Appendix b — Data Card — HMDA (Home Mortgage Disclosure Act Loan Application Register)

Provider & what it is

The HMDA Loan Application Register (LAR) is the public, near-census record of mortgage applications in the United States. Under the Home Mortgage Disclosure Act, most lenders must report nearly every mortgage application they receive — not just loans made, but ones denied, withdrawn, or fallen through. Per application you get the action taken (originated, approved-not-accepted, denied, withdrawn, closed-for-incompleteness), the loan amount, the property location (state, county, census tract), and applicant demographics (race, ethnicity, sex, income). Since 2018 you also get a richer file: interest rate, points, debt-to-income, loan-to-value, property value, and more.

The provider is the CFPB (Consumer Financial Protection Bureau), which runs the modern public HMDA Data Browser, with the FFIEC (Federal Financial Institutions Examination Council) publishing the per-year bulk files. It is one of the largest public microdatasets in U.S. finance — tens of millions of records per year.

The reveal is also the limitation the whole fair-lending literature turns on: HMDA records who applied, where, for how much, and what happened — but not the applicant’s credit score. It can show you that denial rates differ by race; it cannot tell you whether that gap is discrimination or unseen differences in creditworthiness. That single missing column is why fair-lending work leans on design, not more controls (Ch 6.4; Sun & Gao, 2019).

Coverage

Years: modern Data Browser coverage from 2007 forward, with the 2018 schema expansion (Dodd-Frank/Regulation C) as a hard break: pre-2018 and post-2018 files have different fields and codings. The new fields (rate, DTI, LTV, property value, NMLS id) start in 2018.
Universe: depository and non-depository lenders above a reporting threshold; smaller lenders are exempt, so HMDA is near-census, not total. The threshold has changed across years — [CHECK] the exact loan-count threshold for your pinned vintage, because it affects which small lenders appear.
Size: multiple gigabytes per nationwide year. Never download the full national LAR to a laptop.
Revisions: a year is published, then re-published with corrections. The vintage matters; pin it.

Key identifiers

Geography: the county is the 5-digit FIPS code; the tract is the 11-digit census-tract code; plus the two-letter state. These are your panel keys for aggregating to county-year or tract-year.
Lender: the institution is identified by LEI (Legal Entity Identifier) in the modern file, replacing the older respondent ID + agency code pair used pre-2018.
Action taken: the coded outcome (1 = originated, 2 = approved-not-accepted, 3 = denied, 4 = withdrawn, 5 = closed-incomplete, …). The denial-rate outcome is built from these codes.
There is no person identifier — HMDA is application-level, not borrower-panel; you cannot follow an individual across applications.

Access path

Two disciplined routes, both free, no key:

CFPB Data Browser API — best for a handful of states/years. The aggregation endpoint lets the server count actions by your chosen dimensions so you never move the microdata: base https://ffiec.cfpb.gov/v2/data-browser-api/. There is also a filtered-records endpoint for pulling raw rows when you must.
FFIEC bulk files — best for the whole country. Per-year pipe-delimited LAR files; download once to the camp container’s read-only data area and aggregate in chunks with pyarrow/pandas. The container mirrors these so you are not re-downloading gigabytes.

import requests, pandas as pd
# CFPB HMDA Data Browser AGGREGATION endpoint (public, no key).
# Ask the server to count actions by state, so the multi-GB microdata never moves.
HMDA_AGG = "https://ffiec.cfpb.gov/v2/data-browser-api/view/aggregations"
def fetch_hmda(states, year):
    params = {"years": year, "states": ",".join(states),
              "actions_taken": "1,3"}          # 1 = originated, 3 = denied
    r = requests.get(HMDA_AGG, params=params, timeout=60); r.raise_for_status()
    df = pd.json_normalize(r.json()["aggregations"]); df["year"] = year
    return df
# raw = fetch_hmda(["VA","MD","NC"], 2021)   # run on the container, pinned vintage

Aggregate server-side whenever you can. CFPB pre-counts by state × action (and finer geographies), so pull counts, not microdata, and build the denial-rate panel from those. [CHECK] exact parameter names against the live API docs for your pinned vintage — CFPB has renamed fields across releases, so the snippet is the pattern, not a stable contract.

License & note

Free and public. HMDA LAR is published by the FFIEC/CFPB as public data; cite the CFPB HMDA Data Browser (and the FFIEC bulk file if used) plus your pinned vintage and access date. No key, no redistribution restriction. The non-negotiable obligation: name the vintage, because the API serves whatever snapshot CFPB currently hosts and the same query returns different numbers across re-publications. Per CONVENTIONS §5, the gigabyte-scale bulk files stay on GMU infrastructure; you aggregate down first.

Gotchas

No credit score. The single most important caveat. A raw denial-rate or pricing gap by race confounds discrimination with unobserved creditworthiness (Week 2 OVB). This is why Sun & Gao (2019) reached for a clean design, and why Bartlett et al. and BHR are notable for getting at the score from outside HMDA (Ch 6.4).
The 2018 schema break. Pre- and post-2018 files differ in fields and codings. Do not splice a pre-2018 series to a post-2018 series without reconciling the schema — the “new” rate/DTI/LTV fields simply do not exist before 2018.
It’s enormous. Never pull the national file to a laptop; aggregate server-side or in chunks on the container.
Thin cells are noise. A rural county with 40 applications has a wildly swinging denial rate. Filter out cells below a minimum decided-application count (e.g., 100) so sampling noise in tiny geographies does not dominate the panel.
The denominator is a choice you must state. “Denied / (originated + denied)” excludes withdrawn and incomplete applications — defensible, but a referee will ask. Write it into the spec (CONVENTIONS §4).
Vintage drift. Quote no figure without naming the snapshot it came from. [CHECK] the exact snapshot date on the camp container before citing a number.

First 10 rows (illustrative schema sketch — not real values)

Aggregation response (CFPB Data Browser), shape only — counts, not microdata:

state	actions_taken	count	year
VA	1	412903	2021
VA	3	51277	2021
MD	1	298140	2021
MD	3	44011	2021
…	…	…	…

Raw LAR record (post-2018 filtered pull), shape only — one row per application:

activity_year	lei	state_code	county_code	census_tract	action_taken	loan_amount	derived_race	applicant_income	interest_rate
2021	EXAMPLELEI000000000	VA	51059	51059482101	3	325000	Black or African American	78	NA
2021	EXAMPLELEI000000000	VA	51059	51059482102	1	410000	White	142	3.125
…	…	…	…	…	…	…	…	…	…

Derived county-year denial-rate panel (what you actually regress on), shape only:

county	year	denied	apps_decided	denial_rate
51059	2021	1842	19204	9.59
51013	2021	1130	14072	8.03
…	…	…	…	…

Which chapter / lab / capstone uses it

Week 4, Lab 4 — Clean DiD on HMDA + a state policy shock. The lab’s Path A builds exactly the county-year denial-rate panel above from the Data Browser API, then runs TWFE → event study → Callaway–Sant’Anna against clean controls. The card’s aggregation pattern is the lab’s recipe.
Week 6, Ch 6.4 — fair lending in the algorithmic era (Bartlett, Morse, Stanton & Wallace; Bhutta, Hizmo & Ringo). HMDA is the spine of both papers, and the chapter’s whole pivot is what HMDA cannot see (the credit score) and how those authors get it from outside HMDA.
Capstone 1 — Fair Lending on HMDA. The terminal use: a real disparity-plus-clean-design paper. The student pins a vintage, builds the panel on GMU infrastructure, and defends a parallel-trends or matched-design identifying assumption.
Sun & Gao (2019, PNAS) is the anchor paper across this thread (Mentor Session 4, Lab 4, Capstone 1): same-sex-borrower lending studied with a design precisely because HMDA-family data lacks the score.

# Data Card — HMDA (Home Mortgage Disclosure Act Loan Application Register) ## Provider & what it is The **HMDA Loan Application Register (LAR)** is the public, near-census record of mortgage applications in the United States. Under the Home Mortgage Disclosure Act, most lenders must report **nearly every mortgage application they receive** — not just loans made, but ones denied, withdrawn, or fallen through. Per application you get the **action taken** (originated, approved-not-accepted, denied, withdrawn, closed-for-incompleteness), the **loan amount**, the **property location** (state, county, census tract), and **applicant demographics** (race, ethnicity, sex, income). Since 2018 you also get a richer file: interest rate, points, debt-to-income, loan-to-value, property value, and more. The provider is the **CFPB** (Consumer Financial Protection Bureau), which runs the modern public **HMDA Data Browser**, with the **FFIEC** (Federal Financial Institutions Examination Council) publishing the per-year bulk files. It is one of the **largest public microdatasets in U.S. finance** — tens of millions of records per year. The reveal is also the limitation the whole fair-lending literature turns on: **HMDA records who applied, where, for how much, and what happened — but not the applicant's credit score.** It can show you *that* denial rates differ by race; it cannot tell you whether that gap is discrimination or unseen differences in creditworthiness. That single missing column is why fair-lending work leans on *design*, not more controls (Ch 6.4; Sun & Gao, 2019). ## Coverage - **Years:** modern Data Browser coverage from **2007** forward, with the **2018 schema expansion** (Dodd-Frank/Regulation C) as a hard break: pre-2018 and post-2018 files have different fields and codings. The new fields (rate, DTI, LTV, property value, NMLS id) start in 2018. - **Universe:** depository and non-depository lenders above a reporting threshold; smaller lenders are exempt, so HMDA is *near*-census, not total. The threshold has changed across years — `[CHECK]` the exact loan-count threshold for your pinned vintage, because it affects which small lenders appear. - **Size:** multiple gigabytes per nationwide year. **Never download the full national LAR to a laptop.** - **Revisions:** a year is published, then re-published with corrections. The vintage matters; pin it. ## Key identifiers - **Geography:** the county is the **5-digit FIPS** code; the tract is the **11-digit census-tract** code; plus the two-letter **state**. These are your panel keys for aggregating to county-year or tract-year. - **Lender:** the institution is identified by **LEI** (Legal Entity Identifier) in the modern file, replacing the older **respondent ID + agency code** pair used pre-2018. - **Action taken:** the coded outcome (1 = originated, 2 = approved-not-accepted, 3 = denied, 4 = withdrawn, 5 = closed-incomplete, …). The denial-rate outcome is built from these codes. - There is **no person identifier** — HMDA is application-level, not borrower-panel; you cannot follow an individual across applications. ## Access path Two disciplined routes, both **free, no key**: 1. **CFPB Data Browser API** — best for a handful of states/years. The **aggregation endpoint** lets the server count actions by your chosen dimensions so you never move the microdata: base `https://ffiec.cfpb.gov/v2/data-browser-api/`. There is also a filtered-records endpoint for pulling raw rows when you must. 2. **FFIEC bulk files** — best for the whole country. Per-year pipe-delimited LAR files; download once to the **camp container's read-only data area** and aggregate in chunks with `pyarrow`/`pandas`. The container mirrors these so you are not re-downloading gigabytes. ```python import requests, pandas as pd # CFPB HMDA Data Browser AGGREGATION endpoint (public, no key). # Ask the server to count actions by state, so the multi-GB microdata never moves. HMDA_AGG = "https://ffiec.cfpb.gov/v2/data-browser-api/view/aggregations" def fetch_hmda(states, year): params = {"years": year, "states": ",".join(states), "actions_taken": "1,3"} # 1 = originated, 3 = denied r = requests.get(HMDA_AGG, params=params, timeout=60); r.raise_for_status() df = pd.json_normalize(r.json()["aggregations"]); df["year"] = year return df # raw = fetch_hmda(["VA","MD","NC"], 2021) # run on the container, pinned vintage ``` **Aggregate server-side whenever you can.** CFPB pre-counts by state × action (and finer geographies), so pull *counts*, not microdata, and build the denial-rate panel from those. `[CHECK]` exact parameter names against the live API docs for your pinned vintage — CFPB has renamed fields across releases, so the snippet is the *pattern*, not a stable contract. ## License & note **Free and public.** HMDA LAR is published by the FFIEC/CFPB as public data; cite the CFPB HMDA Data Browser (and the FFIEC bulk file if used) plus your pinned vintage and access date. No key, no redistribution restriction. The non-negotiable obligation: **name the vintage**, because the API serves whatever snapshot CFPB currently hosts and the same query returns different numbers across re-publications. Per CONVENTIONS §5, the gigabyte-scale bulk files stay on GMU infrastructure; you aggregate *down* first. ## Gotchas - **No credit score.** The single most important caveat. A raw denial-rate or pricing gap by race confounds discrimination with unobserved creditworthiness (Week 2 OVB). This is *why* Sun & Gao (2019) reached for a clean design, and why Bartlett et al. and BHR are notable for getting at the score from outside HMDA (Ch 6.4). - **The 2018 schema break.** Pre- and post-2018 files differ in fields and codings. Do not splice a pre-2018 series to a post-2018 series without reconciling the schema — the "new" rate/DTI/LTV fields simply do not exist before 2018. - **It's enormous.** Never pull the national file to a laptop; aggregate server-side or in chunks on the container. - **Thin cells are noise.** A rural county with 40 applications has a wildly swinging denial rate. Filter out cells below a minimum decided-application count (e.g., 100) so sampling noise in tiny geographies does not dominate the panel. - **The denominator is a choice you must state.** "Denied / (originated + denied)" excludes withdrawn and incomplete applications — defensible, but a referee will ask. Write it into the spec (CONVENTIONS §4). - **Vintage drift.** Quote no figure without naming the snapshot it came from. `[CHECK]` the exact snapshot date on the camp container before citing a number. ## First 10 rows (illustrative schema sketch — not real values) *Aggregation response (CFPB Data Browser), shape only — counts, not microdata:* | state | actions_taken | count | year | |---|---|---|---| | VA | 1 | 412903 | 2021 | | VA | 3 | 51277 | 2021 | | MD | 1 | 298140 | 2021 | | MD | 3 | 44011 | 2021 | | … | … | … | … | *Raw LAR record (post-2018 filtered pull), shape only — one row per application:* | activity_year | lei | state_code | county_code | census_tract | action_taken | loan_amount | derived_race | applicant_income | interest_rate | |---|---|---|---|---|---|---|---|---|---| | 2021 | EXAMPLELEI000000000 | VA | 51059 | 51059482101 | 3 | 325000 | Black or African American | 78 | NA | | 2021 | EXAMPLELEI000000000 | VA | 51059 | 51059482102 | 1 | 410000 | White | 142 | 3.125 | | … | … | … | … | … | … | … | … | … | … | *Derived county-year denial-rate panel (what you actually regress on), shape only:* | county | year | denied | apps_decided | denial_rate | |---|---|---|---|---| | 51059 | 2021 | 1842 | 19204 | 9.59 | | 51013 | 2021 | 1130 | 14072 | 8.03 | | … | … | … | … | … | ## Which chapter / lab / capstone uses it - **Week 4, Lab 4 — Clean DiD on HMDA + a state policy shock.** The lab's Path A builds exactly the county-year denial-rate panel above from the Data Browser API, then runs TWFE → event study → Callaway–Sant'Anna against clean controls. The card's aggregation pattern *is* the lab's recipe. - **Week 6, Ch 6.4 — fair lending in the algorithmic era** (Bartlett, Morse, Stanton & Wallace; Bhutta, Hizmo & Ringo). HMDA is the spine of both papers, and the chapter's whole pivot is what HMDA *cannot* see (the credit score) and how those authors get it from outside HMDA. - **Capstone 1 — Fair Lending on HMDA.** The terminal use: a real disparity-plus-clean-design paper. The student pins a vintage, builds the panel on GMU infrastructure, and defends a parallel-trends or matched-design identifying assumption. - **Sun & Gao (2019, *PNAS*)** is the anchor paper across this thread (Mentor Session 4, Lab 4, Capstone 1): same-sex-borrower lending studied with a *design* precisely because HMDA-family data lacks the score.