Monitoring bank risk around the world using unsupervised learning

Abstract

This paper provides a transparent and dynamic decision support tool that ranks clusters of listed banks worldwide by riskiness. It is designed to be flexible in updating and editing the values and quantities of banks, indicators, and clusters. For constructing this tool, a large set of stand-alone and systemic risk indicators are computed and reduced to fewer representative factors. These factors are set as features for an adjusted version of a nested k-means algorithm that handles missing data. This algorithm gathers banks per clusters of riskiness and ranks them. The results of the individual banks' multidimensional clustering are also aggregable per country and region, enabling the identification of areas of fragility. Empirically, we rank five clusters of 256 listed banks and compute 72 indicators, which are reduced to 12 components based on 10 main factors, over the 2004–2024 period. The findings emphasize the importance of giving special consideration to the ambiguous impact of banks' size on systemic risk measures.