Predicting Profitability of Peer-to-Peer Loans with Recovery Models for Censored Data

Abstract

Peer-to-peer lending is a new lending approach gaining in popularity. These loans can offer high interest rates, but they are also exposed to credit risk. In fact, high default rates and low recovery rates are the norms. Potential investors want to know the expected profit in these loans, which means they need to model both defaults and recoveries. However, real-world data sets are censored in the sense that they have many ongoing loans, where future payments are unknown. This makes predicting the exact profit in recent loans particularly difficult. In this paper, we present a model that works for censored loans based on monthly default and recovery rates. We use the Bondora data set, which has a large amount of censored and defaulted loans. We show that loan characteristics predicting lower defaults and higher recoveries are usually, but not always, similar. Our predictions have some correlation with the platform’s model, but they are substantially different. Using a more accurate model, it is possible to select loans that are expected to be more profitable. Our model is unbiased, with a relatively low prediction error. Experiments in selecting portfolios of loans with lower or higher Loss Given Default (LGD) demonstrate that our model is useful, whereas predictions based on the platform’s model or credit ratings are not better than random.