Thesis (BS Applied Mathematics) -- University of the Philippines Mindanao, 2009

Credit scoring is a classification problem wherein loan applicants are classified in good or bad depending on their characteristics. Mounting losses from the delinquent loans have urged researchers to improve the accuracy of the loan classifiers. One way of improving the accuracy is through the tuning of parameters of the classifiers such that the number of accepted applicants that will violate the credit agreement is minimized. Thus, credit scoring is an optimization problem as well. One of the most effective optimization tools is particle swarm optimization (PSO). hence, PSO was used in the study to find the weights in the commonly used credit scoring classifiers, logistic regression and distance weighted k-nearest neighbors (DWKNN). The effectiveness of PSO in finding these weights was compared to the commonly used tools in weight estimation, maximum likelihood estimation for logistic regression (MLE-LR) and the inverse of the computed distances for DWKNN (ID-DWKNN). Four classifiers were developed: PSO-LR, PSO-DWKNN, MLE-LR and ID-DWKNN. It was found that the accuracy of PSO-DWKNN in detecting the applicants that will default was better than ID-DWKNN when the training set was composed of applicants that were picked randomly. The classifiers MLE-LR and PSO-LR were found to perform better than the two classifiers when the training set had equal number of applicants from both classes. Both were able to classify all the bad applicants correctly. Nevertheless, it is noteworthy that PSO was not able to find the optimal set of weights for logistic regression and DWKNN. PSO, however should not be given up entirely as its performance is also dependent on the training set and the nature of the data.