Assessing risk in life insurance using ensemble learning

Abstract

Risk assessment is an important aspect of decision making while granting policy to an applicant. In the vast economy with enormous feature criteria for everyone, it is an ongoing challenge for the insurance companies to assess each applicant based on various factors to provide right policies on the basis of a risk score. We propose a method of ensemble learning as a solution to this problem where the predictions from pre-existing supervised learning algorithms can be used to enhance the accuracy of prediction. A real-world dataset having 128 attributes has been used to study the risk value associated with a policy applicant. Machine learning algorithms were applied to the dataset to predict the risk associated with the applicant. Two ensembles have been used for classification of risk level assigned to a person which further leveraged our approach to an optimized and efficient class of predictors namely ANN and gradient boosting algorithm XGBoost. As a result, we discovered that the XGBoost algorithm with optimized hyperparameters gave us the best results in terms of Quadratic Weighted Kappa Score. The proposed methodology outperforms other existing methodologies as discussed in the later sections of the paper.