Abstract
<title>Abstract</title> <p>Credit card fraud detection models are typically evaluated using predictive performance alone, despite the computational and environmental costs associated with large-scale deployment. This study investigates whether selective computation can improve the trade-off between predictive performance and computational and environmental efficiency in highly imbalanced fraud detection. A two-stage cascade architecture comprising Logistic Regression, an AdaBoost router, and a Stage-2 gradient-boosted classifier (XGBoost, LightGBM, or CatBoost) is evaluated against four standalone models on the ULB credit card fraud dataset (283,726 transactions; 0.167% fraud rate). Experiments are repeated across five independent random seeds using a 70/15/15 train-validation-test split. Predictive performance is assessed using AUPRC, F1-score, and a cost-sensitive metric. Carbon emissions are measured, and floating-point operations (FLOPs) are estimated analytically from each model's structure. Two efficiency metrics, the Carbon Efficiency Ratio (CER) and FLOPs Efficiency Ratio (FER), are introduced to quantify environmental and computational efficiency, respectively. Pareto analysis is employed to identify favorable trade-offs among predictive performance, computational efficiency, environmental impact, and financial cost. Raw XGBoost achieves the highest AUPRC (0.833), whereas the proposed Cascade XGBoost reduces total misclassification cost by 5.2% while maintaining a comparable AUPRC (0.831) and reducing inference FLOPs by 72.7%. In contrast, Cascade Blended substantially increases computational complexity without improving overall efficiency. The results suggest that a carefully designed single-booster cascade can effectively balance predictive performance with computational and environmental efficiency, while additional architectural complexity should be justified by measurable gains across the evaluated objectives.</p>