Credit risk modeling on data with two timestamps in peer-to-peer lending by gradient boosting

Abstract

It is extremely important for a Peer-to-Peer (P2P) lending platform to provide their lenders with a reliable model to predict borrowers’ credit risk. This study introduces four different gradient boosting methods to model the credit risk of borrowers in P2P platforms and estimates the performance of models in a more practical scenario. The timeliness of observations and features is fully investigated in modeling. Each observation is associated with two timestamps that guide the splitting of data set for training and testing models, while no timestamp or only one timestamp was used in most previous research. The hyperparameters of gradient boosting models are optimized by the Bayesian approach, and the models are evaluated in different scenarios. The importance of features implied by gradient boosting models is examined as well. The empirical result shows that the models’ performance from k−fold cross validation is overestimated in comparison to that from rolling windows guided by two timestamps of observations, which indicates that the timestamps plays a significant impact on evaluating the performance of credit risk models in P2P lending.