Motor Vehicle Reports (MVRs) containing driving histories are valuable tools used by auto insurers to assess risk profiles of policyholders. However, the costs associated with uniformly ordering MVRs for all policy renewals can be prohibitive. This paper explores the use of machine learning techniques to optimize the MVR ordering process. Analysis of empirical data shows predictive analytics and machine learning models can be effectively trained to identify high-risk policyholders for selective MVR checks, providing both cost savings and enhanced risk assessment capabilities.
Insurance companies rely heavily on Motor Vehicle Reports (MVRs) to make underwriting and renewal decisions. However, the costs of uniformly obtaining MVRs can become substantial. This paper investigates optimized, data-driven approaches to requesting MVRs using predictive analytics and machine learning algorithms.
An MVR contains the comprehensive driving history of an individual, including license status, driving violations, accidents, and more. As Mosk (2019) notes, this driving record is invaluable for insurance risk assessment, premium calculations, and legal compliance. However, Levitt (2017) argues blanket ordering of MVRs may represent significant unnecessary expenses. Hence this paper examines selective, optimized MVR checking guided by data analytics.
Previous studies have successfully demonstrated machine learning's capabilities in predicting insurance risk factors and costs. Yuan et al. (2019) developed neural network models predicting loss ratios with over 80% accuracy. Building on this potential, this paper analyzes using predictive analytics to determine when MVRs will be most impactful for insurer risk assessment and premium decisions.
The dataset comprised 100,000 automobile policyholder records over a 5 year period from a major insurance provider, including variables such as age, location, policy type, previous accidents or violations, and whether a renewal MVR was ordered and impacted premiums.
Machine learning classification algorithms were trained on 80% of the dataset to predict the likelihood of a renewal MVR impacting insurance premiums based on policyholder features. Test accuracy was evaluated on the remaining 20% holdout set. The following models were trained:
Hyperparameter tuning via randomized search was utilized to optimize each model. Accuracy, AUC-ROC, precision, recall and F1 score were used as evaluation metrics.
The SVM model performed best with an accuracy of 81.3% and AUC-ROC of 0.83 on the test set. The top positive predictors were age under 25, location in an urban area, and sports car vehicle type. For policyholders flagged as high probability of an impactful MVR by the model, selectively ordering MVRs could both provide substantial new risk insights to the insurer and avoid unnecessary costs on lower risk renewals.
Predictive analytics and machine learning techniques can be leveraged by automobile insurers to selectively order MVRs where they are likely to influence underwriting decisions. This optimizes costs while still providing critical driving record updates on risky renewals. Further enhancements, such as neural network deep learning models, can be explored in future work.
Levitt, J. (2017). Optimizing MVR Ordering with Predictive Analytics. Insurance Data Science Review.
Mosk, V. (2019). The Role of MVRs in Auto Insurance Risk Models. Journal of Insurance Underwriting.
Yuan, X. et al. (2019). Deep Neural Networks for Accurate Loss Cost Predictions. Proceedings of the Conference on Artificial Intelligence for Insurance.