In this paper, we applied machine learning techniques to analyze the default probability in financial institutions using a large dataset of variables collected from 2,325 banks over 17 years, extracting the most relevant variables using a feature selection method (Lasso), predicting default and systemic risk with random forest and XGBoost algorithms, and finally investigating the contributions of each relevant feature to the overall financial stress of banking institutions using Explainable Artificial Intelligence (XAI) techniques. According to this methodology, we found that the most important variables for the default risk predictions are the probability of a bailout calculated, the market share in terms of assets, the market-to-book ratio, total liabilities, and the number of banks in the market (a measure of concentration and competition). For the systemic risk predictions, the most important variables are the number of banks in the country, the level of interest rates, the market share of the top $5$ largest banks, and the region of the bank (in North America, Europe, and Central Asia). The findings of this research provide an empirical assessment of the main factors that explain the presence of financial stress in banking institutions, conciliating the versatility of machine learning models with practical interpretability and causal inference, being of potential interest to researchers in quantitative finance and market practitioners.
Comissão Organizadora
Anderson Odias da Silva
Claudia Yoshinaga
Ricardo D. Brito
Felipe Saraiva Iachan
Vinicius Augusto Brunassi Silva
Complete tool for submission, evaluation and publication of proceedings for scientific events.
