Artificial Neural Networks and Deep Learning for stress testing a banking system

Abstract

This thesis is focused on the use of artificial neural networks (ANNs) and deep learning techniques for solving financial problems that can be encountered by banks and regulatory authorities. Banking systems are characterized by intricate linkages that increase the overall systemic risk and have the potential to undermine the financial stability of the whole economy. For this reason, the measurement of the resilience of banking systems to adverse shocks and the identification of the main sources of risk are the main focus for banks and regulators. Several tools have been developed for this purpose. However, many real-world financial problems have a non-linear behavior, which is difficult to capture with classical statistical tools. With the aim to address these nonlinearities, the present work discusses the implementation of ANNs in areas related to financial stability and in particular to stress-testing. As a final case study analysis, the application of a deep neural network to dynamic balance sheet stress-testing is performed by using real US data. At the same time, the choice of a suitable architecture and hyperparameters is examined with the purpose of enhancing the model’s generalization and predictive capabilities.