PEER-TO-PEER LOAN FRAUD DETECTION: CONSTRUCTING FEATURES FROM TRANSACTION DATA

Although financial fraud detection research has made impressive progress because of advanced machine learning algorithms, constructing features (or attributes) that can effectively signal fraudulent behaviors remains a challenge. In recent years, a new type of fraud has emerged on peer-to-peer (P2P) lending platforms, where individuals can borrow money from others without a financial intermediary. In these markets, the information asymmetry problem is seriously elevated. Inspired by the fraud triangle theory and its extensions, and using the design science research methodology, we construct five categories of behavioral features directly from P2P lending transaction data, in addition to the baseline features regarding borrowers and loan requests. These behavioral features are intended to capture the fraud capability, integrity, and opportunity of fraudsters based on their loan requests and payment histories, connected peers, bidding process characteristics, and activity sequences. Using datasets from real users on two large P2P lending platforms in China, our evaluation results show that combining these additional features with the baseline features significantly enhances detection performance. This design science research contributes novel knowledge to the financial fraud detection literature and practice.