A Variable Neighborhood Search Algorithmic Approach for Estimating MDHMM Parameters and Application in Credit Risk Evaluation for Online Peer-to-Peer (P2P) Lending

Abstract

Online peer-to-peer (P2P) lending is a new financing channel on which lenders are matched with borrowers using internet platform. Borrowers can get financing more easily, but it means higher credit risk to lenders, making credit scoring models a key tool for lending P2P platforms. The goal is to estimate the level risk (being good or bad borrower), from the collected informations of each applicant. One of the classification approaches is Multi dimensional Hidden Markov Model (MDHMM). The MDHMM parameters are usually estimated using Baum-Welch algorithm (BW). However, the Baum-Welch algorithm tends to arrive at local optimal points. In this paper a strategy called Variable neighborhood search (VNS), is proposed to addresses this problem. The hybrid model in which VNS algorithm is coupled with Baum-Welch algorithm for parameter estimation of MDHMM, is applied in credit scoring domain, using real peer-to-peer lending data. The experiments results show the performance efficiency of our model in comparison with classical and alternative machine learning models for credit scoring.