Abstract

This research presents a novel two-step model leveraging machine learning techniques to predict loan defaults in peer-to-peer (P2P) lending, utilizing data from the Bondora P2P lending platform. The first step involves constructing a graph of loans to extract graph features: pagerank, betweeness, closeness, eigenvector, katz, authority, and hub—thus expanding the pool of information beyond traditional credit risk factors. In the second step, we employ three distinct machine learning models: elastic net regression, random forest, and a deep neural network, to exploit these features. Subsequently, a feature importance analysis is conducted to assess the contribution of the seven centrality measures to the model performance. This study not only offers significant contributions to the literature on P2P lending default prediction but also provides practical insights for industry stakeholders. We suggest further research into other network-based features, alternative machine learning models, and possible applications of our method to other financial industry domains.