Domain-Driven Feature Engineering and Multi-Machine Learning Models for Credit Card Default Prediction
TL;DR
This work proves the value of financial domain knowledge in feature construction and provides a stable, reproducible benchmark scheme for credit risk assessment.
Abstract
The paper conducts a systematic study on credit card default prediction using the UCI credit card default dataset with 29,965 samples and 23 original features. The paper proposes six categories of domain-knowledge-driven engineered features, including credit utilization ratio, repayment ratio, delinquency severity score, financial trend metrics, interaction features, and binning features, which convert raw financial variables into effective predictive signals. Under unified experimental settings with an 80 / 20 stratified train-test split and random state set to 42, the paper benchmarks six classifiers: logistic regression, K-nearest neighbors (KNN), random forest, gradient boosting, XGBoost, and multi-layer perceptron (MLP). The gradient boosting model achieves the best performance, with an AUC-ROC of 0.7808 and an F1-score of 0.4693. Exploratory data analysis reveals that defaulted cardholders show a monotonically increasing trend of payment delays, and the distribution of credit utilization ratios differs significantly between defaulters and non-defaulters. Ablation experiments verify that the engineered features improve the AUC value by approximately four percentage points. This work proves the value of financial domain knowledge in feature construction and provides a stable, reproducible benchmark scheme for credit risk assessment.