How magic a bullet is machine learning for credit analysis? An exploration with fintech lending data

J. Christina Wang and Charles B. Perkins

Abstract

Fintech lending to consumers has grown rapidly since the 2007–9 Great Recession. This study applies machine learning (ML) methods to loan-level data from the largest fintech lender of personal loans to assess whether these techniques can produce more accurate out-of-sample default predictions than standard regression models, as fintech advocates claim. To explain loan outcomes, the analysis incorporates the economic conditions faced by borrowers after origination—an element typically absent from other ML studies of default. For the given data, ML methods do improve prediction accuracy, especially over shorter horizons within a year. However, having more data—up to a point—enhances their relative accuracy, likely due to data or model drift over time, which can cause more complex models to misfire out of sample. Adding standard credit variables beyond a core set offers only marginal gains, implying that unconventional data must be sufficiently informative to benefit consumers with limited credit history. Finally, the study finds little statistically significant evidence that ML methods yield unequal benefits across borrower subgroups defined by risk, income, or location.