Goldmann, Sebastian Hendrikus (2024) Enhancing Credit Risk Prediction in Retail Banking: Integrating Time Series and Classical ML Algorithms.
Abstract
This thesis investigates the application of Time Series Classification (TSC) algorithms to enhance credit risk models, focusing on the historical balance data of retail customers. The primary aim was to explore how TSC models could improve the discriminatory power of these risk models. Employing various TSC techniques, including deep learning, shapelets, and Canonical Interval Forest (CIF) algorithms, the study rigorously evaluated their performance in predicting the Probability of Default (PD). Additionally, this thesis contributes to the evolving field of credit risk modeling by introducing a novel approach that integrates TSC with traditional credit risk assessment methods. The research revealed that TSC algorithms, particularly when applied to end-of-day balance data, have the potential to significantly enhance the predictive accuracy of credit risk models. The CIF model, in particular, demonstrated notable efficacy, rivaling the performance of existing credit risk models. However, applying TSC algorithms to multivariate monthly data showed limited effectiveness, suggesting the removal of critical information in such data aggregation. The study also highlighted the interpretability challenges with complex TSC models and the need for more holistic data inclusion for a comprehensive credit risk assessment.
Supervisors
UT Supervisors
- Dr. Jörg Osterrieder
- Dr. Marcos Machado
ING Supervisors
- Voichita Etesse