Every question below is answered from the complete public CFPB HMDA record, with the source, the limit that must travel with the figure, and — where the number can be recomputed from the raw file — the query to do it.
22.08% — 262,250 denials out of 1,187,606 decisioned FHA forward applications. The denominator is originated plus approved-not-accepted plus denied; reverse mortgages (HECM) are excluded. An earlier figure of 21.7% included them and was corrected on 2026-07-26.
Limit: Associational, not causal. HMDA contains no credit scores. Never a prediction about an individual application.
FRC-001 · source page · reproduce this
Among the 100 largest FHA lenders in 2025, observed denial rates ranged from 1.8% to 78.7% — a 44-fold spread on the same federally insured program in the same year, against a national rate of 22.1%.
Limit: Associational, not causal. HMDA contains no credit scores. Never a prediction about an individual application.
FRC-002 · source page · reproduce this
AmeriSave Mortgage Company, at 78.7% — 18,056 denials out of 22,944 decisioned FHA applications. A high observed rate reflects applicant mix, channel and documentation practice as well as underwriting, and is not evidence of wrongdoing.
Limit: Associational, not causal. HMDA contains no credit scores. Never a prediction about an individual application.
FRC-003 · source page
Flat Branch Mortgage at 1.8%, tied with Lakeview Community Capital, among the 100 largest FHA lenders. Within the eleven-lender panel we track over time, CrossCountry is lowest at 6.4%.
Limit: Associational, not causal. HMDA contains no credit scores. Never a prediction about an individual application.
FRC-004 · source page
About 38% of the explainable variation attaches to lender identity rather than the applicant file. Adding a single lender-identity variable to a model of applicant and loan characteristics raises McFadden pseudo-R-squared from 0.171 to 0.276 across 859,090 decisions.
Limit: Associational, not causal. HMDA contains no credit scores. Never a prediction about an individual application.
FRC-005 · source page · claim passport
It persists. Across 2018-2025, 80.8% of the variance in large-lender denial rates sits between lenders rather than within them, and the most lenient large lender was the same institution in seven of eight years.
Limit: Associational, not causal. HMDA contains no credit scores. Never a prediction about an individual application.
FRC-006 · source page
Cleveland-Elyria, Ohio: high-volume lenders ranged from 6.4% to 80.1% in 2025, a 73.7-point gap, computed over lenders with at least 100 decisioned applications in the metro. This has never been corrected or withdrawn.
Limit: Associational, not causal. HMDA contains no credit scores. Never a prediction about an individual application.
FRC-007 · source page · reproduce this
Idaho: FHA applications under $150,000 were denied at 53.4% against 12.0% for loans of $250,000 or more — a 4.45-times penalty. Utah follows at 3.87x and Washington at 3.44x.
Limit: Associational, not causal. HMDA contains no credit scores. Never a prediction about an individual application.
FRC-008 · source page · reproduce this
On observed rates, yes, but the category is the wrong unit. In 151 metros where both compete, national-footprint lenders denied 23.6% against 16.7% for local and regional lenders, and the national lender was stricter in 112 of them. However the softest door in the country (1.8%) and the strictest (78.7%) are both narrow-footprint lenders: going local raises the average and the variance at the same time.
Limit: Associational, not causal. HMDA contains no credit scores. Never a prediction about an individual application.
FRC-009 · source page · claim passport
Not for every lender. Inside identical cells of state, loan amount, income, DTI and leverage, one large wholesale lender runs 14.2 points softer than its peers below 80% CLTV and -36.6 points stricter above 96.5% — a 50.8-point swing, with a within-lender permutation p of 0.0. Other lenders are flat. The data locate the transition between published leverage bands and never at a point.
Limit: Associational, not causal. HMDA contains no credit scores. Never a prediction about an individual application.
FRC-010 · source page
Yes, in two directions at once. Among the 50 largest FHA lenders, the six owned by homebuilders price -1.33 points below their peers relative to APOR (permutation p=0.0001) while denying at 1.26 times their own applicant-mix expectation, against 0.74 for everyone else. The raw denial-rate difference is not statistically distinguishable; only the peer-adjusted measure separates them, and it is not a size effect.
Limit: Associational, not causal. HMDA contains no credit scores. Never a prediction about an individual application.
FRC-011 · source page
Debt-to-income, but the mix shifts systematically with loan size. Collateral is cited in 25.9% of denials on loans under $100,000 and 14.1% at $300-400,000, dropping out of the top three above $400,000, where DTI dominates at 38.3%. Reason codes are lender-reported and a file may cite up to four.
Limit: Associational, not causal. HMDA contains no credit scores. Never a prediction about an individual application.
FRC-012 · source page
The median large FHA lender cites it in 2.2% of its denials, but two lenders cite it in more than 72%, and volume-weighted it accounts for roughly 17.9% of FHA denials nationally. It is the one common denial reason that says nothing about the borrower's finances. It correlates with denying above own-mix expectation (Pearson 0.452) but the rank correlation is only 0.201, consistent with a filing convention used by a small extreme group.
Limit: Associational, not causal. HMDA contains no credit scores. Never a prediction about an individual application.
FRC-013 · source page
Yes, and steeply. Nationally, FHA applications under $100,000 were denied at 46.9% against 18.5% for loans above $400,000. The pattern holds in every state we can measure, with penalties ranging from about 1.6 times to 4.45 times.
Limit: Associational, not causal. HMDA contains no credit scores. Never a prediction about an individual application.
FRC-014 · source page
Of 238 lenders with at least 500 decisioned applications in 2025, 49 have an observed/expected denial ratio whose 95% confidence interval lies entirely above 1.0, once thin-coverage rows are excluded. The screen is keyed by LEI and joins to a counterparty list in one line.
Limit: Associational, not causal. HMDA contains no credit scores. Never a prediction about an individual application.
FRC-015 · source page
No detectable relationship. Pooled within-lender across 77 lender-years and 11 lenders, the correlation between log annual volume and denial rate is -0.208, the 95% interval crosses zero, and a permutation test gives p=0.148. We publish no capacity indicator because we do not have one.
Limit: Associational, not causal. HMDA contains no credit scores. Never a prediction about an individual application.
FRC-016 · source page
No detectable relationship. Across 56 lenders, the correlation between the observed/expected denial ratio and the median rate spread over APOR is -0.179, with a 95% interval of [-0.422, 0.088] and a permutation p of 0.195. The softest lender in the comparison carries one of the highest spreads.
Limit: Associational, not causal. HMDA contains no credit scores. Never a prediction about an individual application.
FRC-017 · source page
Not identifiable from the available sample. Across six lenders with both books mapped, Pearson is 0.054 while Spearman is 0.543, the 95% interval runs from -0.79 to +0.83, and deleting one institution moves Pearson to 0.835. We previously published this as evidence that lender quality is product-specific; that claim was withdrawn on 2026-09-06.
Limit: Associational, not causal. HMDA contains no credit scores. Never a prediction about an individual application.
FRC-018 · source page
Yes, repeatedly, and the log is public and dated. It includes the July 2026 HECM universe correction (21.7% to 22.1%, 29,691 records removed), a withdrawn cross-program correlation claim, and three publisher errors found in September 2026 — stale lender rates on sixteen pages, wrong state small-loan figures, and a superseded universe left in our machine-readable files. No Cleveland metro finding has ever been corrected, narrowed or withdrawn; if a correction is not listed on that page, it did not happen.
Limit: Self-reported. No figure we publish has yet been independently reproduced.
FRC-019 · source page
Almost, and the residual is published rather than hidden. Summing the 52 state rows gives 1,185,337 decisioned applications against a national 1,187,606. The 2,269-record difference carries no usable state assignment, and it is denial-heavy: 83.2% of those records are denials. State-level rates are therefore very slightly understated relative to the national figure.
Limit: Associational, not causal. HMDA contains no credit scores. Never a prediction about an individual application.
FRC-020 · source page
Not necessarily. Denial rates describe applications that were filed, and say nothing about buyers who were redirected before applying. The New York metro recorded only 7,144 decisioned FHA applications in 2025 — 31st of 319 metros, behind markets a fraction of its size — which is consistent with exclusion happening upstream of the record, though the public data cannot demonstrate the mechanism.
Limit: Associational, not causal. HMDA contains no credit scores. Never a prediction about an individual application.
FRC-021 · source page
Yes, and the queries are published. Each headline figure has a recipe — the SQL, the expected result and the mistake people usually make — against the raw public CFPB file. If a query produces a different number, we ask to be told and it goes in the corrections log.
Limit: Associational, not causal. HMDA contains no credit scores. Never a prediction about an individual application.
FRC-022 · source page
It is measured monthly and published. On the 15th of each month, eight AI systems answer twelve frozen questions whose correct answers we publish ourselves, graded against a rubric published before any answer is collected. In the September 2026 administration, scores ranged widely, one system reported a correction we never issued for the third month running, and three systems could not locate our public corrections log.
Limit: Associational, not causal. HMDA contains no credit scores. Never a prediction about an individual application.
FRC-023 · source page
Nobody. No lender, vendor or AI company funds this work, there is no advertising and no affiliate placement, and we accept no payment from any institution appearing in our data. The underlying datasets are free under CC BY 4.0; optional research products are listed openly and cannot alter, delay or remove a published figure.
Limit: Self-reported.
FRC-024 · source page
Method: how every figure is computed · reproduce them yourself · how our own tables reconcile · what we tested and found nothing · four DOI-registered papers