Churn models at mobile providers: The importance of network measures

Telecom session

Benedek Gábor (Thesys) — Ágnes Lublóy (Corvinus University of Budapest)
Churn models at mobile providers: The importance of network measures

Mobile providers all over the world face the phenomenon of customer churn on a daily basis. In their endeavour to expand their customer base, or at least maintain the number of customers at a constant level, the providers must stay on their toes in a fierce competition for new customers. Attracting new customers is more expensive than retaining old ones. In order to keep the number of switching customers at a minimum, mobile companies must identify the customers at risk and target marketing campaigns at them. Providers’ internal churn models are specifically constructed for this purpose. The data mining methods used in these models, however, disregard network effects. The value of each customer is determined on an individual basis, and the potential spillover effects resulting from social linkages between customers are not taken into consideration. However, if one key customer switches from one service provider to another, several customers having strong links to the key customer may do the same. Should this process start an avalanche of service switching, the provider’s loss will be far more substantial than originally predicted.

This study looks into the effects of customers’ network topological properties on churn probability, that is, the probability of a customer switching to another provider. The present study contributes to the churn literature by demonstrating that the network characteristics of customers are also important churn factors.

The research relies on real life phone call and SMS records of approximately 26 thousand customers calling or texting almost 800 thousand people within a period of six months. The records cover almost 15 million phone calls and 1.3 million short messages. The analysis employs a unique snowball-type sampling method. Starting with an initial database of 0.04% of customers resident in a city, the final sample covers 30% of the customers living in that city. The degree distribution of the core network has proved to be almost scale-free, and the network exhibits the small-world property. These two network topological properties justify the sampling method, as both of them are well-known characteristics of social networks.

The study explores the essential question of whether a substantial difference exists between retained and churned customers in terms of network characteristics. After defining seven customer-level network characteristics, the study segments the customers into two distinct groups with significantly different churn ratios. Meaningful segmentations have been achieved with reference to the network topological properties capturing the number of relations and the embeddedness of customers. By defining an appropriate threshold, the network topological description of each segment could be given. Thus, the study provides clear evidence that individual network characteristics have a considerable impact on churn probabilities. In addition to allowing a longer-term projection of churners, the inclusion of network-related measures in the churn model also improves the accuracy of the model.

Customers addressed by marketing campaigns and non-targeted customers are segmented separately. The findings suggest that targeted customers are less threatened by churn than their non-targeted peers. In order to have a churn probability comparable to that of targeted customers, non-targeted customers need to meet more restricted conditions in terms of degree, proportion of in-network call duration and (weighted) embeddedness. This phenomenon highlights the importance and effectiveness of the provider’s marketing actions.

The comparison of traditional churn models and risk assessment models (terminology proposed in this paper for the first time) emphasizes the difference in how far ahead they can make churn predictions. The viability of the risk assessment models is shown by demonstrating the significance of certain network metrics. Risk assessment models incorporate network topological properties, which are more stable over time than are phone usage patterns. Relatively long-term prediction is therefore unrealistic without network-related metrics. Even in the short run, the accuracy of the churn models might be improved by incorporating customer network characteristics into the models. The results suggest that mobile companies should update their current churn models by the addition of some network related metrics.