Credit score model via GAS model and logistic regression

This thesis examines and compares the effectiveness of credit scoring models and intro- duces a new one with time-varying parameters, the so-called GAS-logistic regression model. The model integrates logistic regression, the traditional approach for predicting the probability of default in financial institutions, with dynamic features of Generalized Autoregressive Score (GAS) models. The objective is to find out whether integrating GAS models enhances the prediction of risk over time. By conducting both, simulations and empirical analysis on real-world U.S. mortgage data, the performance of the GAS-logistic model is analyzed and benchmarked against that of the standard logistic regression model. The findings suggest that integrating GAS models provides more nuanced predictions of risk over time. However, given the optimization of nonlinear functions through gradient algorithms, results vary with different initial values. That suggests that further research could be conducted to refine this model.