Abstract:
As markets have become increasingly saturated, companies have acknowledged that
their business strategies need to focus on identifying those customers who are most
likely to churn. To address this, a method is required that can identify these customers,
so that proactive retention campaigns can be deployed in a bid to retain them. To
further complicate this, retention campaigns can be costly. To reduce cost and
maximise effectiveness, churn prediction has to be as accurate as possible to ensure that
only the customers who are planning to switch their service providers are being targeted
for retention.
Current techniques and research as identified by literature focus primarily on the
instantaneous prediction of customer churn. Much work has been invested in this
method of churn prediction and significant advancement has been made. However one
of the major drawbacks of current research is that the methods available do not provide
adequate time for companies to identify and retain the predicted churners. There is a
lack of time element in churn prediction. Current research also fails to acknowledge the
expensive problem of misclassifying non-churners as churners. In addition, most
research efforts base their analysis on customer demographic and usage data that can
breach governing regulations. It is proposed in this research that customer complaints
and repairs data could prove a suitable alternative.
The doctoral research presented in this thesis aims to develop a customer profiling
methodology for predicting churn in advance, while keeping the misclassification levels
to a minimum. The proposed methodology incorporates time element in the prediction
of customer churn for maximising future churn capture by identifying a potential loss of
customer at the earliest possible point. Three case studies are identified and carried out
for validating the proposed methodology using repairs and complaints data. Finally, the
results from the proposed methodology are compared against popular churn prediction
techniques reported in literature. The research demonstrates that customers can be
placed into one of several profiles clusters according to their interactions with the
service provider. Based on this, an estimate is possible regarding when the customer
can be expected to terminate his/her service with the company. The proposed
methodology produces better results compared to the current state-of-the-art techniques.