Freddie Mac Single-Family Loan-Level Dataset

The Freddie Mac Single-Family Loan-Level Dataset is the Commercial & Residential Mortgage Finance catalog's deepest credit record: every mortgage Freddie Mac purchased or guaranteed from January 1, 1999 through March 31, 2026 (Release 47) - approximately 56 million loans - disclosed at loan level with a monthly performance trail behind each one. This is the guarantor's own ledger, not a survey of it.

The record runs on two grains. An origination file carries one record per loan per vintage quarter: the underwriting snapshot of FICO, debt-to-income, loan-to-value and combined LTV, note rate, channel, occupancy, property type and the seller that delivered the loan. A performance file carries one record per loan per reporting period: balance and rate as they actually stood, delinquency status, loan age, modification flags, zero-balance codes when a loan exits, and - at disposition - the full loss anatomy down to net sales proceeds, recoveries and itemized expenses.

The population splits into a Standard book resembling credit-risk-transfer eligibility (fully amortizing fixed-rate mortgages, full documentation, Relief Refinance and Home Possible loans funded on or after March 1, 2015, government-insured loans excluded) and a Non-Standard book holding the remainder. Definitions were verified field-by-field against the published July 2026 file layout during the August 2026 research pass. Get a sample of this dataset cut to the vintages and states you care about.

One origination row per loan per vintage quarter, one performance row per loan per month. Three rows exactly as they ship:

# origination record - 2000 vintage, one row per loan per vintage quarter
loan_identifier      : F26Q50000001   # consistent across origination and performance files
first_payment_date   : 200007         maturity_date    : 203006    # 360-month term
original_upb         : 253000.00      original_rate    : 7.875     # note rate at acquisition
original_ltv         : 65             original_cltv    : 65        dti_ratio : 21
occupancy_status     : S              # second home
property_state       : OR             property_type    : CO        # Oregon condominium
amortization_type    : FRM            loan_purpose     : P         # purchase
channel              : T              # third-party / broker
first_time_homebuyer : N              mi_percentage    : 0         super_conforming : N
seller_name          : WASHINGTON MUTUAL BANK

# origination record - the 2000-era credit box in two ratios
loan_identifier      : F26Q50000002   occupancy_status : I         # investment property
original_upb         : 122000.00      original_rate    : 9.250     cltv : 80    dti_ratio : 44
property_state       : IL             property_type    : CO        amortization_type : FRM
loan_purpose         : P              channel          : R         # retail
seller_name          : BANK OF AMERICA, N.A.

# monthly performance record - one row per loan per reporting period
loan_identifier      : F26Q60000001   period           : 200009
current_actual_upb   : 52000.00       current_rate     : 9.250
delinquency_status   : 00             # current on schedule
loan_age             : 0              months_to_maturity : 360
non_interest_bearing_upb : 0.00       actual_loss      : 0.00      # stands until disposition

Read together, the record's logic shows up plainly. The first row is a 2000-vintage Oregon condominium bought as a second home: $253,000 at 7.875 percent, 65 LTV and CLTV alike (no subordinate financing), a debt-to-income ratio of just 21, broker-delivered by Washington Mutual Bank - a super-prime loan in an era when such terms existed alongside everything else. The second row is the same vintage's other face: an Investment-property condominium in Illinois at 9.250 percent with an 80 CLTV and a 44 debt-to-income ratio, retail-channelled, sold by Bank of America - the looser underwriting the loss years later made famous, sitting in the same schema as its conservative neighbor. The third row is the performance side: a loan current on schedule (00), balance and rate restated for the period, loan age zero, and actual_loss standing at 0.00 - a column that stays flat until a disposition ever writes something into it.

Live rows land in this identical shape - name the states, vintage quarters and Standard-versus-Non-Standard cut when you get a sample of this dataset.

Thirty-three documented fields define the record, every definition checked against the source's published file layout effective July 2026. They divide into three groups that map onto a loan's biography:

Underwriting anatomy - CLASSIC FICO, ORIGINAL DEBT-TO-INCOME (DTI) RATIO, ORIGINAL LOAN-TO-VALUE (LTV) and ORIGINAL COMBINED LOAN-TO-VALUE (CLTV), ORIGINAL INTEREST RATE, ORIGINAL UPB, amortization type, loan purpose, channel, occupancy status, property type and state, mortgage insurance percentage, first-time-homebuyer flag, first payment and maturity dates, and SELLER NAME - the covariates every default model wants, present on day one.

The monthly ledger - PERIOD, CURRENT ACTUAL UPB, CURRENT LOAN DELINQUENCY STATUS coded in months delinquent, LOAN AGE, REMAINING MONTHS TO LEGAL MATURITY, CURRENT INTEREST RATE, MODIFICATION FLAG and DUE DATE OF LAST PAID INSTALLMENT (DDLPI) - enough to build delinquency transition matrices and prepayment curves straight from the rows rather than asserting them from aggregates.

Disposition economics - ZERO BALANCE CODE and effective date distinguishing prepayments from short sales, charge offs, REO dispositions, note sales and reperforming-loan sales, then NET SALES PROCEEDS, MI RECOVERIES, NON MI RECOVERIES, TOTAL EXPENSES and ACTUAL LOSS - realized severity, decomposed.

The complete dictionary follows below, one row per field with type, definition and example.

Geography - United States collateral identified by property state, postal code and Metropolitan Statistical Area or metropolitan division, so state-level rollups and metro-level cohort comparisons assemble without any geocoding project.

Temporal - originations from January 1, 1999 through March 31, 2026, with monthly performance disclosed through the same cutoff in Release 47. The window spans the entire modern credit cycle: a 1999-vintage purchase, the 2006-2007 peak book whose losses defined an era, the refinance waves, and the post-2020 rate shock - all in one schema, all at loan grain.

Granularity - one origination record per loan per vintage quarter plus one monthly performance record per loan per reporting period, split Standard versus Non-Standard. There is no aggregation layer between you and the loans: delinquency transitions, prepayment speeds and severity distributions are computed, not quoted.

A sample dataset of 50,000 randomly selected loans per vintage year mirrors the full tape's identical fields for work that does not need all 56 million. Set against the wider Datadory catalog - average quality score 7.81 across 1,744 datasets - this record scores 10/10, carried by exhaustive published documentation, verified field definitions and twenty-seven years of unbroken history.

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

Your cadence is your call regardless of the rhythm of the underlying disclosure. Most teams load their vintages once and append performance periods as they arrive, so each new reporting month diffs cleanly against the last and models retrain without re-plumbing. Deliveries arrive normalized to the field dictionary above, keyed on the loan identifier that holds constant across origination and performance files, as JSON, CSV, spreadsheet-ready Excel tables or XML - the headerless pipe-delimited raw shape handled for you. Scope the extract however the work needs: named states, MSAs or postal codes, chosen vintage quarters, Standard-only, Non-Standard-only, or the 50,000-loan-per-vintage sample instead of the full tape. Every delivery ships the dictionary, the sample rows and the coverage profile mapped to exactly what you asked for.

• Default and prepayment modeling - millions of loan-months of delinquency transitions, modification flags and zero-balance codes make rolling transition matrices and prepayment speed curves an estimation exercise rather than an assertion; see ML model training.
• Loss-given-default calibration - actual loss decomposed into net sales proceeds, MI recoveries, non-MI recoveries and itemized legal, preservation, tax-and-insurance and miscellaneous expenses is the quantity severity models are benchmarked against, on real outcomes rather than assumptions.
• Credit-box drift studies - FICO, LTV, CLTV and DTI distributions by vintage quarter and state quantify exactly when underwriting loosened, by how much, and what it cost.
• Portfolio stress testing - seasoned curves drawn from the whole 1999-2026 cycle give any book that looks like Freddie's a benchmark for what a rate and unemployment path actually did to comparable loans.
• Seller and channel analytics - SELLER NAME and origination channel reveal who fed the guarantee book, in what mix, and how that flipped between purchase and refinance eras.
• Policy and academic research - two decades of GSE credit performance at loan level, citable, reproducible and reconcilable row-by-row against published aggregates.

Traders running rate-path scenarios work the same rows from the other direction; see quant backtesting and credit risk screening.

Investors and quant researchers (relevance 3/3) get the loan-level tape behind agency credit performance - delinquency, prepayment and severity on the book their securities reference; see investors quants use cases. Data scientists and ML engineers (3/3) get a ready-made tabular corpus of roughly 56 million labeled outcomes with clean covariates - the rare dataset where the target variable and the features ship together; see data scientists use cases. Market researchers and consultants (2/3) get vintage-by-state credit-box benchmarks no survey can reconstruct; see market researchers use cases. Journalists, academics and students (2/3) get a guarantor's own record to hold aggregate claims against; see journalists academics use cases. Developers building data products (2/3) get a fixed, documented schema keyed on a stable loan identifier that drops into mortgage-analytics tools; see developers builders use cases. Sales and growth teams find the angle in seller-level volume patterns across vintages.

Within commercial and residential mortgage finance data, this record owns loan-level depth on one guarantor's book: roughly 56 million loans, monthly grain, actual losses decomposed. The neighbors own different jobs. FHFA Public Use Database discloses acquisitions for both Enterprises at census-tract level with borrower demographics - breadth and context where this gives depth, but banded categories instead of continuous underwriting variables. FHFA National Mortgage Database (NMDB) Aggregate Statistics samples the whole first-lien universe including non-GSE lenders yet publishes aggregates only, where this hands over rows. FHFA House Price Index prices the collateral these loans sit on. American Housing Survey (AHS) National Microdata sees households the GSE tapes never touch - FHA, VA and portfolio loans included. If the question is what happened to the loans themselves, this is the record; the two lenses are worked side by side in Freddie Mac loan-level vs AHS national microdata.

The record reads best beside notes that price its collateral and frame its borrowers. Pair it with the FHFA House Price Index to put a collateral-value track under every LTV band, the FHFA Public Use Database for the borrower-demographic view of acquisitions, Ginnie Mae Disclosure Data and Reports for the government-guaranteed counterpart book, and the American Housing Survey for household context. Glossary notes on delinquency rate, conditional prepayment rate and loan-to-value ratio define the quantities the tape computes. Background lives in the Freddie Mac source profile, the commercial and residential mortgage finance data hub and the best commercial and residential mortgage finance datasets ranking.