Fair and Transparent Underwriting: Advanced AI Models for Thin-File Vehicle Insurance

The insurance industry struggles to predict insurance claims accurately. This prediction is crucial for running operations smoothly, preventing fraud, and setting the right premium prices. Traditional manual methods of claim evaluation are slow, subjective, and not capable of effectively handling the large volume of data from policyholders. This work comprises a machine learning (ML) framework for the prediction of vehicle insurance claims that supports the resolution of the hard-core problem of the identification of claim probabilities based on the different policyholder demographics and vehicle characteristics. A pipeline was systematically designed that features data preprocessing with cleaning, one-hot encoding, and outlier removal, followed by feature engineering and SMOTE-based class imbalance mitigation to tackle the issue of severe imbalance when no-claim instances account for 93.6% while claim cases make up only 6.4%. A hybrid CNN-GRU model was trained and evaluated using a variety of performance criteria. This model combines convolutional neural networks for the extraction of spatial features with gated recurrent units for the learning of sequential patterns. Using Random Forest (86.77%), Naive Bayes (95.28%), and MLP (76%) models, the proposed model was comprehensively assessed. CNN-GRU outshined other methods in terms of performance and attained the following measures: accuracy of 98.34%, precision of 98.45%, recall of 99.21%, F1-score of 98.56%, and AUC of 1.00, thus greatly surpassing traditional models. SHAP analysis showed that car age, policyholder demographics, and vehicle specifications are the main factors contributing to the prediction. This confirms that hybrid deep learning (DL) architectures are robust, scalable solutions for real-time insurance claim prediction systems.