CFPB Home Mortgage Disclosure Act (HMDA) Data

CFPB Home Mortgage Disclosure Act (HMDA) data is the loan-level record of the American mortgage market: every application a covered lender takes and every loan it purchases, reported under HMDA and Regulation C as machine-readable rows rather than press releases. Each record describes one loan application or purchased loan across roughly 100 fields - the lender's Legal Entity Identifier, geography down to the census tract, loan type (conventional, FHA, VA, RHS), purpose, lien status, amount, rate spread, HOEPA status, points and fees, property value, applicant age, ethnicity, race and sex, credit score type, automated underwriting system and principal denial reason, plus Census-appended tract demographics such as minority population share.

Two properties make this franchise hard to replicate. Denials are recorded with the same completeness as originations, so the loans that never funded sit in the table beside the ones that did. And the geography runs finer than almost any commercial mortgage panel reaches - tract-level codes on every row, which turns neighborhood lending maps into a group-by rather than a data-collection project.

Get a sample of this dataset and we will return real loan rows cut to the lenders, states or tracts you name.

One row per application, straight from a 2023 filing - a conventional first-lien cash-out refinance in Harris County, Texas:

activity_year            : 2023
lei                      : 549300WN576XI5QFH005
derived_msa_md           : 26420                  # Houston-The Woodlands-Sugar Land, TX
state_code               : TX
county_code              : 48201                  # Harris County
census_tract             : 48201430800
conforming_loan_limit    : C                      # conforming
derived_loan_product_type: Conventional:First Lien
derived_dwelling_category: Single Family (1-4 Units):Site-Built
derived_ethnicity        : Hispanic or Latino
derived_race             : Race Not Available
derived_sex              : Male
action_taken             : 5                      # closed for incompleteness
loan_type                : 1                      # conventional
loan_purpose             : 32                     # cash-out refinancing
loan_amount              : 495000
loan_term                : 360                    # 30 years
income                   : 155                    # thousands, as reported
applicant_age            : 45-54

Read the row closely and the panel's honesty shows up twice. action_taken is 5 - the application was closed for incompleteness, never became a loan, and is in the table anyway, which is exactly why denial-rate and fall-through work computes cleanly here and nowhere else. And derived_race reads 'Race Not Available': privacy modification replaces sensitive values with explicit categories rather than blanks, so absence is coded, not silent. Scale one row to millions and every lender's year becomes countable by product, tract and outcome.

Eighteen documented groups anchor the dictionary below, verified against the released data structure. The spine is identity (activity_year, lei) and a four-step geography ladder (derived_msa-md, state_code, county_code, census_tract) that lets you roll the same rows up to metro, state, county or neighborhood without a separate crosswalk. Outcomes and attributes arrive as documented enumerations - action_taken in eight states, loan_type, loan_purpose, denial_reason-1 - with derived convenience columns such as derived_loan_product_type spelling the combinations out in words:

Three chips summarize the footprint:

• Geography: United States - all 50 states, DC, Puerto Rico and the Virgin Islands - geocoded to census tract, with MSA, state and county codes on every row and Census-appended tract demographics riding along.
• Temporal: 2007 through the latest filing year (filing year 2025) - a historic series from 2007-2017 stitched onto the expanded modern collection from 2017 onward, giving a nineteen-year panel that spans a full housing cycle, crash to recovery to rate shock.
• Granularity: one row per loan application or purchased loan per institution - no sampling, no aggregation layer. Nationwide first-lien owner-occupied files alone ran roughly 4.8M-8.3M rows per year across the historic era, each carrying around 100 fields.

Set against the wider Datadory catalog - where the average quality score across all 1,744 datasets is 7.81 - this slice scores 9/10, carried by exhaustive official field documentation and a volume no private mortgage panel can match at this geographic depth.

API, files, or your warehouse. Daily, weekly, or hourly.

You choose the channel and the cadence; the field dictionary above travels unchanged across all three. Full-year extracts suit teams loading entire filing periods once and joining them to their own portfolio or servicing tables, scoped feeds suit products surfacing lender or tract metrics inside an app, and warehouse delivery suits analysts running fair-lending and market-share models in SQL. Cadence changes are a settings conversation, not a re-integration project.

• Lender competitive tracking - LEI-keyed application and origination counts by county, MSA and product line, so a rival's retreat from cash-out refinancing shows up in the rows quarters before it reaches an earnings call.
• Fair-lending analysis - denial reasons, rate spreads and pricing set beside applicant race, ethnicity, sex and age at tract resolution: the canonical evidence base for disparate-treatment work.
• Mortgage market sizing - sum loan_amount by geography, product type or lien status to size addressable markets for lenders, insurers and proptech from observed transactions, not survey recall.
• Credit-model feature engineering - millions of labeled outcomes (originated, denied, withdrawn, incomplete, purchased) with borrower attributes attached, ready for approval-probability modeling.
• Community lending and CRA narratives - tract minority population share and income measures show where lending reaches and where it stalls, street by street.
• Expensive-credit surveillance - rate spread and HOEPA flags isolate high-priced lending pockets by lender and neighborhood, historically the earliest warning regulators and journalists both watch.

Market researchers and consultants get the definitive citation for any claim about who originates what, where - when a deck says a metro's refi wave peaked, this is the table that dates it. Competitive intel and product teams get lender-level volume and mix shifts watchable filing period by filing period. Investors and quant researchers get a nineteen-year, cycle-spanning panel of lender behavior for factor construction and event studies. Data scientists and ML engineers get outcome-labeled rows at a scale that makes supervised approval models trivial to train and honest to evaluate. Journalists and academics get the numbers regulators themselves cite, with tract-level geography for neighborhood storytelling.

Three things worth knowing upfront. First, privacy modification is built in: records are altered to protect applicants and borrowers before they exist in analyzable form - ages arrive in bins, some demographic cells read as explicit 'not provided' categories rather than blanks, and geographic detail carries its own suppression rules. Treat demographic precision as deliberately bounded, not degraded. Second, the analytical load sits in enumerated codes: action_taken runs 1 through 8, loan_type, loan_purpose, applicant_sex and friends each carry their own value lists, and several reserve 'exempt' as a value for filers with partial obligations. Third, the series has two labeling eras: filings before 2017 use plain-language column names and a narrower field set, while 2017 onward uses code-style names and the expanded schema. Long-run panels should be assembled on the fields both eras share; we deliver the mapping either way.

Notes and adjacent reading:

• World Bank Global Findex Database 2025 - demand-side self-reports of borrowing behavior against this file's regulator-recorded applications.
• Findex vs CFPB HMDA - the comparison page that works the two lenses on credit markets together.
• FDIC National Survey of Unbanked and Underbanked Households - who sits outside the credit system entirely, the population HMDA rows can never see.
• Federal Reserve Survey of Consumer Finances (SCF) - household balance sheets behind the application behavior recorded here.

Browse the whole vertical on the Consumer Finance data hub or the ranked shortlist of the best consumer finance datasets.