In this master's thesis, we focus on building models for the assessment of credit risk using actual data of credit transactions obtained from a Slovenian bank. We use two approaches to construct predictive models from data: traditional machine learning methods and their combination with survival analysis methods. The main question of our work is whether survival analysis can improve the predictions of credit defaults compared to classic methods. Within the research framework, we compared the predictive performance of both approaches using models based on decision trees and random forests. The results showed that, on the training set, traditional machine learning algorithms reach higher predictive capability than survival analysis models. Traditional algorithms showed slightly higher predictive performance on the test set, but both approaches were comparable here. There was significant overfitting with traditional machine learning algorithms, while the survival analysis method proved to be quite resistant to it. We, therefore, conclude that despite certain advantages, survival analysis does not guarantee an improvement in predictions compared to traditional machine learning methods.
|
