Evaluation of cost-sensitive machine learning methods for default credit prediction

Consumer Peer-to-Peer lending market has grown rapidly during the recent years. This alternative finance business model gives many advantages for all participants - investors, borrowers, and internet-based Peer-to-Peer lending platforms. However, the risk arising from defaulted credits should not be underestimated as this is one of the main concerns for investors and platforms. Machine learning algorithms are successfully used to predict the credit risk in the process of credit approval, but for better decision making they should consider the highly imbalanced nature of the data and misclassification costs associated with wrong predictions.The purpose of this research paper is to explore different techniques for dealing with imbalanced datasets and different misclassification costs related to the target classes. We propose a set of measures to evaluate model performance both in term of accuracy and profit optimization. Our approach is implemented on real dataset from the Lending Club company. We use case-based data about the profit and loss for every loan and calculate more precisely the misclassification costs. Cost-sensitive models are trained using extremely gradient boosted decision trees (XGBoost) with hyperparameters optimization. Empirical results show that cost sensitive XGBoost model with weights derived from the cost matrix gives and threshold adjustment best results both in terms of accuracy and profit optimization.