This study aims to examine whether machine learning (ML) models improve the prediction of bank financial performance relative to traditional econometric techniques, using evidence from Islamic and conventional banks in the Middle East and North Africa (MENA) region over the extended period 2004–2024.
An unbalanced panel of 108 banks (35 Islamic, 73 conventional) across 15 MENA countries yields approximately 2,160 bank-year observations spanning three major global shocks: the 2008 Global Financial Crisis (GFC), the 2014–2016 oil price collapse, and the COVID-19 pandemic. Three ML architectures – Multilayer Perceptron Artificial Neural Networks, Random Forest, and Extreme Gradient Boosting (XGBoost) – are benchmarked against multivariate ordinary least squares (OLS) and stepwise regression. Endogeneity is addressed via dynamic panel system generalized method of moments (GMM) (Blundell and Bond, 1998). SHAP (SHapley Additive exPlanations) value decomposition provides model interpretability, and LASSO regularisation confirms variable selection robustness.
XGBoost consistently outperforms all competing models across every banking-type and performance-metric combination. R² improvements over OLS range from 63% to 408%, with the largest gains for conventional bank Return on Equity. SHAP analysis identifies credit risk and the Z-score as universal dominant predictors, while revealing threshold nonlinearities invisible to linear models. COVID-19 imposed asymmetric profitability shocks: conventional banks sustained larger equity-return losses while Islamic banks demonstrated relative resilience attributable to lower leverage and profit-and-loss sharing (PLS) structures. Post-pandemic recovery was faster for conventional banks as the interest-rate normalisation cycle restored net interest margins.
The 2020–2024 data extension relies on Refinitiv Eikon coverage, which may be uneven across smaller MENA markets. SHAP values provide interpretable attribution but do not establish causality. Future research should incorporate Sharia governance quality indices, deep-learning architectures and digital-transformation variables.
XGBoost-based early warning models, calibrated with SHAP-identified credit risk thresholds, offer substantially more accurate performance prediction than regression-based supervisory benchmarks. MENA central banks managing dual-banking systems should model interest-rate policy transmission separately for Islamic and conventional bank segments.
The asymmetric resilience of Islamic banks during the COVID-19 pandemic – attributable to PLS risk-sharing structures – underscores the social value of Sharia-compliant financial intermediation in crisis periods. Findings support targeted regulatory frameworks that recognise the distinct risk profiles of Islamic banking in MENA.
To the best of the authors’ knowledge, this is among the first studies to apply SHAP-interpreted XGBoost to comparative Islamic–conventional bank performance prediction across a two-decade MENA panel encompassing all three major global shocks of the period. The integration of dynamic GMM endogeneity correction with ML-based prediction constitutes a methodological advance over prior literature.
