Designing an Innovative Risk Management Framework for Fraud Detection in Automobile Insurance: A Case Study of Saman Insurance

Abstract

This study aimed to design and validate an innovative fraud detection risk management framework for automobile insurance at Saman Insurance by integrating fuzzy Delphi-based expert consensus with an XGBoost machine-learning model using real claims data. This applied mixed-method study was conducted in two sequential phases. In the first phase, dimensions, components, and indicators of automobile insurance fraud detection were extracted through literature review, claims-file analysis, and expert consultation and were subsequently screened and validated through a three-round fuzzy Delphi process involving 50 insurance, fraud-detection, machine-learning, and data-mining specialists. In the second phase, 4,860 automobile insurance claims files were analyzed, including 772 fraudulent and 4,088 non-fraudulent claims. Of 47 expert-validated indicators, 24 measurable predictors were retained after preprocessing and entered into an XGBoost binary classification model. Model development incorporated an 80:20 training-test split, five-fold cross-validation, hyperparameter optimization, and early stopping. Performance was evaluated using accuracy, precision, recall, specificity, F1 score, AUC-ROC, precision-recall AUC, and balanced accuracy. The fuzzy Delphi process resulted in seven validated dimensions and 47 indicators, with technical and expert damage indicators receiving the highest fuzzy mean (0.892), followed by policyholder behavioral indicators (0.874) and accident characteristics (0.853). The final XGBoost model achieved a test accuracy of 0.927, precision of 0.826, recall of 0.792, specificity of 0.952, F1 score of 0.809, AUC-ROC of 0.946, precision-recall AUC of 0.838, and balanced accuracy of 0.872. The most influential predictors were inconsistency between reported accident and observed damage, contradiction among submitted documents, previous claims frequency, suspicion of staged accidents, and delay in claim reporting. The integration of expert-derived fraud indicators with XGBoost produced a robust and interpretable framework capable of supporting probability-based fraud risk scoring, claims prioritization, and more efficient allocation of investigative resources in automobile insurance.