<![CDATA[Purpose: This study examines the use of the Rare-Event Logit approach to improve fraud detection under conditions of extreme class imbalance. The topic is important because fraud cases usually represent only a very small proportion of total financial transactions, which may reduce the accuracy of conventional classification models. Design/methodology/approach: This study uses a simulation-based quantitative design to evaluate fraud detection performance in large-scale imbalanced data settings. The analysis compares standard logistic regression and Rare-Event Logit with bias-corrected estimation, including Firth’s penalized likelihood approach. Model performance is assessed using the Area Under the Precision-Recall Curve and F1-Score. Findings/Results: The results show that standard logit and Rare-Event Logit perform similarly under moderate imbalance conditions. However, Rare-Event Logit provides a stronger theoretical advantage in handling rare-event bias and stabilizing parameter estimation as data sparsity increases. This indicates that bias-corrected probabilistic models are more suitable for fraud detection in highly imbalanced environments. Originality/Value: This study highlights the value of Rare-Event Logit as an alternative approach for fraud detection in rare-event settings. The findings imply that financial institutions can improve fraud risk identification by adopting bias-corrected models that are more robust to class imbalance.]]>
Copyrights © 2026
