Title: Forecasting U.S. bank failures with machine learning techniques
Authors: Alexander Kostrov - University of St. Gallen (Switzerland) [presenting]
Lyudmila Grigoryeva - University of Konstanz (Germany)
Stefanie Bertele - University of Konstanz (Germany)
Abstract: After consolidation and waves of failures, there are still about 5000 operating banks in the United States. Many of them remain weak and enter the official problem bank list. The main question is whether a Support Vector Machine (SVM) beats well-established models used to predict bank failures (namely, Naive Bayes classifier, discriminant analysis, and logit model). Using FDIC call reports with bank-specific statistics for 2004-2016 we construct a set of CAMELS proxies to predict failures. We consider the class imbalance problem in data, that is, when one class of observations is severely undersampled. We apply the synthetic minority oversampling technique (SMOTE) to solve the problem. The use of SMOTE brings a statistically significant improvement in the forecasting performance of all four models. It means that the class-imbalance problem must be addressed and mitigated in predicting bank failures. SVM is found to be very competitive in comparison to three classical classifiers. It is more accurate for different forecasting horizons. In line with the literature, improvements in classifying historical bank failures are likely to remain in the future.