DeepCredit: Exploiting user cickstream for loan risk prediction in P2P lending

Peer-to-peer (P2P) lending or crowdlending, is a recent innovation allows a group of individual or institutional lenders to lend funds to individuals or businesses in return for interest payment on top of capital repayments. The rapid growth of P2P lending marketplaces has heightened the need to develop a support system to help lenders make sound lending decisions. But realizing such system is challenging in the absence of formal credit data used by the banking sector. In this paper, we attempt to explore the possible connections between user credit risk and how users behave in the lending sites. We present the first analysis of user detailed clickstream data from a large P2P lending provider. Our analysis reveals that the users' sequences of repayment histories and financial activities in the lending site, have significant predictive value for their future loan repayments. In the light of this, we propose a deep architecture named DeepCredit, to automatically acquire the knowledge of credit risk from the sequences of activities that users conduct on the site. Experiments on our large-scale real-world dataset show that our model generates a high accuracy in predicting both loan delinquency and default, and significantly outperforms a number of baselines and competitive alternatives.