Challenges in Using High-Dimensional Clustering Methods to Identify Banks’ Business Models

The present case study deals with a clustering exercise conducted on 382 balance sheet indicators observed for a sample of 365 systemically important euro area banks. The aim of our research was to identify the business models followed by those banks, starting from a highly granular dataset. The technical challenges behind the clustering problem were related to the need to reduce the problem dimension, thus allowing for cluster interpretation, to account for an unknown number of outlying banks, very distant from all the others, to cope with the input “fat” dataset, with more indicators than banks. We solved the problem by adapting the existing factorial k-means method by optimally combining dimensionality reduction and outlier detection with clustering. This procedure resulted in the classification of euro area banks into four clusters: wholesale funded, securities holding, traditional commercial and complex commercial banks. A subsequent statistical analysis was performed to examine whether banks classified into different clusters differ, on average, with respect to their risk-return characteristics. We found that the ensuing business models exhibit clearly identifiable differences in this respect.