Enhanced HMDA Report — Regulator View

STRATA proxies race/ethnicity on records without self-reported demographics, from applicant name + property address. It returns a five-class probability vector, a dominant race, an index of uncertainty (iu), and an address lookup-quality tier (block-group+tract, tract, ZCTA, or name-only). Raw geographic identifiers never leave the STRATA container — only anonymized cohort tokens are emitted.

CREST estimates each disparity with a de-biased, Fisher-consistent weighted method-of-moments estimator that inverts the proxy's misclassification using the calibrated posterior P(true race | proxy) — correcting the attenuation (regression dilution) that biases a plain plug-in of the proxy toward zero. The 95% interval is a tract-cluster-robust interval anchored to the property census tract, so it reflects proxy + finite-sample + within-tract correlation and tests for a systematic disparity rather than a chance in-sample gap. It is closed-form and deterministic, keyed to a published seed, so results are reproducible and comparable across institutions.

Reading the result. "Actionable" means the 95% CI excludes zero — a statistically stable finding. "Indeterminate" means the CI includes zero — the point estimate is not reliable and should not be acted on alone. A large gap can be indeterminate (a proxy artifact); a small gap can be actionable.

Screening vs. confirmatory analysis. These disparity estimates are marginal (unconditional) statistics — the correct tool for identifying which institutions warrant closer examination, analogous to a raw HMDA disparity ratio. The estimator assumes lending outcomes are independent of the proxy conditional on true race (Y ⊥ proxy | true race). Geography can challenge this: neighborhood factors that affect lending (property values, local income, LTV distributions) are correlated with race through census tract, and the proxy encodes the tract. Sensitivity analysis shows that under race-correlated geography confounding (the realistic case in segregated markets), the marginal estimator may overstate the true race effect by approximately 1–3.5 percentage points depending on the strength of the neighborhood–race correlation. This does not disqualify the screening statistic — every standard HMDA disparity analysis faces the same issue — but it means these estimates flag where to look, not the final adjudicated effect size. For confirmatory analysis controlling for legitimate underwriting factors (LTV, property value, loan purpose), use CREST's covariate-adjusted mode, which recovers the within-geography race effect stably across confounding levels.

Method reference: arXiv:2504.21259.

All data is synthetic. No real lender, applicant, name, or address appears; LEIs are anonymized. The records are assembled from real, auditable public components so they read plausibly:

Scope: geography is New York only in this example (the vendored Census tract table covers NY's 5,411 tracts). The disparities are illustrative, designed to exercise the actionable-vs-indeterminate distinction — not empirical findings.