Credit risk assessment using bayesian networks and machine learning approaches

Credit risk assessment is vital for the financial stability and operational efficiency of lending institutions. However, traditional models often fail to capture the complexity and systemic risks embedded in modern financial data. This study aims to enhance both the predictive accuracy and interpretability of credit risk models by proposing a dual-framework approach that compares machine learning (ML) techniques with Bayesian network (BN) modeling. Using LendingClub data across multiple economic cycles, the study applies rigorous data preprocessing, class balancing, and feature engineering. Two modeling phases are developed: a baseline model using borrower and loan features, and an enriched model incorporating macroeconomic indicators such as inflation, unemployment, and GDP per capita. The ML component applies ensemble methods, deep neural networks, and gradient boosting to identify default patterns. Simultaneously, a BN is constructed through feature discretization, Tabu search-based structure learning, and Expectation-Maximization for parameter estimation. Main findings demonstrate that macroeconomic integration significantly improves performance, with the enriched BN’s AUC increasing from 0.725 to 0.757. While ensemble ML models achieve marginally higher discrimination (e.g., XGBoost AUC 0.770), the macro-enriched BN delivers highly competitive accuracy (86.6%) and superior probability calibration, advantages critical for regulatory compliance. Sensitivity analysis, mutual information, and forward–backward inference highlight debt-to-income ratio, sub-grade, interest rate, loan term, and unemployment rate as the most influential drivers of default. Overall, the findings demonstrate that macro-augmented BNs can complement high-performing ML ensembles by providing interpretable and regulatorily aligned credit risk models that support both portfolio-level decision-making and granular, case-level explanations.