Strategic Default Detection Leveraging Card Spending: Static-Feature Ensemble and Cluster-Augmented Signals for Non Time-Series Modeling

초록

This study presents an ensemble approach for detecting strategic defaulters using transaction data from department-store-exclusive credit cards. Strategic defaulters are individuals who intentionally default after a short-term spending surge, despite having sufficient repayment capacity, making them difficult to detect using conventional credit scoring systems. Instead of modeling temporal sequences, we extract static features that capture not only abrupt changes, concentrations, and irregularities in monthly spending patterns but also overall consumption levels. These features allow us to quantify behavioral anomalies without relying on time-series structures. Additionally, we introduce a cluster-augmented stage that fits K-Means clustering and a Gaussian mixture model to derive geometry- and density-based signals and append them to the static features. To address extreme class imbalance, we employ oversampling techniques such as random oversampling, SMOTE, and ADASYN. We then constructed a soft voting ensemble model that integrated logistic regression with tree-based classifiers, including Random Forest, XGBoost, and LightGBM. Experimental results show that the ensemble approach significantly enhances the recall and F1-score for the minority class compared to the individual models. Under the same data splits and oversampling settings, these cluster-augmented features improved minority-class detection and were supported by label-wise univariate analyses revealing clear differences between strategic defaulters and non-strategic cases. The key predictive signals include spending volatility, transaction sparsity, seasonally concentrated patterns and cluster-derived geometry- and density-based indicators. Our findings demonstrate that non time-series modeling using static features can effectively capture strategic delinquency risks, providing a lightweight yet robust solution for imbalanced classification tasks in credit risk detection.

키워드