Can K-Means Work For Detecting Bank Risk?
“Banks’ Risk Clustering Using K-Means: A Method Based on Size and Individual & Systemic Risks” by Mathieu Mercadier, Amine Tarazi, Paul Armand, and Jean-Pierre Lardy offers an impressive and novel approach to assessing the riskiness of listed banks using advanced machine learning techniques.
Methodology?
The paper introduces an unsupervised learning algorithm, specifically an adjusted version of a nested k-means algorithm, to cluster listed banks based on their risk factors. This method is unique in its ability to handle missing data, a common challenge in financial datasets. The researchers utilized a large set of both stand-alone and systemic risk indicators, reduced these through dimensionality reduction to fewer representative factors, and then used these factors as features for the algorithm.
One of the critical aspects of this study is its dynamic and transparent decision support tool, which is updated weekly and available on a dedicated website. This tool allows for the clustering results of individual banks to be aggregated per country and region, providing a global perspective on banking risk.
Empirical Analysis and Findings?
The empirical analysis is based on data from 161 listed banks and 85 indicators, reduced to 11 main factors, over the period from 2004 to 2021. An interesting finding of the study is the ambiguous impact of bank size on systemic risk measures. Contrary to some existing literature, this paper suggests that size does not consistently determine a bank’s systemic risk.
The methodology detailed in the paper includes dynamic data extraction, cleaning procedures, calculation of indicators and factors, and the choice and setting of algorithms. The authors emphasize the importance of their method in distinguishing the impact of the Global Financial Crisis (GFC) compared to the more recent COVID-19 pandemic on the banking industry.
Potential Improvements?
The authors propose several improvements for future iterations of this risk clustering approach. These include dealing with multiple principal components for low explained variance and exploring other clustering algorithms like fuzzy k-means, which provides probabilities for each observation to belong to various clusters.
In conclusion, “Banks’ Risk Clustering Using K-Means” is an impressive work that utilizes machine learning to enhance the understanding of banking risks. Moreover, dynamic, transparent, and easily editable methodology sets a new standard in the field of financial risk analysis, providing valuable insights for financial institutions, regulators, and policymakers.
The complete research paper is available Banks’ Risk Clustering Using K-Means: A Method Based on Size and Individual & Systemic Risks by Mathieu Mercadier, Amine Tarazi, Paul Armand, Jean-Pierre Lardy :: SSRN
