Lending Club Credit Study

Did the Rate Cover the Risk?

Credit losses, realized investor returns and a profit-based approval policy across 2.26 million Lending Club loans

Will Edwards, PhDOctober 2026Code on GitHub

Executive summary

Results

    About the data

    Lending Club was the largest US marketplace lender of the 2010s. It graded each borrower from A to G, set an interest rate for each of 35 sub-grades, and sold the loans to investors. Its public loan file covers every loan issued from 2007 through 2018 ( loans, ), with each loan's outcome as of March 2019, plus rejected applications.

    A lender's pricing is a bet that each grade's interest rate will more than cover its losses. This study checks that bet loan by loan: how losses built up by vintage, what investors actually earned in each grade once defaults and recoveries came in, and whether a model built only on information available at application could have picked out the loans that lost money.

    1Lending volume and loss timing

    Figure 1. Loans funded per year ($ billions), with the share of applications approved
    Figure 2. Share of loan dollars lost to default, by months since the loan was made (36-month loans)

    2Realized returns by grade

    The matured book is every 36-month loan issued from 2010 through 2015 ( loans, ), all of which had reached their final status by the data cutoff. For each loan, monthly cash flows are rebuilt from its installment, the total it paid, and when it stopped paying. Recoveries arrive six months after charge-off, and the investor pays Lending Club's 1% servicing fee. The realized return is what investors actually earned per year, like an interest rate after losses.

    Figure 3. Interest rate charged vs. realized annual return to investors, by sub-grade (36-month loans issued 2010–2015)
    Interest rate chargedRealized annual return (IRR)
    Sub-grades with fewer than 500 matured loans are omitted. The gap between the lines is the share of the rate eaten by defaults, net of recoveries and fees.
    Table 1. Matured book by grade
    Charged off: share of loans that defaulted. Net loss: unrecovered principal as a share of dollars lent. Realized return: annual IRR to investors after defaults, recoveries and the 1% servicing fee.

    3Screening out money-losing loans

    The model learned from loans made in 2013 and 2014, then was asked to score every 2015 loan it had never seen, using only what Lending Club knew on the day each application came in. Because every 2015 loan has since been paid off or charged off, its predictions can be checked against what actually happened, in dollars.

    Figure 4. The riskiest tenth of loans lost money: lifetime return to investors on 2015 loans, grouped by the model's risk score
    Figure 5. Lifetime return within each grade, split into fifths by the model's risk score
    Figure 6. What turning down the riskiest loans would have done to investor profit on 2015 loans
    Ranked by the modelModel without Lending Club's gradeRanked by Lending Club's grade alone
    Model performance and key inputs
    Table 2. Out-of-time performance on 2015 loans
    Figure 7. Inputs with the largest influence on predicted default: mean absolute SHAP value (log-odds) across 20K 2015 loans

    The model is LightGBM, trained on 36-month loans issued January 2013 to September 2014 ( loans), tuned on the fourth quarter of 2014, and scored once on all 36-month loans issued in 2015 ( loans, ). Inputs are credit bureau attributes, income, debt-to-income, employment, loan purpose and state. It never sees payment history or later credit pulls.

    4Data and methods

    Data and definitions

    Lending Club accepted and rejected loan files, 2007–2018Q4, as published on Kaggle (wordsforthewise/lending-club), with loan status as of March 2019. A loan is bad if it was charged off, in default, or 31–120 days late at the cutoff. The file gives the last payment month but not the charge-off date, so charge-off is placed five months after the last payment, matching Lending Club's 150-day charge-off rule. Vintage curves count a loan at month m only if it had been on the books for at least m months by the cutoff.

    Returns

    Each loan's payments are spread as regular installments up to its last payment month, scaled to the total actually received, with any excess treated as a prepayment in the final month. Recoveries net of collection fees arrive six months after charge-off. Pool IRRs solve for the monthly rate that equates pooled cash flows with dollars lent and are annualized. Lifetime return in Section 3 is total cash received minus dollars lent, divided by dollars lent, without discounting.

    Modeling and policy

    Validation is out-of-time: the model never sees 2015 loans while training or tuning. Bureau fields that Lending Club only began reporting in late 2015 are excluded. The approval policy declines the riskiest k% of 2015 loans under each ranking and sums the realized profit of the rest. Because Lending Club's sub-grade has only 35 values, ties within the marginal sub-grade are broken at random. SHAP values are LightGBM's built-in tree contributions.