key, email terry@aequum.ai.
Getting this wrong is expensive — and it's not hypothetical
Someone has to estimate race when it isn't self-reported — for HMDA fair lending analysis, a CRA exam, or a discrimination claim in litigation. If that estimate is wrong in a way that systematically favors the institution, examiners and opposing counsel will eventually find it. “Our vendor's tool produced the number” is not a defense that holds up once it's actually tested — and the tools built to make this simple were never built to survive that test.
Where PPP forgiveness went — and where it didn't, by neighborhood.
Within five miles of a community bank — and the loan went elsewhere anyway.
During the Paycheck Protection Program, minority-owned businesses disproportionately borrowed through a nationwide platform channel that forgave loans at markedly lower rates than branch-based community banks. Most were not remote — they sat inside the everyday footprint of a community bank that, on its own record, forgave at a far higher rate.
Chalavadi, Leitch & Pastor. Decomposes the forgiveness gap into within-lender and between-lender components, and traces the between-lender share to sorting into a nationwide platform channel — not to physical distance from a community bank.
Read the PaperThe compliance layer, and the engine beneath it
CREST produces an auditable, reproducible disparity estimate. STRATA is the inference engine that feeds it — available on its own as a hosted API or a private no-egress deployment.
CREST accounts for how much legacy proxy methods can understate a measured disparity — producing auditable, reproducible disparity estimates for compliance self-testing, regulatory exam support, and litigation.
- Financial institutions for fair lending self-testing and exam defense
- Applicable across fair lending, employment, insurance, housing, healthcare, and criminal justice
- Enterprise — includes full STRATA imputation$100,000 / yr
- Standalone — litigation & consulting$15,000 / matter
The ML inference engine beneath CREST: an LSTM + geographic → XGBoost pipeline. On a held-out voter-registration sample (986,801 records), STRATA assigned 18.0% of non-white people to white, against 28.5% for surname + ZIP BISG: a 37% relative reduction. Available as a hosted API or a private no-egress deployment for regulated institutions.
- BISG / BIFSGFree
- STRATA — with a free keyto 10,000 records / dayUse it from Python or R with the aqEthnicity package.
- Starter$1.00 / 1K rows
- Growth — 1M–10M$0.75 / 1K rows
- Scale — 10M+$0.50 / 1K rows
- AWS Marketplace$0.03 / 1K rows
- STRATA Premium — no-egress$20,000–$45,000 / yr
Keys by email to terry@aequum.ai until self-service registration is live.
Why existing tools get this wrong
BISG's errors are SES-correlated, not random
Minorities in wealthier areas are misclassified as white; whites in poorer areas are misclassified as non-white. The result: in small-business lending, black vs non-black approval gaps are understated by over 43% (44% black vs white) (Greenwald et al., J. Financial Economics 157 (2024), 103857). CREST accounts for this structure, not just the headline accuracy number.
A validated alternative to BISG
On a held-out voter-registration sample (986,801 records), STRATA assigned 18.0% of non-white people to white, against 28.5% for surname + ZIP BISG: a 37% relative reduction. In our PPP working paper (Sorted Out, SSRN 7262782), on the common sample of 2,841,057 PPP records that every classifier can score, STRATA's block-group/tract tier reaches 87.6% black recall against 75.2% for BISG. The durable value is the measurement layer, not the benchmark.
CREST's bias simulation and disparity estimation methods are patent pending.
Two deployment modes for every buyer
Hosted API for researchers, fintechs, and advocacy organizations. Private AWS container for banks, insurers, and mortgage lenders where PII cannot leave the institution's environment. Data never leaves your infrastructure in the Premium tier.
Published and peer-reviewable
Ready to talk?
Whether you're a compliance team preparing for exam, counsel quantifying discrimination harm, an insurer testing models for unfair discrimination, or a researcher — we'd like to hear from you.