Rethinking Profitability Dynamics in the EU Banking Sector: An Explainable Machine Learning Approach to Bank Sector-Specific, Macroeconomic, and Institutional Quality Factors

This study comprehensively examines the banking, macroeconomic, and institutional quality dynamics determining the return on equity (ROE) of the banking sector in 27 European Union countries over the 2014–2024 period. Adopting a purely explanatory framework rather than a predictive exercise, the primary aim is to identify and rank the factors driving cross-country profitability variability. The traditional multiple linear regression (MLR) method and three machine learning models (CatBoost, Extra Trees, Gradient Boosting) are comparatively analyzed, with model transparency ensured via the Shapley Additive Explanations (SHAP) algorithm. Empirical findings provide evidence consistent with strong, non-linear interactions among profitability dynamics that traditional econometric models tend to overlook. Comparative analyses indicate that the best-performing CatBoost algorithm possesses notably higher explanatory power compared to the MLR model, an advantage that persists when the linear benchmark is augmented with country fixed effects. According to SHAP results, the non-performing loan (NPL) ratio is the most dominant factor eroding profitability. Conversely, inflation is associated with a positive impact on ROE through the repricing channel up to a certain threshold, after which its marginal contribution flattens, exhibiting a concave structure. These thresholds should be read as model-implied patterns within the present sample rather than as general economic constants. The direct explanatory power of the institutional quality indicators employed here—and of a principal-component composite of the broader governance set—remains relatively limited, suggesting an indirect role operating through macroeconomic channels. These findings, supported by leave-one-country-out (LOCO), fixed-effects, and lagged-regressor robustness checks, suggest that explainable machine learning offers a valuable analytical infrastructure for characterizing the asymmetric effects of macro-financial shocks on bank performance.