A Feature Elimination Machine Learning Model for Credit Assessment and Repayment Behavior Prediction in Marketplace Lending

With the rapid development of the credit industry and the advent of marketplace lending, credit scoring models play a vital role in reducing the risk exposure for lenders. However, traditional credit scoring models like the FICO Score make it hard for people with weak credit history to acquire credit services. Credit scoring models based on machine learning can provide accurate assessments for such thin-credit people, but a lot of private data, like social media activities, are used during the evaluation procedure. In this work, a credit scoring approach with a focus on marketplace lending is proposed that combines machine learning with a novel feature selection method that follows a backward elimination approach. Thus, many irrelevant features are eliminated from the dataset during the feature selection, and private data are not used or remain limited. The model is trained and tested in a large loan dataset available in the public domain. It performs pretty well compared to traditional credit scoring method and can be used to provide credit assessment for thin-credit history individuals without using personal private data. The approach has also explanatory power, as the feature selection approach offers a perspective for understanding how each feature affects individual loan repayment behavior.