Abstract
Researching the determinants of bank failure is an important task, yet the extant literature on bank failure early warning models fail to identify which model technique, sampling methodology, or set of coefficients provides the most accurate model when predicting failure on out-of-sample data. In this two-essay study, I examine previously published studies on bank failure prediction to determine with statistical significance which among the chosen set is most accurate. I also examine the effects of bias-adjusting models from the Machine Learning literature to determine if bias-correcting sampling algorithms improve accuracy. In the first essay, I replicate three bank failure models (Martin (1977), Cole and White (2012), and DeYoung and Torna (2013) and use them to demonstrate the importance of out-of-sample predictive accuracy using bias-adjusting metrics and the use of McNemar’s Test to show, with statistical significance, that one set of predictive variables is better than the rest. Future researchers may use this framework to demonstrate significant contributions to the field, and regulators may apply these strategies to choose between candidate early warning models. I also test whether including savings banks (in addition to commercial banks) affects out-of-sample predictive accuracy. In the second essay, I extend the first essay by using oversampling techniques to correct for the inherent bias present in bank failure prediction datasets. I also examine the predictive accuracy of gradient boosting decision tree models and show with statistical significance that machine learning techniques can improve out-of-sample predictive accuracy for early warning bank failure models. I find that bias-adjusting techniques do improve accuracy, primarily among healthy banks.