Personal Loan Default Prediction Dataset: 2.26M Loans, Free

Start with the honest count: among the 24 primary consumer-finance datasets Datadory catalogs - every one of them free to access - exactly one ships loan-level records with repayment outcomes attached. That record is the LendingClub Loan Data (Kaggle Mirror): roughly 2.26 million accepted personal loans issued between June 2007 and 2018 Q4, alongside about 2.76 million rejected applications, packed into CSV/gz files totalling approximately 1.36 GB zipped.

Accepted-loan rows carry monthly performance and settlement fields, which is what turns a static snapshot into a supervised learning problem: you derive the default indicator from observed payment history rather than trusting a single pre-computed flag. The Kaggle listing declares the package commercial delivery terms: commercial delivery terms, credits creator Nathan George, and carries the tags credit-risk, default-prediction and machine-learning-benchmark.

Everything else in the industry plays a supporting role. Federal Reserve releases contribute macro covariates, household surveys contribute borrower-resilience features, CFPB complaint data contributes servicer-conduct signals, and HMDA contributes application decisions - valuable, but a decision is not a default. Discovery beyond this catalog runs through Data.gov's consumer-finance catalog search: the GSA index spans 552,271 datasets catalog-wide and returns 21 results for the 'consumer finance' query as of 2026-08-21, each linking federal, state and local publishers.

Each row is one loan application, and the package documents 151 fields. The accepted-loan files are where labels live: their monthly performance and settlement fields let you define a charge-off window consistently across nine years of vintages, which is why a frozen snapshot still anchors benchmark leaderboards. The rejected-application files hold roughly 2.76 million rows - enough for approval modelling - but they stop at application attributes and three-digit ZIP prefixes, whereas accepted loans span all US states.

Because performance fields accumulate with loan age, early vintages have had years to mature while 2018 originations have barely any observed history. A naive random split therefore leaks time-to-event information between train and test sets; validation grouped by issuance period is the correction practitioners apply to this corpus.

The industry's recommended recipe pairs the loan file with Federal Reserve aggregates. The Federal Reserve G.19 Consumer Credit release publishes 118 series totalling about 78,000 monthly observations spanning January 1943 to June 2026: revolving and nonrevolving credit outstanding broken out by holder type - depository institutions, finance companies, credit unions, the federal government and nonprofits - plus terms-of-credit rate series reaching back as early as June 1971. Student loan and motor vehicle memo items appear quarterly only.

Joined onto issue months, those aggregates hand the model a credit-cycle variable instead of pretending 2007-2018 was one regime. Logistics are trivial: the full SDMX XML package runs about 6.0 MB zipped, the largest preformatted CSV package 669.8 KB, all commercial delivery terms with no registration.

A second enrichment channel is conduct. The CFPB Consumer Complaint Database indexes roughly 17.24 million complaints (17,242,643 as of 2026-08-21) refreshed daily under commercial delivery terms, and its no-authentication REST API supports field-level filters, aggregations and search_after pagination over pages of 1-100 records - enough to attach servicer-response patterns to a lender dimension.

For balance-sheet depth, the Federal Reserve Survey of Consumer Finances (SCF) ships triennial family microdata from 1989 through 2022. The 2022 public extract holds 22,975 records - 4,595 families times five multiply-imputed records - with replicate weights and a summary extract at 357 variables under DOI doi.org/10.17016/8799. Use the published replicate weights: the catalog notes that ignoring multiple imputation and the complex sample design yields incorrect standard errors.

Banking-status segments come from the FDIC National Survey of Unbanked and Underbanked Households: roughly 30,000 households per biennial wave across eight waves (2009-2023), with 567 documented variables in the 2023 wave.