This project used three machine learning models to analyze credit card default risk without relying on credit scores or credit history. The models were Logistic Regression, Random Forest, and Deep Learning, and the dataset consisted of 30,000 credit card users and 26 features. The Random Forest model performed the best, with a precision score of 0.80 and a recall score of 0.65. The most important predictors of default were found to be the most recent two months' payment status and credit limit.
The models were intended to aid in human decision-making rather than automate it, and it is suggested that the model output probabilities rather than predictions to increase accuracy and provide human managers with more discretion. These models could be useful for credit card firms, loan lenders, and banks in making educated decisions on creditworthiness.