Modelling of Life Insurance Cancellation

This study focuses on prediction of life insurance contracts cancellation. The information base of this research is a set of 15 688 life insurance contracts which spans the period from 2008 to 2015 from one of the largest Russian insurance company. In this study, the following mathematical models were used: binary logistic regression, Cox regression, Random Forest method and Neural Network model (multilayer perceptron). In addition optimal cut-off thresholds were determined for each model. As a result, the best performance were shown by models based on the Cox regression and Random Forest method. The accuracy of the Cox regression is 69.5%, and accuracy of the Random Forest method is 88.9%. In the future, these models may find its application in the practice of insurance companies to determine the customers most likely to cancel the contract, and therefore decrease the churn rate.