Improving Credit Scoring : A Rescaled Cluster-Then-Predict Approach

Credit scoring models predict a person's creditworthiness. Accurate and interpretable credit scoring models help maximize a financial institution's risk-adjusted return and hence are of considerable importance. Although the XGBoost is the state-of-art classifier for this task, its complexity makes it challenging to interpret. This paper proposes a novel approach, the rescaled cluster-then-predict method, which rescales features before clustering the data into subgroups. Our approach leads to a more interpretable, yet still competitive, Logistic Regression that even outperforms XGBoost in some cases. Empirical results on two datasets highlight the benefits of the proposed method. We also find that clustering positive cases (rather than all cases) leads to similar results while requiring substantially less computation time. These findings have important implications for the finance industry, where accurate, interpretable, and efficient credit scoring models are in high demand