Forecasting bank failures

Forecasting bank failures is of great interest to investors, analysts and policy makers. Banks are a source of systemic risk and the early identification of banks in distress may mitigate the negative effects of bank failures on the economic system. The objectives of the study are twofold. First, we attempt to identify the most important variables for the identification of the insolvent and the solvent banks and second we construct a forecasting model for bank failures over the 2007-2012 period.

The methodology employed is a Support Vector Machine (SVM) based structural model in order to forecast the failure or survival of a banking institution based on its 4 most recent years of financial data. Our sample consists of 300 U.S. banks selected to represent 200 solvent and 100 insolvent ones. The dataset spans from 2003 to 2012. We collect for each one of the 300 banks in our sample 37 individual variables and financial ratios that come from their publicly reported financial statements. To select the most informative variables we employ a variable selection method based on local learning. This selection procedure results in a small set of only four explanatory variables that are most important for the identification of soon to be failed banks. Then we train the SVM model to classify banks as solvent or insolvent and use a test sample to evaluate the forecasting accuracy of the model in out of sample data.

For this purpose, we used four different kernels: the linear, the radial basis function (RBF), the polynomial and the sigmoid. The best results are achieved using the polynomial and the RBF kernel. The out-of sample overall forecasting accuracy of the model is 94%. The solvent forecasting accuracy is 100% and the insolvency forecasting accuracy is 82.35%.