Metody modelování pravděpodobnosti defaultu firemních zákazníků banky

The aim of the thesis is the prediction of the probability of default using four models and comparison of predictive accuracy of models. The models used for the prediction of binary dependent variable are logistic regression, generalized additive model, support vector machine and decision tree. The models are applied on real data of corporate clients of a Czech bank. ROC, AUC and confusion matrix are used to compare models. The support vector machine seems to be the best model based on AUC, although the model shows the lowest value of overall accuracy compared to other models. The generalized additive model has the second largest value of AUC and the value of overall accuracy, which are only a few points higher than the corresponding values of logistic regression. Decision tree has the lowest AUC, but also the model has the highest value of overall accuracy, because the model correctly classified the largest number of clients into classes.