Towards Profitability: A Profit-Sensitive Multinomial Logistic Regression for Credit Scoring in Peer-to-Peer Lending

Abstract

This paper proposes a profit-sensitive learning method for loan evaluation in the peer-to-peer (P2P) lending market that could provide better investment suggestions for the lenders. Currently, the most widely utilized loan evaluation method is credit scoring, which focuses on evaluating the loans’ defaulting risk and formulates a binary classification problem. It screens out the non-default loans from the default ones and thus defines the best loans as those with a low probability of default (PD). However, the conventional credit scoring totally ignores the profit information while solely focusing on the risk. To address the above issue, we propose a profit-sensitive multinomial logistic regression model that incorporates the profit information into the credit scoring approach. More specifically, we first transform the binary classification problem in traditional credit scoring to a multi-level classification task by further dividing the default loans into two sub-classes: “default and profitable” and “default and not profitable”. Then we design a multinomial logistic regression model with a novel loss function to solve the above-defined multi-level classification task. The loss function weights loans differently according to their varying profits as well as their occurrence frequencies in the real-world practices. The effectiveness of the proposed method is examined by the real-world P2P data from Lending Club. Results indicate our approach outperforms the existing credit scoring only approach in terms of identifying more profitable loans while ensuring the low risk. Therefore, the proposed profit-sensitive learning method serves as an innovative reference when making investment suggestions in P2P lending or similar markets.