US mortgage lending / Underwriting standards
How tight is US mortgage underwriting?
Approval standards for conventional first-lien home-purchase mortgages, 2007 to 2024, from the full HMDA public Loan/Application Register: a mix-adjusted approval-odds index, the raw approval rate, the debt-to-income and loan-to-value mix of new lending, income multiples, and what lenders say when they deny. In 2024, 2,727,441 such applications reached a credit decision.
What this page is, and is not. HMDA records applications and their outcomes; it carries no loan performance: no delinquency, no default, no prepayment. Everything here measures underwriting standards (who gets approved, at what leverage), not credit outcomes. Vintage default curves require the GSE loan-level performance datasets (Fannie Mae Data Dynamics, Freddie Mac), which need a free registration; that layer is planned and will be added once the registration is done.
This is not a fair-lending analysis. The approval model below deliberately excludes every protected-class field in the LAR: race, ethnicity, sex and age are not predictors, and nothing on this page compares approval odds across demographic groups. Fair-lending conclusions require methods (and confidential data) this page does not attempt. The FFIEC’s own release language for these files applies:
“HMDA data alone cannot be used to determine whether a lender is complying with fair lending laws. The data do not include some legitimate credit risk considerations for loan approval and loan pricing decisions.”FFIEC, press release announcing the 2019 HMDA data, June 24, 2020
The underlying files are the public (privacy-modified) LAR: DTI is published only in bins, CLTV and income carry “Exempt” and “NA” values, and 2018-2024 are Snapshot-tier files. Every measure below states its denominator and how much of it is actually reported. See the mortgage methodology for the action-code and denial-rate foundations shared with the rest of the mortgage layer.
The approval-odds index, 2018–2024 (2019 = 100)
Year fixed effects from a logistic approval model on the modern-era LAR: the odds that an observationally identical application (same DTI bin, CLTV bin, income bucket, loan size, purpose, occupancy and state) is approved, relative to 2019 and scaled so 2019 = 100. Below 100 means tighter underwriting. The index peaked at 109.9 in 2021 during the pandemic-era boom, then fell to 71.5 by 2024: the same observable application faced roughly 29% lower approval odds than in 2019.
Source: HMDA public LAR (CFPB/FFIEC), 2007-2024 Logistic model on 2,100,000 sampled applications (300,000 per year, reservoir sample, seed 42), conventional first-lien decisioned applications of all purposes, 2018-2024. No protected-class predictors. Methodology
Honest fit disclosure: holdout AUC 0.830 on 420,000 held-out applications. Approval models on the public LAR are limited by construction: the file carries no credit scores, no reserves, and only binned DTI, so the model explains lender decisions only partially. The index reads the year coefficients, not individual predictions. Because the model is fitted on a per-year sample (300,000 applications per year) rather than all 50.3 million records in the model universe, the index carries sampling noise: redrawing every year’s sample under a different seed moves it by at most 2.1 index points (AUC 0.830 either way) and changes no reading on this page, so treat the level as the signal and the decimal as indicative.
Approval rate, 2007–2024
The unadjusted share of decided conventional first-lien home-purchase applications that were approved: loans originated plus approvals not accepted by the applicant, over those two plus denials (action codes 1+2 over 1+2+3, the exact complement of the denial-rate definition on this narrower universe). The post-2008 tightening, the long 2010s easing, and the post-2021 tightening are all visible without any model.
Source: HMDA public LAR (CFPB/FFIEC), 2007-2024 Era break at 2018: the 2007-2017 legacy files carry no reverse-mortgage, open-end or business-purpose flags, so those exclusions apply only from 2018. Filter: loan_type = 1, lien_status = 1, loan_purpose = 1, action codes 1-3. Methodology
The DTI mix of applications, 2018–2024
Debt-to-income is published only in bins in the public LAR (“<20%”, “20%-<30%”, “30%-<36%”, single percents 36-49, “50%-60%”, “>60%”), so these are shares of published bins, never a continuous ratio. Shares are of decided applications with a reported bin; in 2024, 5.5% of applications carried “Exempt” or “NA” instead and are excluded from the denominator. The squeeze is visible: as rates rose after 2021, the share of applications above 43% DTI climbed from 22.2% in 2021 to 33.7% in 2024.
Source: HMDA public LAR (CFPB/FFIEC), 2007-2024 Published DTI bins grouped as labeled; '>43%' sums the bins strictly above 43 (44-49, 50%-60%, >60%), so a loan at exactly 43% is not counted as high-DTI. The legacy era (2007-2017) has no DTI field. Methodology
Leverage of originated loans, 2018–2024
Two leverage margins among loans that were actually made: the share with combined loan-to-value above 90 (small downpayments) and the share with DTI above 43% (stretched incomes). Both are computed only where the field is reported: in 2024, CLTV is reported for 92.2% of originations and DTI for 94.7% (the rest are “Exempt” or “NA”).
Source: HMDA public LAR (CFPB/FFIEC), 2007-2024 Originated loans (action code 1) in the same universe. The 2018 file names the CLTV column loan_to_value_ratio; 2019-2024 name it combined_loan_to_value_ratio. Methodology
Median income multiple, 2007–2024
Loan amount over the annual income relied on in the credit decision, median across originated loans where income is reported and positive. The multiple bottomed at 2.00x in 2011 as post-crisis underwriting demanded income headroom, then climbed to a 2.93x peak in 2021 when low rates stretched budgets against rising prices; it stands at 2.60x in 2024.
Source: HMDA public LAR (CFPB/FFIEC), 2007-2024 Income is reported in thousands of dollars in both eras; the multiple uses loan_amount / (income x 1000) (legacy: loan_amount_000s / applicant_income_000s). 2,142,069 of 2,195,689 2024 originations carry a usable income. Loan amounts in the public LAR are privacy-binned midpoints, so the multiple inherits that rounding. Methodology
Why lenders deny, 2024
The first denial reason lenders report (denial_reason-1) on the 417,830 in-universe denials of 2024, using the FFIEC code sheet verbatim. Debt-to-income leads, consistent with the post-2021 squeeze in the DTI mix above.
| Code | Denial reason (FFIEC label) | Denials | Share |
|---|---|---|---|
| 1 | Debt-to-income ratio | 137,645 | 32.9% |
| 3 | Credit history | 128,719 | 30.8% |
| 4 | Collateral | 41,089 | 9.8% |
| 6 | Unverifiable information | 31,172 | 7.5% |
| 7 | Credit application incomplete | 23,101 | 5.5% |
| 9 | Other | 21,546 | 5.2% |
| 5 | Insufficient cash (downpayment, closing costs) | 21,093 | 5.0% |
| 1111 | Exempt | 7,154 | 1.7% |
| 2 | Employment history | 6,176 | 1.5% |
| 8 | Mortgage insurance denied | 135 | 0.0% |
| 10 | Not applicable | 0 | 0.0% |
Source: HMDA public LAR (CFPB/FFIEC), 2007-2024 First reported reason only (lenders may report up to four); lender-reported classifications, available 2018 onward. Codes and labels quoted from the FFIEC public LAR data-fields documentation. Methodology
scripts/build/build_underwriting.py from the same verified loan-level files as the rest of /mortgage; the broad-universe denial rate recomputed there reconciles exactly with the published denial-rate series.