guardrail string | records_used int64 | raw_denial_rate float64 | mcfadden_r2_profile_only float64 | mcfadden_r2_with_lender float64 | door_effect_share_of_explained float64 | overlay_residual_top15_strict list | overlay_residual_top15_lenient list | reason_fingerprint_families dict |
|---|---|---|---|---|---|---|---|---|
Aggregate historical analysis only. No individual scoring function exists or is saved. | 859,090 | 0.2188 | 0.1712 | 0.276 | 0.3797 | [
{
"lei": "549300YIQ7S7Z8PIHE53",
"n": 22927,
"denials": 18039,
"actual": 0.787,
"expected": 0.35,
"overlay_residual_pp": 43.68,
"name": "AMERISAVE MORTGAGE COMPANY"
},
{
"lei": "549300NOCASXPA34X033",
"n": 7931,
"denials": 5321,
"actual": 0.671,
"expected": 0.313,... | [
{
"lei": "549300MHH03ICJYEPE85",
"n": 5016,
"denials": 572,
"actual": 0.114,
"expected": 0.327,
"overlay_residual_pp": -21.334,
"name": "549300MH…"
},
{
"lei": "549300S5YL3OH0IVCS62",
"n": 7008,
"denials": 828,
"actual": 0.118,
"expected": 0.311,
"overlay_resi... | {
"aile_1": {
"baskin_sebep_kodu": "4.0",
"sebep_payi": 0.274,
"uye_sayisi": 12,
"ornek_uyeler": [
"254900M2C4Q6DATP2889",
"549300EMNDEK4BA8WB53",
"549300FGXN1K3HLB1R50",
"549300HW662MN1WU8550",
"549300JYXTZDSPJEPI44"
]
},
"aile_2": {
"baskin_sebep_kodu": "1.0... |
Aggregate historical analysis only. No individual scoring function exists or is saved. | 859,090 | 0.2188 | 0.1712 | 0.276 | 0.3797 | null | null | null |
The Door Effect — FHA Denial Variance Decomposition (2025)
38% of the explainable variation in FHA mortgage denial outcomes is the lender — not the file.
This dataset contains the published results of a variance decomposition over 859,090 FHA credit decisions from the complete public CFPB HMDA 2025 record. Two nested logistic models were compared: Model A uses only observable application characteristics in the federal record (income, loan amount, DTI band, small-loan flag, state); Model B adds a single variable — lender identity. Model A explains 17.1% of outcome variation (McFadden R²); Model B explains 27.6%. The difference — 37.97% of everything explainable — is the door. Demographic fields are deliberately excluded; this is a study of lender behavior, not borrowers.
Files
| File | Contents |
|---|---|
summary.json |
Headline statistics: records used, raw denial rate, both R² values, door-effect share |
lender_residuals.csv |
Top-15 strictest and top-15 most lenient doors: actual vs. mix-adjusted expected denial rate and the overlay residual (pp), with lender name and LEI |
reason_families.csv |
Denial-reason fingerprint clusters: dominant reason code, reason share, member count, example member LEIs |
door-effect-2025.json |
Canonical results file as published at financeratecalc.com (byte-identical to the site copy) |
Read this before quoting
HMDA contains no credit scores. "Expected" is computed from the observable characteristics above, so the residual means excess denial not explained by observable federal-record characteristics — it is consistent with lender overlays, channel differences, documentation practices, and unobserved risk mix, and cannot by itself prove any single cause. All figures are historical and aggregate (minimum 500 decisions, 100 denials per lender). This dataset contains no individual predictions and never will: a denial is a data point, not a verdict on any applicant. No individual scoring function exists or is saved.
Provenance & reproducibility
- Source: FinanceRateCalc analysis of the complete public CFPB HMDA 2025 record (https://ffiec.cfpb.gov/). Denominator: actions 1, 2, 3; loan_type 2 (FHA).
- Canonical copies: https://financeratecalc.com/door-effect.html · https://financeratecalc.com/data/door-effect-2025.json
- Companion paper: Yetis, Z. (2026). The Door Effect: Lender Identity and Unexplained Variation in FHA Mortgage Denial Outcomes — Evidence from the Complete 2025 HMDA Record. SSRN Working Paper 7309319. https://ssrn.com/abstract=7309319 · doi:10.2139/ssrn.7309319 (https://doi.org/10.2139/ssrn.7309319)
- Full methodology: https://financeratecalc.com/door-effect.html
- Related: A Public Benchmark for Consumer Mortgage AI Accuracy, SSRN 7156938
Note: the canonical JSON keeps the original field names of the reason_fingerprint_families block (aile_* = family_*, baskin_sebep_kodu = dominant reason code, sebep_payi = reason share, uye_sayisi = member count, ornek_uyeler = example members). reason_families.csv provides the same data with English column names.
License & citation
CC BY 4.0 — reuse freely with attribution and a link back.
FinanceRateCalc (2026). The Door Effect: FHA Denial Variance Decomposition,
2025 Federal Record. https://financeratecalc.com/door-effect.html
- Downloads last month
- 47