Predicting the insolvency of u.s. banks using support vector machines (SVM) based on LL-fs

We construct a Support Vector Machine (SVM) based structural model that predicts the collapse of US Bank Institutions using publicly disclosed elements of their financial statements on a rolling 4-year window. In our approximation, we used SVM as a classifier to discriminate solvent to insolvent banks. For the selection of the input variables we applied a novel local learning technique that decomposes the problem of feature selection into a problem of local learning based on linear models and then evaluates the feature relevance globally. The resulting model exhibits remarkable prediction capabilities and underlines the significance of certain variables in overall accuracy, achieving a 14/17 prediction of actually defaulted banks in out-of-sample prediction.