Comparison of the Prediction of the First Class Automobile Insurance Renewal between Random Forest Model

Abstract

This paper aims to compare the prediction of the first class automobile insurance renewal between the random forest model and adaptive boosting model, and to study the partition proportion between the training and testing dataset. Secondary data were obtained from an insurance company in Thailand from 2018 to 2019 with 10,000 datasets and 8 interested variables. The response variable in this study is the renewal status of each policy, e.g., renew or not renew. The independent variables are the gender of the policyholder, car premium, car year, number of claims in one year, number of incurred losses in one year, car premium discount, and car premium surcharge if the incurred loss excess the policy limit by using RStudio version 1.2.5033 program with rattle package for prediction of the first class automobile insurance renewal. The result showed that the adaptive boosting model with a proportion of training and testing dataset as 85:15 is an appropriate model with 66.90 % accuracy and 78.44 % of overall efficiency. This adaptive boosting model is a little better than the random forest model. The most important independent variables for insurance policy renewal are car premium and the number of claims in one year, followed by the car premium discount.