Comparative Analysis of Machine Learning and Traditional Models in Credit Risk Prediction

Thesis Abstract

Effective credit risk assessment is pivotal for the stability and profitability of financial institutions, yet traditional statistical models such as logistic regression often face limitations in capturing complex, nonlinear patterns inherent in borrower data. This study investigates the comparative effectiveness of machine learning algorithms and conventional models in predicting credit risk, aiming to enhance predictive accuracy and operational efficiency within credit scoring systems. The primary objectives include evaluating the predictive performance of selected machine learning techniques—namely random forests, support vector machines, and gradient boosting machines—against traditional logistic regression, identifying significant predictor variables, and determining the models’ suitability for practical implementation in a banking context. Employing a quantitative research design, the study utilizes a cross-sectional approach on a dataset comprising 10,000 anonymized credit applications obtained from a leading commercial bank over a five-year period. The sample encompasses diverse demographic, financial, and behavioral features relevant to creditworthiness, such as income level, employment status, loan amount, repayment history, and credit bureau scores. Data collection was facilitated through existing bank records, complemented by structured data validation procedures to ensure accuracy and consistency. Ethical considerations, including data anonymization and confidentiality, were strictly adhered to throughout the research process. The analytical framework involves preprocessing steps like feature scaling, missing data imputation, and feature selection using recursive feature elimination. Model performance was assessed through robust validation techniques, including k-fold cross-validation and out-of-sample testing. Quantitative analysis employs a combination of descriptive statistics, ROC curve analysis, precision-recall metrics, and statistical significance testing via paired t-tests and McNemar’s tests to compare model performances. Furthermore, variable importance measures provided insights into the contributing factors influencing credit decisions within each modeling framework. The theoretical foundation integrates the Signal Detection Theory to interpret classification performance and the Prospect Theory to understand decision-making biases in credit lending. Anticipated findings suggest that machine learning models, especially gradient boosting machines, will outperform traditional logistic regression in predictive accuracy, as evidenced by higher AUC-ROC scores, improved sensitivity, and specificity. However, the study also expects to identify trade-offs concerning interpretability and computational complexity. The insights derived are anticipated to contribute to the existing literature by providing a comprehensive comparison based on real-world data, thereby informing practitioners about optimal model selection under varied operational constraints. This research advances knowledge by systematically evaluating modern predictive tools against established statistical methods within the domain of credit risk assessment, emphasizing practical implications for banking institutions seeking to adopt data-driven decision models. The findings are expected to recommend integrating machine learning algorithms into credit scoring processes, tailored with appropriate explainability measures to address regulatory requirements. The study concludes with a set of evidence-based recommendations for financial institutions aiming to optimize credit risk management, alongside suggestions for future research exploring hybrid models and the integration of alternative data sources to further improve credit evaluation processes.