From Black Box to Clarity: Machine Learning and Agnostic Techniques in Credit Risk Management

Abstract

Credit risk, a pivotal bank concern, was initially addressed in the Basel framework through the capital adequacy ratio. The need for superior risk models remains pressing as the banking landscape evolves. This research examines the effectiveness of machine learning in credit risk assessment using a dataset comprising over 280 corporate loan applications from a public Tunisian bank for approved loans in 2023's challenging economic conditions. This study compares logistic regression, neural networks, and random forests, implementing preprocessing steps, including missing value treatment and variable selection through Lasso and Ridge regularization. Our findings reveal that random forests achieve 94% accuracy in predicting default probabilities, outperforming neural networks (92%) and logistic regression (88%). Return on assets emerges as the most significant predictor for random forests, while the debt-to-equity ratio dominates neural networks and logistic regression predictions. We implement a novel three-tier interpretability framework combining SHAP, LIME, and Partial Dependence Plots (PDP) to address the “black box” challenge of machine learning algorithms. This comprehensive approach enhances model transparency and reveals critical financial thresholds specific to emerging markets, which is particularly valuable given Tunisia's economic context. The results demonstrate that sophisticated analytics combined with robust interpretability methods can significantly improve credit risk assessment in challenging economic environments. The implications extend beyond traditional banks to microfinance institutions, offering a framework that balances advanced prediction capabilities with transparent decision-making processes. This approach proves particularly valuable in emerging markets, where default patterns differ significantly from those in developed economies.