1998 Traditional vs Neural Credit Score Modelling
Author: Aleta Y. Lepatan-Florendo
Credit Score Modelling is a sophisticated decision tool used mainly in risk management and consumer lending. Using historical loan payment data and other relevant credit characteristics, a statistical model can be constructed which can efficiently distinguish between good and bad borrowers. This approach practically reduced human intervention in the credit approval process and thus made the system automated and more objective. Tracing its roots in the financial sectors of advanced economies, credit scoring had been gaining attention in the Philippines, especially in the wake of the Asian Financial Crisis. Traditional Credit Score Models were largely based on logistic regression technique to predict the probability of a good or delinquent loan. Traditional approaches also involved other statistical methods such as discriminant and factor analyses. However, modern advancements in credit scoring posed the use of neural networks wherein statistical models can be constructed even for variables with non-linear relationships; −a known shortcoming of traditional credit scoring models. This paper sought to compare and evaluate the efficiency of traditional (logistic regression) credit score approach against more advanced neural models. Separate models for logistic regression and neural approaches were developed using consumer automobile loans data from 1995 to 1996, together with credit card data for year 1997. Predictive power for both models were then evaluated. The study concluded that Neural Networks resulted in better predictive power compared to the logistic model.