Predicting Default Risk on Peer-to-Peer Lending Imbalanced Datasets

• Published in: IEEE access Vol. 9; pp. 73103 - 73109
• Main Authors: Chen, Yen-Ru, Leu, Jenq-Shiou, Huang, Sheng-An, Wang, Jui-Tang, Takada, Jun-Ichi
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01-01-2021; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Datasets; Default; Fraud; History; imbalanced datasets; Loans; Machine learning; Measurement; Neural networks; Peer to peer computing; Peer to peer lending; Random forests; re-sampling; Risk; Small business
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2021.3079701

In the past few years, Peer-to-Peer lending (P2P lending) has grown rapidly in the world. The main idea of P2P lending is disintermediation and removing the intermediaries like banks. For a small business and some individuals without enough credit or credit history, P2P lending is a good way to apply for a loan. However, the fundamental problem of P2P lending is information asymmetry in this model, which may not correctly estimate the default risk of lending. Lenders only determine whether or not to fund the loan by the information provided by borrowers, causing P2P lending data to be imbalanced datasets which contain unequal fully paid and default loans. Imbalanced datasets are quite common in the real worlds, such as credit card fraud in transactions, bad products in the plant and so on. Unfortunately, the imbalanced data are unfriendly to the normal machine learning schemes. In our scenario, models without any adaptive methods would focus on learning the normal repayment. However, the characteristic of the minority class is critical in the loaning business. In this study, we utilize not only several machine learning schemes for predicting the default risk of P2P lending but also re-sampling and cost-sensitive mechanisms to process imbalanced datasets. Furthermore, we use the datasets from Lending Club to validate our proposed scheme. The experiment results show that our proposed scheme can effectively raise the prediction accuracy for default risk.