Empirical Investigation of Deep Learning Architectures for Systematic Credit Risk Classification in Heterogeneous Financial Markets

The dynamic nature of global financial markets necessitates robust methodologies for credit risk classification, particularly as credit portfolios diversify across sectors and geographies. This study presents an empirical investigation into the efficacy of various deep learning (DL) architectures-including Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Transformer-based models-for classifying credit risk across heterogeneous financial environments. Leveraging a large-scale, multi-country credit dataset, we benchmark the performance of DL models against traditional machine learning algorithms. The study introduces an integrated feature engineering pipeline tailored for financial time-series data and accounts for market heterogeneity through sectoral and geographic stratification. Our findings demonstrate that Transformer-based architectures consistently outperform other models in predictive accuracy and generalizability across market segments. We further explore model explainability and interpretability using SHAP values. The proposed framework can inform regulators, financial institutions, and investors in adopting data-driven risk management practices