Predicting bank defaults in Italy: A comparative analysis of conventional and machine learning approaches

Although the probability of default (PD) modeling has reached a great maturity in both academia and business, for the Italian case we demonstrate that banks’ available PD models would be misleading if today applied directly to Italian banks. We argue that what determines the PD of Italian banks, rather than the liquidity, are the return on assets (ROA), the financial leverage and the type of the bank. Furthermore, we demonstrate that the conventional approach dominates the more trendy machine learning (ML) and that model’s performance could be used as a supervisory tool for retrospective analysis of the bank’s position. This work stands out as the only study to consider state aid in defining bank default over the horizon period from 2010 to 2020, examining bank default in Italy across various types of banks, and employing both conventional and machine learning approaches, while also proposing an easy-to-handle three-variable model for predicting bank defaults.