National Mortgage Database Aggregate Statistics: What FHFA Publishes and How to Use It

They are the only nationally representative view of the entire US first-lien mortgage universe that comes without a microdata access fight. The FHFA National Mortgage Database (NMDB) Aggregate Statistics derive from a five percent national sample of closed-end first-lien residential mortgages, jointly developed by FHFA and the Consumer Financial Protection Bureau and maintained by FHFA to satisfy the monthly mortgage market survey mandate in section 1324(c) of the 1992 Safety and Soundness Act as amended by HERA 2008.

Three named products ship from it: New Residential Mortgage Statistics (originations), Outstanding Residential Mortgage Statistics, and Residential Mortgage Performance Statistics. Each row is weighted so the 5% sample represents the full market, which is why a single state-year cell can carry an origination count for every mortgage closed there rather than just the sampled ones. Every product also has a companion interactive dashboard - released June 26, 2026 - that builds, analyzes and exports custom charts without touching a CSV.

Start dates differ by product, and pipelines break when they assume otherwise. New-mortgage statistics reach deepest: annual values for 1998 through 2024, quarterly values from 1998 Q1 through 2025 Q3, and a monthly cut from January 1998 through September 2025. Outstanding residential mortgage statistics begin at 2013 Q1, and performance statistics at 2002 Q1 - both running forward through 2026 Q1 as of the 6/26/2026 release.

That staggered history matters for panel design. A delinquency study reaching the early 2000s can use performance tables all the way back to 2002 Q1, but any balance-based question before 2013 Q1 has no NMDB answer; the Federal Reserve's Z.1 mortgage debt series, verified back to 1945, is the fallback for long-run holder-type totals. The monthly new-mortgage series ends September 2025 in the current release, and whether it is discontinued or merely lagging the annual and quarterly cuts is an open question worth confirming with FHFA (NMDBdata@fhfa.gov) before you automate against it.

Geography runs four levels deep plus a rural split. Tables cover the nation, Census regions and divisions, all states, and - for performance statistics only - the 100 largest metro areas, with rural/non-rural splits also published. If your question needs county or census-tract granularity, NMDB will not answer it; that work belongs to the tract-level FHFA Public Use Database.

Every zip unpacks to one long-format CSV with exactly 14 standardized columns - SOURCE, FREQUENCY, GEOLEVEL, GEOID, GEONAME, MARKET, PERIOD, YEAR, QUARTER, MONTH, SUPPRESSED, SERIESID, VALUE1 and VALUE2 - documented field-by-field in the Data Dictionary and Technical Notes PDF dated 06/26/2026. The SERIESID column carries the actual statistic, with documented examples including TOT_ORIG (originations in thousands), TOT_LOANAMT, AVE_LOANAMT, AVE_INTRATE, PCT_FTHB, PCT_CASHOUT, PCT_ARM, AVE_DTI, AVE_VANTAGESCR and AVE_LTV.

VALUE1 holds the primary weighted statistic; VALUE2 holds a secondary value where applicable, such as a standard error or unweighted count, though its exact meaning varies per SERIESID and should be checked against the dictionary appendix before you compute intervals around it. The SUPPRESSED flag marks cells withheld for borrower privacy, so treat suppression patterns themselves as signal about thin segments - small geographies crossed with narrow credit bands suppress first. Sample rows show the shape immediately: Alaska, Annual, State, All Mortgages, 1998, TOT_ORIG, VALUE1 = 26.

Each federal source answers a different slice of the mortgage question, and picking wrong wastes a week. NMDB is the only one weighted to represent the whole closed-end first-lien universe including bank and nonbank loans outside the Enterprises, and it does so from 14 standardized columns per file rather than a modeling tape. HMDA covers applications as well as originations but only from covered lenders; the FHFA FHFA Public Use Database (PUDB) reaches census tracts with banded demographics but counts Enterprise purchases alone across 2008-2024; and the Freddie Mac tape goes to loan-month performance depth for its own roughly 56 million loans only. Our side-by-side of the two FHFA products works through that split in detail.

Third, series definitions: VALUE1 carries the weighted statistic and VALUE2 a secondary value whose meaning varies by SERIESID, and both are defined only in the Data Dictionary and Technical Notes PDF dated 06/26/2026 - read the appendix for each series you use rather than guessing. Fourth, open questions: whether the monthly new-mortgage cut ending September 2025 is discontinued or lagging should be confirmed with FHFA before automating against it.

Pairing fills what aggregates leave out. The FHFA House Price Index (HPI) supplies repeat-sales collateral values down to census tracts from January 1975 (90,000+ observations in its master file), which turns NMDB's AVE_LTV bands into current equity estimates, while American Housing Survey (AHS) National Microdata contribute household-level mortgage terms across 16,834 mortgage records in the 2023 National PUF. Worked setups live in our market sizing playbook, and economist-and-analyst applications are collected on the data scientists page for this industry.

Start with the commercial residential mortgage finance data guide, the pillar that sorts all ten primary datasets - four of them from FHFA - into four working groups. Then go one level deeper on the neighbors this page leans on: the loan level mortgage performance dataset guide for when aggregates are too coarse, and the house price index by zip code guide for the collateral side of LTV work.