Robust Clustering of Banks in Argentina

Abstract

The purpose of this paper is to classify and characterize 64 banks, active as of 2010 in Argentina, by means of robust techniques used on information gathered during the period 2001-2010. Based on the strategy criteria established in (Wang 2007) and (Werbin 2010), seven variables were selected. In agreement with bank theory, four “natural” clusters were obtained, named “Personal”, “Commercial”, “Typical” and “Other banks”. In order to understand this grouping, projection pursuit based robust principal component analysis was conducted on the whole set showing that essentially three variables can be attributed the formation of different clusters. In order to reveal each group inner structure, we used R package mclust to fit a finite Gaussian mixture to the data. This revealed approximately a similar component structure, granting a common principal components analysis as in (Boente and Rodrigues, 2002). This allowed us to identify three variables which suffice for grouping and characterizing each cluster. Boente’s influence measures were used to detect extreme cases in the common principal components analysis.

Keywords

• robust clustering
• projection pursuit
• common principal components
• robust K-means
• influence measures
• theory of the firm