Explainable credit risk modeling with hybrid tabular deep learning and adaptive feature routing

Well defined and transparent credit risk evaluation continues to be a fundamental issue in financial decision-making, particularly when dealing with intricate borrower profiles in the forms of organized tabular data. While traditional machine learning models often lack interpretability for predictive accuracy, recent deep learning approaches struggle to generalize across such data formats. We propose an innovative hybrid Tabular Deep Learning system that combines feature tokenizer transformer (FT-transformer) and TabNet architectures with an adaptive feature routing (AFR) method. The AFR module dynamically identifies significant aspects for each data instance, facilitating context-aware representation learning and enhancing generalization across other borrower groups. To guarantee explainability and compliance with regulations, our approach integrates a multimodal interpretability package that includes Shapley additive explanations (SHAP)-based attributions, attention mapping, and counterfactual reasoning for actionable what-if analysis. Comprehensive experiments on the benchmark dataset reveal the model’s exceptional performance, attaining an AUC of 0.985, an F1-score of 0.972, and a premier ranking according to the Gini coefficient. Visual metrics of AUC vs. Gini coefficient, cost-benefit curves, and the counterfactual dashboards showcase the model’s transparency and practical applicability. This study observes the application of explainable AI in credit risk modeling by successfully reconciling the balance between higher predictive accuracy and interpretability, hence facilitating explainable financial decision-making scenarios.