SecureWheels: A Cost-Sensitive Hybrid Learning Framework for Motor Insurance Claim Fraud Screening

Motor vehicle insurance fraud accounts for billionsof dollars in global annual underwriting losses, while conventionalclaims adjudication relies predominantly on manual surveyorinspections, heuristic rules, or naive classifiers prone to majorityclass collapse. Claim assessment involves critical decisions undersevere uncertainty: the vast majority of claims ( 94%) arelegitimate, while fraudulent claims ( 6%) are rare, adversarial,and systematically concealed. A screening system must identifyhigh-risk anomalies without treating predictive scores as legallyconclusive proof of fraud. This paper presents SecureWheels, anauditable, cost-sensitive hybrid decision-support framework thatintegrates supervised bagging, gradient-boosted decision trees,and unsupervised anomaly isolation within an investigator-inthe-loop review pipeline. The analytical engine couples Random Forest (wRF = 0.45) and XGBoost (wXGB = 0.45, withscale pos weight = 15.67) to model complex non-linear attributeinteractions under extreme class imbalance, while an IsolationForest (wIF = 0.10) provides an unsupervised anomaly signalsensitive to zero-day fraud tactics. The framework is benchmarked on a standardized historical corpus of 15,420 realworld motor insurance claims using a strict 70:30 stratifiedholdout split (10,794 training vs. 4,626 test records) and 5-foldcross-validation. Operating under a Bayes cost-sensitive decisionthreshold (τ = 0.40) derived from asymmetric economic lossmatrices, the proposed tri-model ensemble achieves a ROC-AUCof 84.91%, Fraud Recall of 75.09%, and overall Accuracy of77.32%, intercepting 208 out of 277 holdout fraudulent claimswhile preserving straight-through processing for 3,369 legitimateclaims. Local and global explainability via TreeSHAP attributesprimary risk drivers, ensuring regulatory transparency andauditability.Index Terms—Motor insurance fraud, cost-sensitive learning,class imbalance, Random Forest, XGBoost, Isolation Forest,TreeSHAP, explainable AI, decision-support systems.