Abstract:
Over the last couple of years, we have seen much advancement in mathematical analysis and computational capabilities. This advancement, coupled with the increased availability of big data, has made it possible to commoditise machines and enable them to act as risk managers and financial analysts. In this dissertation, we will briefly review machine learning and consumer credit risk/scoring. We look at different methods and models proposed in the literature and thoroughly explore the mathematical theory behind deep learning. We then apply this knowledge and other recent advancements in the field to build a fully connected feed-forward deep neural network using open source credit card default data from a large Taiwanese retail bank. Our deep neural network aims to improve upon other models proposed in the literature regarding accuracy and other metrics such as ROC-AUC, Cohen's Kappa, precision, recall and F1-score. We then conclude that deep neural networks are competitive in terms of performance compared to other machine learning models and outperform traditional models. We highlight the potential that deep learning has yet to achieve in finance and pay close attention to the hurdles faced in complexity, development costs and regulatory roadblocks.
