Comparison of Different Versions of a Decision Tree Method for Creditworthiness Assessment of Private Persons

In recent decades, financial institutions have attached increasing importance to credit risk management as an essential tool for controlling their profitability. The idea of risk reduction is the ability to identify potential loan applications prone to default and reject them. Financial institutions use statistical analysis and machine learning models to assess the creditworthiness of potential borrowers. Nowadays there are many of research on credit scoring models that compare models based on statistical metrics. However, the real needs of financial institutions are to generate profits, not to optimize statistical metrics. Logistic regression is the standard method for credit scoring. However, modern machine learning models may outperform it. Decision tree algorithms, developed in the 1980s, have been given a boost with the development of modern methods. This bachelor thesis provides a comparative analysis of different versions of the decision tree algorithm in the context of assessing the creditworthiness of private persons on two real historical data sets. Specifically, this study compares logistic regression, the CART algorithm, and two ensemble methods: Random Forest and Gradient Boosted Decision Tree. The comparison is made using statistical metrics (accuracy, precision, recall, f1-score), the area under the ROC-curve, and modeling the potential gains from applying a particular model. The results demonstrate that the Gradient Boosted Decision Tree algorithm outperforms other models. Also, the results indicate that improving the statistical performance of the model does not always lead to an increase in profits from using this model.