THREE DECADES OF CREDIT RISK MANAGEMENT EVOLUTION: A SYSTEMATIC LITERATURE REVIEW ON MACHINE LEARNING INTEGRATION AND IFRS 9 FORWARD-LOOKING FRAMEWORKS

Abstract

Credit risk assessment has been fundamentally reshaped over the last three decades, progressing from traditional score carding methods to machine learning techniques, and ultimately to forward-looking approaches dictated by IFRS 9. The purpose of this Systematic Literature Review (SLR) is to use published research from 1000 sources over the time frame 1993-2025, collected via Publish or Perish and Google Scholar to examine the evolution, method diversity and regulatory considerations in the literature of credit scoring research. Following PRISMA, five main methodology clusters were discovered: traditional machine learning, deep learning, ensemble methods, statistical/logistic models, and IFRS 9/Expected Credit Loss (ECL) models. It has been revealed that ML adoption rates have sky-rocketed since 2018, with ensemble techniques (XGBoost, Random Forest and gradient boosting being the most widely used methods) possessing superior discriminatory abilities compared to the logistic regression approach. Incorporation of macro-economic variables (GDP growth, inflation, unemployment rates) within a point-in-time probability of default (PD) model results in significant accuracy improvement under the IFRS 9 forward-looking framework. Emerging deep learning models like LSTMs and hybrid CNN-RNN networks prove advantageous for modelling sequence data on loan performance but issues of interpretability remain a crucial obstacle to implementation for regulatory authorities. Key research gaps were identified: a lack of explainability models (XAI) fitting with Basel III/IV capital requirements, sparse research on emerging market banking systems, and limited incorporation of climate related financial risk in ECL modeling. The findings presented in this paper hold significant implications for practitioners in banking, risk modeling specialists and regulators on the implementation of forward-looking credit risk architectures.