Credit Risk Prediction Using Machine Learning Models: A Data-Driven Study Within the Broader Mathematical and Statistical Framework of Banking Risk Management

Credit risk prediction is one of the most consequential applications of quantitative modelling in the banking and financial sector, sitting alongside market-risk pricing models, time-series volatility forecasts, and portfolio-optimization frameworks as part of the mathematical and statistical backbone that underpins modern banking systems. This paper first reviews that broader backbone — the Black-Scholes-Merton option-pricing framework, ARIMA and GARCH time-series models used for volatility and interest-rate forecasting, and the Markowitz mean-variance framework for portfolio risk — and critically discusses the assumptions and limitations of each in today's volatile, structurally-shifting economic environment. Building on this foundation, the paper narrows its focus to credit risk and presents an original empirical study that classifies loan applicants as low-risk or high-risk. A real-world dataset containing 32,581 loan records with 12 borrower attributes was preprocessed by imputing missing values, encoding categorical variables, and normalizing skewed numerical features. Four supervised machine learning algorithms — Logistic Regression, Decision Tree, Random Forest, and Gradient Boosting — were trained on an 80/20 split and evaluated using Accuracy, Precision, Recall, F1-Score, and ROC-AUC. Random Forest achieved the strongest overall performance (93.36% accuracy, 97.14% precision, 82.48% F1-score, ROC-AUC = 0.9313), followed closely by Gradient Boosting (92.70% accuracy, ROC-AUC = 0.926), while both ensemble methods clearly outperformed the linear Logistic Regression baseline (85.12% accuracy). The results demonstrate that tree-based ensemble learning provides materially more reliable credit risk classification than either classical linear scoring or single decision trees, and the paper closes by situating these findings against the well-documented fragility of constant-parameter statistical models such as Black-Scholes, ARIMA/GARCH, and Markowitz mean-variance optimization under regime shifts and market stress.