Kok, Dyon (2024) Stakeholder-centric approach to applying machine learning to probability of default models.
Abstract
This thesis explores the integration of Explainable Machine Learning (XML) into Probability of Default (PD) models, focusing on the Explainable Boosting Machine (EBM) and Generalized Additive Models with Interactions Network (GAMINet), compared against traditional logistic regression within a financial institution. It assesses explainability, regulatory compliance, performance, and operational feasibility, employing thorough model evaluation methods including architecture review, hyperparameter optimization, and cross-validation. Despite not surpassing the existing logistic regression model in performance, the research underscores XML's capacity to enhance PD models' interpretability without sacrificing accuracy. It highlights EBM's simplicity and GAMINet's ability to capture complex interactions, with a particular emphasis on the significance of transparency and reasoned feature selection for stakeholder trust. The study also considers computational and integration challenges, ethical concerns, and bias management, advocating for a balanced XML integration strategy that combines EBM's efficiency with GAMINet's analytical depth. Future research directions include Python-SAS integration and advanced feature transformation to further refine model accuracy and interpretability. This work contributes to the broader conversation on leveraging machine learning in finance, striving to align technological progress with regulatory and ethical standards.
Supervisors
UT Supervisors
- Dr. Marcos Machado
- Dr. Jörg Osterrieder
ING Supervisors
- Leon Dusée
- Dr. Markus Haverkamp