How can lenders prosper? Comparing machine learning approaches to identify profitable peer-to-peer loan investments

Successful Peer-to-Peer (P2P) lending requires an evaluation of loan profitability from a large universe of loans. Predictions of loan profitability may be useful to rank potential investments. We investigate whether various types of prediction methods and the types of information contained in loan listing features matter for profitable investment. A range of methods and performance metrics are used to benchmark predictive performance, based on a large dataset of P2P loans issued on Lending Club. Robust linear mixed models are used to investigate performance differences between models, according to whether they assume linearity, whether they build ensembles, and which types of predictors they use. The main findings are that: linear methods perform surprisingly well on several (but not all) criteria; whether ensemble methods perform better than individual methods is measure dependent; the use of alternative text-based information does not improve profit scoring outcomes. We conclude that P2P lenders could potentially increase their investment returns by applying linear methods that directly predict the internal rate of return instead of other dependent variables such as loan default.