Customer Churn Prediction in Insurance Sector using Machine Learning and Optimization Techniques

Abstract

The term Customer Relationship Management, or CRM for short, states to a strategy for handling an organization s relations and interactions with existing and prospec- newlinetive clients. The components of customer relationship management (CRM) are as fol- newlinelows: identifying customers, attracting new customers, keeping current clients, and de- newlineveloping relationships with existing clients. In this context, keeping current clients (also known as Client Retention) is one of the most important elements in the profitable oper- newlineation of any business. According to the research that has been conducted, it is between four and five times more lavish for a firm to attain a new client than it is to keep the customers that they already have. Churn can be defined as the breaking of a convention newlinebetween a client and a company, or in other words, the discontinuation of a client s use newlineof a item or service provided by the relevant organization. The primary necessity for newlineany organisation is to make predictions about clients who are most possible to churn newlinefrom a service in advance. The ability to accurately find customer churn is extremely newlineimportant in certain fields, particularly those concerned with the provision of tangible or intangible services, as well as insurance services. newlineThe majority of the existing predictive modeling methods are based on supervised newlinelearning. These methods include Artificial Neural Networks (ANN), Decision Trees(DT), newlineInductive Rule Learning(IRL), Logistic Regression(LR), K-Nearest Neighbour(KNN), newlineNaive Bayesian(NB) classifier, Random Forest(RF), and Support Vector Machines(SVM). newlineThese techniques are used for the problem of predicting how many customers will leave newlinean insurance company. Unsupervised learning techniques have been engaged in a lim- newlineited number of studies. Existing works in churn prediction demonstrate that classi- newlinefication based on a single model fails to yield satisfactory outcomes and that hybrid newlineand ensemble models are good alternatives for improved performance in classification. newlineThese finding