Forecasting Bank Failure in the U.S.: A Cost-Sensitive Approach

Abstract

Preventing bank failure has been a top priority among regulatory institutions and policymakers driven by a robust theoretical and empirical foundation highlighting the adverse correlation between bank failures and real output. Therefore, the importance of creating early signals is an essential task to undertake to prevent bank failures. We used J48, Logistic Regression, Multilayer Perceptron, Random Forest, Extreme Gradient Boosting (XGBoost), and Cost-Sensitive Forest (CSForest) to predict bank failures in the U.S. for 1482 (59 failed) national banks between 2008 to 2010 during the global financial crisis and its aftermath. This research paper stands as a prominent contribution within the existing literature, employing contemporary machine learning algorithms, namely XGBoost and CSForest. Distinguished by its emphasis on mitigating Type-II errors, CSForest, a novel algorithm introduced in this study, exhibits superior performance in minimizing such errors, while XGBoost performed as one of the weakest among the peers. The empirical findings reveal that Logistic Regression maintains its relevance and efficacy, thus underscoring its continued importance as a benchmark model.