An Ensemble Machine Learning Approach for Detecting Fraudulent Insurance Claims

Abstract

Fraudulent insurance claims impose substantial financial losses and operational inefficiencies on the global insurance industry. Detecting such claims is particularly challenging due to severe class imbalance, evolving fraud tactics, and heterogeneous claim characteristics. This study proposes a structured ensemble machine learning framework for detecting fraudulent insurance claims using a stacked generalization approach. The proposed architecture integrates logistic regression, Random Forest, and XGBoost as base learners, while a meta-level classifier optimally combines their predictive outputs. The framework emphasizes robust preprocessing, class-weight adjustment, and cross-validation-based stacking to mitigate information leakage and enhance minority-class detection. Experimental evaluation on an imbalanced insurance claim dataset demonstrates that the stacked ensemble significantly outperforms individual base models in terms of Area Under the Curve (AUC), recall, and F1-score. The ensemble approach effectively reduces false negatives, thereby improving operational fraud sensitivity without substantially increasing false positives. The findings confirm that ensemble learning enhances predictive stability and generalization in insurance fraud detection tasks. The proposed model offers a scalable, interpretable, and deployment-ready solution for modern insurance analytics environments.