Evaluating The Performance Of Credit Scoring Models Using Data Mining Techniques : Case Of The Ghanaian Banking Industry

Banks and other financial institutions cannot do away with the outstanding credit balances with customers and the speed at which customers file for bankruptcy presents a risk to the credit industry sustainability. The demand for banks and financial institutions to break even in the credit industry and their quest to reduce the cost to the minimal in the future has attracted various research interests from industrial communities, research laboratories, and academics. To save losses in the future, it requires that a more accurate and robust model with a consistent predictive ability to evaluate credit either using existing models or through a proposal of new models. In this work, we develop a linear model using logistics model. We evaluated four algorithms based upon their generational history in terms of performance. Support vector machine, logistic regression, artificial neural networks and random forest as an alternative if all three models fail to predict accurately in the Ghanaian dataset and we benchmark the experiment with the German and Australian dataset. We developed a linear discriminant model for determining the probability of defaults. We adopted 10-fold cross-validation and data partition as two major data splitting technique to ensure efficiency of all the classifiers. In our findings, it was evident that Support vectors machine, in general, have higher accuracy in terms of classification accuracy compared to logistic regression and artificial neural networks using AUC, Type I and II errors and Risk charts. Random Forest was better than all three classifiers in the case of Ghanaian dataset but fail to give higher predictive accuracy in the German and Australian dataset. RF, SVM, and LR can be used as alternatives to each other from our study and proven by other related works. This study did not consider the default probability or creditworthiness of applicants who were rejected in the process of applying for loans. In other words, our sample dataset contains information of only applicants who were granted the loans and defaulted. This could lead to bias in our analysis even though it is acceptable in this field of study