Research: Comparative Performance of Machine Learning Models and Logistic Scorecards in...

Research question: How do advanced machine learning models compare to traditional logistic regression scorecards in predicting credit risk on real-world consumer loan datasets?

Background & Motivation: Credit risk modelling is central to modern banking and financial services, underpinning loan approval, pricing, and regulatory capital decisions. Traditionally, logistic regression-based scorecards have dominated industry practice due to their interpretability and regulatory acceptance.

Research Gap: Recent advances in machine learning (ML) promise improved predictive accuracy but raise questions about transparency, overfitting, and practical deployment. Existing studies often use synthetic data or limited real-world datasets and rarely benchmark state-of-the-art ML methods directly against logistic scorecards under realistic constraints.

Approach & Expected Contribution: This project will perform a systematic empirical comparison of multiple ML algorithms (e.g., random forests, gradient boosting, neural networks) and logistic scorecards on large, open-access consumer credit datasets. Techniques for model interpretability and fairness will also be assessed. Performance will be evaluated across metrics relevant for financial institutions (e.g., AUC, precision-recall, stability over time).

Significance: The findings will inform practitioners and regulators about the practical benefits and limitations of adopting machine learning for credit risk, balancing predictive power with transparency and compliance requirements.