Credit scoring with advanced deep learning techniques

Credit scoring is a crucial task in the financial industry. Accurate credit scoring models enable lenders to assess the creditworthiness of applicants more effectively, allowing them to make informed lending decisions. Although traditional machine learning algorithms have recently excelled in credit scoring, the application of deep learning models remains limited due to challenges in credit data and interpretability concerns. This PhD thesis aims to address these challenges to advance credit scoring by improving both predictive performance and interpretability. The first analysis chapter proposed a novel tabular-to-image approach, Tabular Image, incorporating weight of evidence and informative value with a correlation-based feature allocation strategy. The results showed that Tabular Image with two-dimensional convolutional neural networks outperformed traditional machine learning algorithms, highlighting the potential of the proposed method for credit scoring tasks. The second analysis chapter addresses the issues of data scarcity and population shift during credit market expansion. To tackle these challenges, a novel transfer learning framework is proposed, integrating the Tabular Image transformation technique with the weight of evidence technique and pretrained two-dimensional convolutional neural networks. The results demonstrated that the proposed transfer learning framework significantly improved predictive performance in new markets, indicating the potential of this approach in enhancing risk management in emerging markets. The third analysis chapter proposes a novel interpretability framework that combines the Tabular Image with SHapley Additive exPlanations to provide interpretable explanations for two-dimensional convolutional neural networks in credit scoring tasks. The results showed that the proposed interpretability framework provided meaningful and actionable insights into the predictions of two-dimensional convolutional neural networks, ensuring interpretability and regulatory compliance.