Enhanced Churn Prediction Using Stacked Heuristic Incorporated Ensemble Model

Introduction

Due to peer competition and market saturation, telecommunication sector is facing crucial challenges in terms of customer retention and acquisition. The cost of acquiring a new customer is very high compared to retaining existing customers (Hadden 2007). It is estimated that new acquisition costs 5 to 6 times more than retaining existing customers (De Bock 2011). In advanced countries, the number of telecom subscriptions is found to exceed the total population by a huge margin and that explains the saturation rate in the industry (Hung 2006). Given the market scenario, telecom operators are shifting their focus from acquiring new customers to retaining existing customers. The process of predicting which customers will leave the organization is called as churn prediction. It is the process of assigning a churn probability to each customer based on the historical information. The customers are then ranked with decreasing probability levels and marketing campaigns are performed on customers with high churn probability levels.

According to Huang (2013) improvement in customer retention will impact profit positively and further determine the sustainability of telecom business. Hence identifying churners is a crucial aspect of a company’s retention campaign. This process, if performed well can result in a substantial increase in the profits compared to providing offers to random individuals or to all the customers. However, customer churn prediction is a very complex process specifically due to the large number of dimensions involved in the customer data. Further, the large number of decision points that are to be considered for the decision-making process also increases the complexity of the prediction mechanism. Further, the domain is also imbalanced, with large number of non-churners and very few churners. Since customer churn is also based on several temporal factors, data used for training can only contain recent data. Hence imbalance tends to play a vital role in complicating the prediction process.

This work on customer churn prediction concentrates on a cost-based approach to minimize loss due to customer churn and increase the probability of retaining high value customers. The major contributions of this work are

• •

  To develop an effective churn prediction model that can demarcate between churners and non-churners with high precision

• •

  To effectively handle the data imbalance contained in the churn prediction domain

• •

  To minimize the loss occurring due to churn

• •

  To provide a cost-based prediction model that can effectively reduce the prediction cost