LendingClub Loan Data (Kaggle Mirror)

The LendingClub Loan Data (Kaggle Mirror) preserves the loan-level record of LendingClub's peer-to-peer personal lending platform: about 2.26 million accepted loans issued from June 2007 through 2018 Q4, plus roughly 2.76 million rejected applications from the same window. Accepted loans carry the full LoanStats schema - 151 documented fields spanning origination terms, borrower attributes, credit-file snapshots and settlement outcomes, each defined in the lender's own official data dictionary and verified during our research pass.

Two design features make this the reference panel for consumer-credit modeling. First, closed-loop outcomes: every loan resolves to Fully Paid, Charged Off, Default or Current, with payments received and post-charge-off recoveries on the same row, so probability-of-default and loss-given-default models train on one single already-joined table. Second, an expert baseline built in: the LC-assigned grade ladder, A through G, subdivided into subgrades A1-G5, sits beside the interest rate actually charged. On the Consumer Finance shelf it scores 7/10 on our quality rubric against a catalog average of 7.81 across all 1,744 datasets we hold.

One row per loan application, wide and flat: origination terms up top, borrower attributes beneath them, outcome columns resolving to the right as the loan seasons. Two formatting conventions matter before you parse anything. Percent signs were stripped from int_rate and revol_util at the file level - 11.44 means 11.44 percent - and term collapses to a bare integer, 36 or 60, never anything else.

The block below reproduces one accepted loan end to end, using values exactly as they appear in the file. Origination facts first, the bureau-derived block in the middle, and the settlement trio - payments received, recoveries, final status - closing the row.

Think of the 151 columns as four blocks. The terms block fixes the deal: amount applied for, 36- or 60-month term, interest rate, monthly installment. The borrower block describes who asked: employment length, home ownership, self-reported annual income, verification status, debt-to-income ratio, and the FICO range pulled at origination. The credit-history block adds recent delinquencies and revolving utilization. The assignment block records the platform's judgment - grade and sub-grade - while the outcome block settles the story with total payments received, post-charge-off recoveries and final loan status.

The dictionary below lists the twenty-one columns that do most of the modeling work. The remaining payment and settlement fields, and the whole nine-column rejected-application sheet, fold out with your sample.

Geographically this is a national book: borrowers from every U.S. state appear, with the state code reported by the borrower on the accepted side and three-digit zip codes on the rejected side. Temporally, originations run June 2007 through 2018 Q4 - eleven full calendar years of unsecured consumer lending. The snapshot is frozen at version 3, dated April 2019, which is precisely what makes it useful: a fixed laboratory where no later print can revise your backtest.

Granularity is one row per loan application. Accepted loans additionally accumulate monthly performance and settlement fields on that same row, so a vintage table needs no reshaping before it groups.

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

You pick the channel and the cadence; the field dictionary above travels unchanged across all three. Full-history loads suit a team calibrating a model once against the whole 2007-2018 book. Cohort-scoped feeds suit dashboards and feature stores tracking one grade band, one state, or one purpose code. Warehouse delivery lands typed one-row-per-loan tables next to your own portfolio data so the joins stay local. Cadence changes are a settings conversation, not a re-integration project.

• Default-prediction benchmarks. Train probability-of-default models against terminal loan_status labels and score every challenger against the LC grade baseline.
• Loss-given-default and recovery work. Payments received and post-charge-off recoveries settle each charged-off loan on the same row, with no outside servicing data needed.
• Vintage and cohort analysis. issue_d slices the book into monthly vintages; grade, purpose and state cut each vintage thinner.
• Reject inference and policy research. The rejected-application book pairs amount requested, risk score, DTI and geography for loans that never funded.
• Rate-setting forensics. Interest rate against sub-grade quantifies, in basis points, what one notch of assigned risk was worth at origination.

Data Scientists & ML Engineers get the cleanest classification target in consumer credit: 151 candidate features, a labeled outcome, a benchmark everyone recognizes (their use cases). Investors & Quants rebuild marketplace-lending economics - rate setting, grade migration, realized losses - origination to settlement (their use cases). Market Researchers & Consultants read borrowing behavior straight from stated purposes, income bands and state distributions instead of survey panels. Developers & Data-Product Builders prototype credit products against realistic loan shapes - terms, installments, outcomes - before wiring production systems. Journalists, Academics & Students cite actual loan-level outcomes, millions of them, instead of anecdote.

Four honest caveats. First, the book is frozen: nothing originates after 2018 Q4, so treat this as a fixed laboratory for backtesting and benchmarking, and pair it with a live source when current-vintage work matters. Second, formatting ships as-is: percent signs are gone from int_rate and revol_util, and term is a bare 36-or-60 integer - parse types deliberately rather than trusting defaults. Third, label maturity: a loan issued in mid-2018 cannot yet read Fully Paid or Charged Off, so outcome-ready cohorts run earlier than origination-ready ones, and mixing the two inflates apparent performance. Fourth, asymmetric halves: the rejected side carries nine descriptive fields and no outcome, ever - use it for reject inference and policy research, never for loss modeling. None of these bite unexpectedly; they ship flagged against the analysis you plan to run.

Because the raw artifact is a multi-part archive of period-split compressed CSVs behind a 151-column header, assembled for interactive notebooks rather than pipelines. Datadory reshapes it into typed, one-row-per-loan tables with the data dictionary attached, cuts the cohort you name - grade band by issue years by state by status - and keeps consecutive deliveries reconcilable in your warehouse. Browse the rest of the shelf on the Consumer Finance data hub, see how the surveyed side complements it in the Federal Reserve Survey of Consumer Finances, or read the Kaggle source profile for the wider mirroring catalog.