Auto insurance fraud detection using machine learning

Abstract

Rising vehicle insurance fraud significantly undermines the profitability of insurers and unfairly increases premiums for honest policyholders. To combat this growing threat, advanced machine learning (ML) techniques offer a promising solution for detecting fraudulent claims with greater accuracy and efficiency. This study develops and evaluates an ML-based fraud detection system using rich claim datasets that capture policyholder details, vehicle specifications, and claim attributes such as accident history and claim values. Four ML algorithms—Logistic Regression, Random Forest, K-Nearest Neighbors (KNN), and XGBoost—were trained and assessed using key performance metrics, including accuracy, precision, recall, and F1-score. The results indicate that while Random Forest and XGBoost achieved high accuracy, they exhibited lower recall, making them less effective in identifying fraudulent claims. In contrast, KNN and Logistic Regression demonstrated superior recall, essential for minimizing undetected fraud. Further optimization through hyperparameter tuning and ADASYN resampling improved KNN’s recall to 0.52 and its AUC score to 0.71, while Logistic Regression maintained a recall of 0.60. Based on its balanced performance and interpretability, Logistic Regression was selected for deployment in a web-based fraud detection system. The study concludes that implementing ML-driven fraud detection can significantly reduce fraudulent payouts, streamline claims processing, and enhance customer satisfaction. Future research should explore the use of more recent datasets, deep learning techniques, and alternative resampling methods to further refine fraud detection accuracy. Expanding the model to include other types of insurance, such as life and health, could enhance its applicability across the industry.