Churning in the telecommunications industry
Source: www.spss.com
Copyright SPSS, Inc. 2004


Churn is the process of customer turnover. In the mobile telecommunications market estimated the cost of churn at around $400 per new subscriber. Churn is not restricted to the telecommunications market – it occurs wherever stiff competition provides incentives for customers to switch providers. Industries such as credit card issuers, insurance companies and ISP’s are all subject to increasing levels of churn. ISP’s have hinted at churn rates as great as 50 percent, the mobile telecommunication rate appears to vary around the 30 percent mark. As the market plateaus in these industry sectors (save perhaps for ISP’s) the problem of churn is becoming ever more critical.

Churn analysis using data mining

In order to combat the high cost of churn, increasingly sophisticated techniques can be employed to analyze why customers churn and which customers are most likely to churn in the future. Such information can be utilized by marketing departments to better target recruitment campaigns and by active monitoring of the customer call base to highlight customers who may, by the signature in their usage pattern, be thinking of migrating to another provider.

Data mining techniques bring powerful modeling analytics to the identification of the “Why” and “Most Likely” groups. This paper describes a typical application of such techniques to a mobile telecommunication provider’s customer base.

Customer churn analysis 

Based on an intuitive visual data flow process of analysis, models and reports can be created to tackle otherwise difficult and time consuming problems. You may consider the data set representing a sample of customers with descriptors such as number of minutes on long distance and local calls, financial information such as billing types and payments methods, some demographics such as age, sex, income. The final field “Churned,” indicates whether the customer has ceased usage on their own accord (Vol), or are still a current paying customer (Current). The InVol customers are those who have been removed by the telecommunications provider for reasons not due to the churn we are interested in, i.e. payment defaults, etc.

The task is to profile those customers most likely to be voluntary churners.

Types of analysis

Decision trees and neural networks are the staple of advanced knowledge discovery workbenches. These techniques can be effectively applied to the churn problem. Neural networks are able to uncover complex patterns in the types of customers and rank the customer base based on a score or likelihood, to churn. Although powerful, neural networks are difficult to interpret and do not readily give up the process by which a customer has been scored.

Decision trees on the other hand build very open and interpretable models that show the analyst the patterns discovered. For example, a decision tree algorithm may discover the following rule.
If international call time is 10 minutes and the long distance bill type is standard, then churn likelihood high.

This rule can lead to actionable knowledge, it may be worth calling this customer and suggesting a change of billing type to one more suited to this level of international calling.

Neural networks, decision trees and even clustering algorithms can be helpful in the analysis of customer churn.

Feature selection

In any data analysis there is a process. Specifically:

· Business question definition (in this case churn analysis – who is churning? And can we develop an action plan to combat that churn?)
· Data gathering – build the data sample. Summarize customer-calling behavior and join with relevant demographics and billing information
· Feature selection – a combination of identifying poorly populated (noisy?) attributes and attributes not deemed relevant to the task at hand. This is done by visually inspecting the data and, together with a domain expert, identifying those attributes most likely to impact the churn decision
· Modeling – build the models
· Analysis and report generation

After exploring the data the next stage is to begin modeling. The process of feature selection is important. The problem is that some modeling algorithms cannot cope with large numbers of attributes. In many real data sets a user may have hundreds of attributes and this can lead to poor performance of the modeling algorithm. One technique is to reduce the number of attributes to the core set of highly indicative attributes by using decision trees. Decision trees do not have a problem with large numbers of attributes. 

Once the core set of attributes has been selected, a neural network (or other scoring algorithm) may be trained using the remaining attributes. Once trained the customers can be sorted into a ranked list of high probability to churn down to the least likely to churn (the loyal customers).

Value and profitability analysis

Even when scoring models have been built, the analysis is by far from over. In some ways it has only just begun. The reason is that there are many ways to use a churn-scoring model – all governed by the business task and problem being tackled. If the goal is to “touch” all customers no matter the cost, then we don’t even need a model. If the goal is to maximize the return on a marketing budget of a fixed amount, then modeling is critical.

Other issues such as customer value begin to play a part. What to do with customers, who are likely to churn, but do not in fact use their mobile phone? If the telecommunications provider is not making money from this customer, should they not just let this customer go? This problem is perhaps even more apparent with credit card issuers.