Tools
Tools
Garay, Patricio Andrés (2023) Comparative Analysis of Interpretability and Accuracy between Gradient Boosting Machine and Explainable Boosting Machine on Default Credit domain. Masters thesis, Dublin, National College of Ireland.
Preview |
PDF (Master of Science)
Download (678kB) | Preview |
Preview |
PDF (Configuration manual)
Download (325kB) | Preview |
Abstract
Machine learning (ML) algorithms have gained ground in credit default modelling due to their great predictive power. Gradient Boosted Machines (GBM) have excellent prediction performance. However, their lack of interpretability challenges their use in the finance sector. This study investigates to what extent GBM can be made explainable in the context of credit default modelling. To assess GBMs, this study contrasts GBM with Explainable Boosted Machines (EBM), a glass box model. The evaluation examines accuracy, precision, recall, F1 score, AUC-ROC, and confusion matrix. The accuracy of both models was around 82%; the other performance metrics were similar.
LIME compared local interpretability, showing that GBM and EBM had a Default prediction of around 0.72. Also, LIME illustrated that PAY_0 (September payment records) contributes most to these predictions, followed by PAY_6 (April payment records) and PAY_2 (July payment records). However, LIME does not explain how GBM decides to reach the prediction. GBM can be used in the credit default domain with an acceptable accuracy level. Using the LIME technique helps to know what variables are relevant in the prediction, but understanding how GBM makes decisions through post-hoc method of local interpretability is not helpful. Future studies should investigate global interpretability techniques or other local interpretability methods. Further research into the relationship between outside variables like the Mid-Autumn Festival and the likelihood of credit default could yield insightful results. The study of the interpretability of GBMs in the credit default domain has relevant implications for financial institutions and regulators.
| Item Type: | Thesis (Masters) |
|---|---|
| Supervisors: | Name Email Cosgrave, Noel UNSPECIFIED |
| Subjects: | H Social Sciences > HG Finance > Credit. Debt. Loans. H Social Sciences > HG Finance > Fintech T Technology > T Technology (General) > Information Technology > Fintech Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning |
| Divisions: | School of Computing > Master of Science in FinTech |
| Depositing User: | Tamara Malone |
| Date Deposited: | 09 Aug 2024 10:28 |
| Last Modified: | 09 Aug 2024 10:28 |
| URI: | https://norma.ncirl.ie/id/eprint/7020 |
Actions (login required)
 |
View Item |