Published Paper Details:

The process of credit risk evaluation has undergone a paradigm shift owing to the growing adoption of machine learning and deep learning technology in the past few years. In this survey, we conduct a systematic review of 29 peer-reviewed articles published between 2021 and 2025 and develop a comprehensive taxonomy of nine methodological categories that cover conventional machine learning, deep learning models, generative adversarial networks, interpretable artificial intelligence models, reinforcement learning, graph-based learning, multimodal fusion, large language models, and method aggregation. By performing quantitative meta-analysis, we distil three critical findings: Firstly, that ensemble methods have reached optimal performance with a mean AUC of 0.87 (? = 0.07), which represents 11.5% relative improvement compared to logistic regression baselines while being compliant with regulatory requirements. Secondly, that Explainable AI integration through SHAP and LIME frameworks enables transparent audit trails with at most 1% degradation in model accuracy: Random Forest achieves 99% accuracy on benchmark datasets. Lastly, zero-shot large language models underperform traditional methods by up to 26% (0.67 vs. 0.89 mean AUC) due to limitations in structured data reasoning, but allow semantic risk extraction from unstructured sources. Graph-based architectures with CTGAN augmentation yield the top discrimination of 0.99 AUC but incur a training cost of 1000x compared to ensemble methods. We also show that deep learning offers significant advantages only when datasets contain more than 50,000 samples, and that temporal validation is still critically under-explored, with only 15% of studies including out-of-time testing. The current work presents quantitative benchmarks and identifies deploymentcritical research gaps in causal inference, federated learning, and adversarial robustness.