Modelling loss given default in peer-to-peer lending using random forests

Abstract

Modelling credit risk in peer-to-peer (P2P) lending is increasingly important due to the rapid growth of P2P platforms’ user bases. To support decision making on granting P2P loans, diverse machine learning methods have been used in P2P credit risk models. However, such models have been limited to loan default prediction, without considering the financial impact of the loans. Loss given default (LGD) is used in modelling consumer credit risk to address this issue. Earlier approaches to modelling LGD in P2P lending tended to use multivariate linear regression methods in order to identify the determinants of P2P loans’ credit risk. Here, we show that these methods are not effective enough to process complex features present in P2P lending data. We propose a novel decision support system to LGD modeling in P2P lending. To reduce the problem of overfitting, the system uses random forest (RF) learning in two stages. First, extremely risky loans with LGD = 1 are identified using classification RF. Second, the LGD of the remaining P2P loans is predicted using regression RF. Thus, the non-normal distribution of the LGD values can be effectively modelled. We demonstrate that the proposed system is effective for the benchmark of P2P Lending Club platform as other methods currently used in LGD modelling are outperformed.