We analyzed eight years of federal Home Mortgage Disclosure Act data covering eleven major FHA lenders across all fifty-one U.S. jurisdictions. The question: is a lender's denial behavior driven by market conditions, or by the institution itself?
Finding 1 โ The 80/20 law. Roughly 80% of the variance in FHA denial rates is explained by which lender received the file. Only ~20% is explained by when it was submitted (ICC = 0.796).
Finding 2 โ Hierarchy persistence. The ranking of lenders by denial rate survived every market condition from 2018 to 2025 โ including COVID, the fastest rate-hike cycle in four decades, and the refi boom. Year-over-year rank correlation averaged 0.905 and never fell below 0.77. The same hierarchy holds within 96% of individual states.
Finding 3 โ The 9+2 structure. Statistical model selection (BIC) favors a two-class structure: nine lenders whose behavior is institutionally fixed, and two lenders โ Freedom Mortgage and NewRez โ whose denial behavior couples directly to market friction. A four-class "regime" model was tested and rejected.
A denied FHA application is usually interpreted as a borrower problem. The data suggests it is, to a measurable degree, a routing problem. In 2025, CrossCountry Mortgage denied 6.37% of FHA applicants while NewRez denied 53.27% โ under the same federal guidelines, in the same country, in the same year.
No market crisis in eight years changed which lenders sit at the top or bottom of that range. Institutional identity is the dominant force; market conditions move the level, never the order.
| Class | Lenders | Behavior |
|---|---|---|
| Stable Core (7) | CrossCountry, Guild, UWM*, Planet Home, Mr. Cooper, Rocket, loanDepot | Identity-driven; predictable from 8-yr history (LOO error 2.8pp) |
| Improving Trend (1) | PennyMac | Secular improvement: 52.8% (2018) โ 23.6% (2025); historical average overstates current risk |
| Unpredictable (1) | Wells Fargo | Behavior uncorrelated with market; FHA volume down 63% since 2019 |
| Regime-Sensitive (2) | Freedom, NewRez | Denial behavior couples to market friction (r=0.66โ0.84); LOO error 10.9pp |
*UWM is the system's designated falsification target: its 2025 denial rate sits +6.8pp above its 8-year average. If the 2026 data confirms continued drift, the structure revises from 9+2 to 8+3 โ and we will publish that revision.
Scientific honesty requires listing the casualties. Four hypotheses were tested and rejected: a "convergence predicts shocks" law (confound: mean reversion), an 11-lender lead-lag network (failed false-discovery-rate control), an asymmetric tighten-fast/loosen-slow law (data shows the opposite), and a four-lender regime class (failed cluster validity and bootstrap stability).
The surviving structure passed: two-way variance decomposition, within-state replication across 51 jurisdictions, leave-one-year-out validation (ฯ=0.894), label-randomization testing (p<0.0001), and BIC model comparison against simpler and more complex alternatives.
As a final check, we ran the lender feature matrix through five clustering approaches with no labels supplied โ K-Means, Ward hierarchical, DBSCAN, Gaussian mixtures, and the gap statistic. K-Means, Ward, and DBSCAN independently converge on a two-group structure (silhouette 0.29; weak but consistent). The data-driven volatile group is {Freedom, NewRez, PennyMac} โ and the lag analysis explains the third member: PennyMac's variance is a monotonic improvement trend, not market coupling. Machine learning sees variance; the causal analysis explains it. Bootstrap stability is marginal (ARI 0.58 ยฑ 0.45), which is the honest cost of classifying eleven institutions. GMM-BIC and gap-statistic results at higher k were discarded as small-sample artifacts.
Eight annual observations is a small time series; confidence is bounded at roughly 70โ85% and sub-year dynamics are invisible at this resolution. The classification is FHA-specific and may not transfer to conventional lending. The 2026 HMDA release is the next falsification test, and this page will be updated with the result either way.
Eight-year open dataset behind this research (CC BY 4.0): 2018–2025 denial time series →