Abstract
This paper details the creation of a probability of default (PD) credit scoring model tailored for the Digital Auto Loan product, utilizing loan contracts issued from August 2023 through November 2024. Two distinct modeling strategies are explored: a standard logistic regression model and a hybrid model that merges XGBoost with logistic regression. In the binary classification task, the predicted probability generated by the XGBoost model is added as an extra explanatory variable in the logistic regression model. This approach amalgamates the nonlinear forecasting capability of gradient boosting with the clarity and regulatory compliance of logistic regression, resulting in enhanced predictive accuracy. The modeling process is executed within an automated machine learning framework, which facilitates reproducible training, hyperparameter tuning, version management, and scheduled evaluation of the model. Experimental findings indicate that both the standard and hybrid models exhibit robust discriminatory power, with the combined method attaining a Gini coefficient of 0.72.