Predicting U.S. bank failures and stress testing with machine learning algorithms

Abstract

This study applies multiple machine learning models to forecast the bankruptcy of U.S. financial institutions from the year 2001 to 2023 using data from the Federal Deposit Insurance Corporation. To incorporate time dynamics, this paper employs exponentially weighted moving averages, enhancing the models’ predictive accuracy. The results show that the Random Forest model achieves the highest overall accuracy, while logistic regression, XGBoost, Support Vector Machine, and neural networks offer various levels of performance. Stress testing and sensitivity analysis reveal that model accuracy is heavily reliant on key financial characteristics, and severe stress conditions can significantly reduce predictive capacity.