Abstract
There have been significant challenges in accurately predicting credit risk, which is a vital task in the banking and financial sector due to the rising complexity and dynamic nature of financial sector data. For this purpose, conventional machine learning (ML) classifiers have been employed; however, they usually struggle with imbalances in datasets and limited exploration of how these classifiers generalise across diverse credit environments. To solve these problems, this study builds, constructs, and evaluates an ensemble and deep learning models alongside hybrid data resampling and model interpretability approaches, making use of publicly available datasets.
•
Firstly, an approach was proposed to improve credit risk using an ensemble classifier, Synthetic minority over-sampling, Edited nearest neighbor (SMOTE-ENN) was used to tackle class imbalances, and further integrating SHapley Additive exPlanations(SHAP) for models’ interpretability.
•
Secondly, an approach was developed to investigate how effective deep learning (DL) models are in predicting credit risk by focusing on the challenge of class imbalance and the black-box nature of these models. The approach explores deep learning architectures, SMOTE-ENN, and SHAP.
•
Thirdly, a method was developed using a stacked ensemble method coupled with a hybrid data resampling technique to effectively address class imbalances and enhance prediction across credit risk datasets.
These proposed methods achieved better performance as compared to other ML algorithms and some techniques that are available in the literature. The stacked ensemble achieves an accuracy of 0.930 on the Australian and 0.914 on the German datasets, surpassing all other models in the comparison. The research done further proves that ML algorithms tend to obtain improved performance when trained on relevant and balanced data. Furthermore, the research shows the effectiveness of enhanced ensemble methods in credit risk prediction, as well as the significance of the SHAP values, as they help in the interpretability of the models. Additionally, the thesis provides a direction for future research.