<![CDATA[Extreme class imbalance in online payment fraud detection creates an accuracy paradox and an operational risk in which improving fraud capture can generate costly false alarms. This study uses a quantitative, experiment-based design to evaluate the operational impact of common resampling strategies under extreme skew using interpretable linear decision rules. The Online Payments Fraud dataset (6.36 million transactions) from Kaggle is analysed using six monetary balance/amount variables (amount, oldbalanceOrg, newbalanceOrig, oldbalanceDest, newbalanceDest) plus the rule-based isFlaggedFraud indicator to predict the isFraud label. Five training variants (no resampling, ROS, RUS, SMOTE, ADASYN) are compared with two linear decision rules: an ordinary least squares linear scoring model (thresholded at 0.5) and a linear SVM, using a leakage-free protocol in which resampling is applied only to the 80% training split and performance is assessed on an untouched, highly imbalanced 20% test set. The findings indicate that LinReg–RUS achieves the most balanced operating point (Precision 65.938%, Recall 47.718%, F1 55.367%, ROC-AUC 98.720%), whereas ADASYN increases recall but collapses precision (~2.1%), yielding F1 ≈4.17%. These results contribute controlled, large-scale evidence that under extreme imbalance, simpler resampling–model combinations can provide more deployable precision–recall trade-offs than aggressive synthetic sampling, supporting interpretable baselines for capacity-constrained payment screening.]]>
