Fusing copula-based anomaly detection with ensemble machine learning models for real-time and interpretable financial fraud detection

Abstract

The increasing sophistication and high volume of financial transactions complicate the challenge for financial institutions to manage real-time fraud detection techniques. In the current study, a novel hybrid machine-learning model by the name HAD-IFCX is proposed, which integrates two complementary machine-learning models like Isolation Forest, a copula-based anomaly estimator, and XGBoost classification to increase the detection accuracy of fraudsters in the face of extreme imbalance in classes. The model will have a 5-stage pipeline that includes data preprocessing, class balancing using SMOTE, unsupervised anomaly detection, and supervised classification steps. The comparison is made between the performance of HAD-IFCX and that of the conventional classifiers, which consist of logistic regression, naive Bayes, KNN, decision trees, and neural networks, on the publicly available dataset on Kaggle on credit card fraud. Empirical findings indicate that HAD-IFCX outperforms all the baseline models in every measure of accuracy (98.3%), precision (95.1%), recall (94.0%), and AUC-ROC (0.985). The further screening using the confusion matrix reveals that there is little misclassification, which points to the efficacy of the model. This paper therefore proposes HAD-IFCX as an explainable and scalable model of fraud detection with actionable information that can be applied in real-time financial systems.