A Solution to the Problem of Retail Credit Risk Pricing Problem Based on the Machine Learning XGBoost Algorithm

Machine learning algorithms, represented by logistic regression and decision tree algorithms, have a wide range of applications in retail credit risk management, such as anti fraud models, application scoring models, behavioural scoring models and overdue collection scoring models, which are used to assess the credit risk level of customers at different stages of the retail credit process. The integrated decision tree model, represented by the XGBoost algorithm, constructs the model using more variables and although it lacks the necessary interpretability, its predictions are more accurate. Traditional retail risk pricing typically uses a preapplication scoring model to enforce different product prices for customers with different risk ratings. Based on the XGBoost algorithm and personal credit information from the People’s Bank of China and other tripartite credit bureaus, we attempt to develop a retail risk pricing model based on the actual loan disbursement of the customers, which is used to apply different product prices to customers with different risk ratings, while at the same time remarketing to customers who have not disbursed historically. The data for the construction of the retail risk pricing model are taken from a retail personal loan business of a financial institution, and customers who are approved but not actually withdraw money are marked as 1, while those who are approved and actually withdraw money are marked as 0.