A comparative study of bank distress prediction models

Accurately forecasting bank distress is crucial for maintaining stability in the financial system. This thesis conducts a comparative study of bank distress prediction models in seven Asian economies. The research period covers the years from 2003 to 2019. The main objective of this research is to investigate factors that provide signals of distress in banks and suggest a prediction methodology for use in Asian emerging and developing economies. To achieve these objectives, four quantitative methodologies - multiple discriminant analysis (MDA), logit (or logistic) model, artificial neural networks (ANNs) and extreme gradient boosting (XGBoost) – were employed to predict bank distress using data from seven selected economies. The research incorporates non-conventional variables in the prediction model to explore the asymmetric effect of profitability on the likelihood of bank distress under different macroeconomic conditions. Additionally, the study examines acquired financially distressed banks in the modeling process, which is a novel contribution compared to previous literature. The results of the study show that while logit is preferable due to interpretability, XGBoost has an advantage in predictive accuracy. Hence, a combination of logit and XGBoost is highly recommended for practitioners in distress prediction. The study emphasis the more sensitive influence of profitability on bank distress in high-GDP growth economy or during non-crisis period. By identifying the factors that contribute to bank distress and suggesting a suitable prediction methodology, this study can help policymakers make informed decisions to mitigate the negative effects of bank failure and prevent a systemic crisis. This thesis also contributes to the literature on bank distress prediction models by providing a comparative analysis across different economic environments.