Banks play an essential role in financing the economy as intermediaries in financial markets. Due to the contagion effects, the bank failure triggers considerably more severe volatility than the failure of other business firms. However, several questions remain unsolved in the existing studies, including which is the most efficient model for predicting bank failure, which indicators are more essential in explaining the bank default risk, and how the bank risk level relates to the ultimate bank insolvency. To answer these questions, this thesis elaborates application of ML models to address the problem of long-term bank failure prediction. To begin with, I provide a comprehensive and structured bibliographic survey of OR- and AI-based research devoted to the banking industry over the last decade. The study reviews the main topics of this research, including bank efficiency, risk assessment, bank performance, mergers and acquisitions, banking regulation, customer-related studies, and fintech in the banking industry. The survey results provide comprehensive insights into the contributions of OR and AI methods to banking. Contributing to the literature, I propose several research directions for future studies that include emerging topics and methods based on the survey results. The bibliography study shows that bank failure prediction is a popular topic that requires highly accurate results. Therefore, I contribute to the literature by determining whether models based on crisis data are suitable for predicting bank failure during stable periods and examining predictors that can be used in long-term forecasting. In the empirical study, I propose a novel multistage AI model to predict U.S. bank insolvencies. Empirical results and robustness checks show the model’s superior ability. Moreover, through the model, I find a consistent impact of liquidity, capital, and asset factors on bank performance over various periods, including crisis and stable times. Additionally, I document a clear difference in failed and non-failed banks’ shock resilience, underscoring the need for regulators to focus on banks’ risk-taking and risk assessment practices. This study demonstrates that AI models can provide insights for bank risk evaluation in multiple ways beyond bank failure prediction accuracy.
- Bank Failure
- Machine Learning
- Financial Crisis
Time is the Witness: Bank Failure Prediction with Machine Learning Models
Zhang, W. (Author). 21 Feb 2024
Student thesis: Doctoral Thesis › PhD