Cluster-based classification approaches to bank failure

In this thesis, we present cluster-based classification methodology as a process of identifying bank failure. A central part of the thesis is an analysis of US commercial banks condition and performance pre- and post 2007 banking crisis. The structure of US commercial banks has changed radically in the forms of funding patterns, securitisation and credit risk transfer mechanisms. The change is reflected in both the results of the analysis we carried out in this thesis and in the cluster-based classification. Cluster-based classification allows us to partition a classification problem through clustering on a subset of features, then assigning and training one or more classifiers individually on each cluster and customizing the feature set for the cluster. The approach we developed for banks rely on two information sources namely, financial data of banks and set of financial features that can be used for clustering. In our implementation of the methodology, financial banks are partitioned into clusters and in order to determine the cluster-specific feature subset, we perform feature selection on each cluster using a fast correlation based feature selection algorithm. Finally, a multilayer perceptron neural network is attached to each of the clusters for classification. Using data set on US commercial banks obtained from the Federal Deposit Insurance Corporation (FDIC) , we found that the cluster-based classification approach achieves improved classification accuracy, in an order of up to 16% higher, compared to non-clustering approaches on both the pre and post-2007 bank crisis. In addition, through a computational analysis of the underlying financial dynamics of banks, we identified two main failure patterns which we labelled gradual and brittle. Gradual pattern encompasses banks whose failure proceeds in small steps and failure is not likely to be abrupt. Unlike banks exhibit- ing a gradual failure pattern, brittle banks may not give ample warning of impending failure.EThOS - Electronic Theses Online ServiceGBUnited Kingdo