<![CDATA[Background: Traditional rule-based systems are increasingly inadequate for addressing dynamic financial threats because of their reactive nature and susceptibility to concept drift. At PT XYZ, this vulnerability resulted in a sharp decline in fraud prevention effectiveness, from 90.4% in 2022 to 75.3% in 2025, leaving a critical 24.7% security gap. Objective: This study aimed to develop a behavioral fraud detection framework for digital banking by integrating a 38-feature account-level taxonomy with Isolation Forest, LightGBM, CatBoost, and the Synthetic Minority Over-sampling Technique (SMOTE) to overcome the limitations of legacy rule-based systems. Methods: The study established a proactive theoretical framework using the Cross-Industry Standard Process for Data Mining (CRISP-DM) methodology to shift from transaction-level monitoring to holistic account-level behavioral profiling. Central to this approach was a 38-feature taxonomy that integrated statistical aggregates, temporal signatures, velocity flags, and volume metrics to map the behavioral “fingerprint” of at-risk accounts. Results: Empirical testing on 402 unique accounts confirmed that the proposed architecture successfully closed the identified 24.7% security gap. Both the LightGBM and CatBoost classifiers achieved a perfect recall score of 1.0000 and an accuracy of 0.9972. LightGBM emerged as the best-performing model, with a receiver operating characteristic area under the curve (ROC AUC) of 0.9996, significantly outperforming traditional logistic regression. Conclusion: The proposed framework effectively improved fraud detection by combining behavioral profiling, hybrid anomaly detection, and balanced ensemble learning. LightGBM and CatBoost achieved 99.72% accuracy, 100% recall, and a 99.96% ROC AUC, demonstrating a practical and explainable approach to proactive digital banking fraud detection.]]>
