A COMPARISON STUDY OF PREDICTIVE MODELS FOR CUSTOMER LIFETIME VALUE AND THE CHANGE OF ITS LEVEL

Abstract

The main aims of this research are to construct and compare models to predict the change of Customer Lifetime Value or CLV level in two consecutive years. The data used in this work is collected by a case study insurance company from 2018 to 2020 containing the following features: gender, the number of consecutive years with the company, the number of active policies, the total insurance premium, the total coverage, the total policy profit and CLV level according to company criteria (Standard, Plus, Extra and Ultima). The data from 2018 and 2019 consisting of 541,371 customers are used to construct the ordinal logistic regression and Random Forest models and to check the model accuracies, respectively. Then the data from 2020 consisting of 1,029,001 customers is used for predicting future CLV.

            The results show that, in term of predicting CLV, the Random Forest model (75.01%) has a higher accuracy than the ordinal logistic regression model (65.60%). However, both models cannot be used in detecting the change of CLV level providing the accuracies of 15.57% and 25.74%, respectively. After adjusting the predicting threshold criterion for the ordinal logistic regression model, the total accuracy and the change of CLV level accuracy are improved to 74.60% and 52.16%, respectively. Therefore, the ordinal logistic regression model with the new threshold criterion is selected as our final model. By using the final model, the 1,029,001 customers are classified as 40,786 Standard customers, 809,951 Plus customers, 168,389 Extra customers, and 9,875 Ultima customers.