Machine Learning versus Traditional Statistical Models in Credit Risk Prediction: Evidence from Peer-to-Peer Lending Markets

Abstract

This paper compares the reliability, transparency, and fairness of traditional statistical methods applied to credit risk assessment, especially logistic regression. Traditional methods dealing with data management are typically inefficient, not accurate enough, and cannot deal easily with multiple datasets, while those based on machine learning face model selection problems and multicollinearity. This study aims to provide financial institutions with some practical guidance on the optimal strategies they should implement according to their risk-management needs. It also examines the impact of integrating heterogeneous and unstructured sources of data on the machine learning performance of credit risk models. Particular attention is dedicated to peer-to-peer lending markets, for which we are not aware of established research that jointly investigates the use of classical and machine learning models. The research follows a deductive approach and uses inferential methods in the evaluations and comparisons of logistic regression vis-à-vis the neural network, specifically a convolutional neural network. Model building and validation are based on the Kaggle peer-to-peer lending data as a secondary source. Expected results are as follows: accurate borrower default prediction, increased access to credit, and smart lending decisions. The importance of the finding is that it shows, in practical terms, how machine learning can be used to improve portfolio management and risk assessment for contemporary financial institutions with an interesting, more advanced analytic method.