Machine Learning in Insurance Underwriting Context

This paper discusses different machine learning models that can be used for insurance underwriting. The underwriting process for insurance is conventionally done by well-trained individuals, but the whole process is time-consuming and tedious, especially as the applications are getting increasingly more complex nowadays. The advancement in technology has enabled us to seek an alternative solution to this tedious and labor-intensive process of underwriting. Many believe that tools from Artificial Intelligence and more specifically, methods in Machine Learning, can provide time-efficient and accurate evaluations of risk. Using the publicly available dataset from Prudential Insurance, several feature engineering methods are implemented and tested for their usefulness. Specifically, kth nearest neighbor, multinomial logistic regression, random forest, and gradient boosting trees are fitted, and gradient boosting is shown to have the best performance and robustness.