Cluster-based Adaptive Generation for imbalanced financial data

초록

Financial institutions routinely predict rare but economically consequential events such as credit defaults and client responses. In such environments, severe class imbalance and asymmetric misclassification costs distort predictive performance and may lead to suboptimal capital allocation. We propose a Cluster-based Adaptive Generation (CAGE) framework that addresses unobserved heterogeneity within minority events. The framework partitions minority observations into homogeneous sub-populations and dynamically calibrates generative capacity according to cluster-level structural complexity. We evaluate CAGE across three financial prediction settings with distinct economic objectives: risk-averse lending, cost-sensitive marketing, and income-based client targeting. In addition to Precision, Recall, and F1-score, we directly evaluate economic performance using Total Expected Cost (TEC) and Average Expected Cost (AEC), computed from held-out test-set false positives and false negatives under a pre-specified asymmetric cost ratio. The empirical results show that CAGE achieves the strongest objective-specific and cost-sensitive performance in the risk-averse credit default setting, where minority-event capture and Type-II-error reduction are the primary concerns. In marketing and income-targeting settings, CAGE provides competitive Precision-Recall and F1-score trade-offs relative to the evaluated global oversampling baselines, with performance gains that vary across datasets, metrics, and cost criteria. These findings suggest that heterogeneity-aware synthetic generation is most valuable when minority events exhibit localized structural complexity and when the economic objective places substantial weight on minority-event capture. The proposed framework offers a cost-sensitive empirical strategy for financial forecasting under structural imbalance.

키워드