Should All Your Customers Be Retained?

There’s an enormous focus on managing customers appropriately at the point of contact across all channels, whether it be direct mail, sales inquiry, customer service call or in the retail environment.

As CRM systems are constructed, companies are able to build a detailed picture of a customer’s behavior using transaction information, often enhanced with demographic data. However, most businesses using CRM systems do not embrace the strategic, long-term functions of marketing that this technology provides, such as planning and forecasting. Nor do they take into account this wealth of intelligence to develop new products or pursue new market opportunities.

As all marketers know, Pareto’s principle states that 80% of a firm’s profit comes from 20% of its customers. The problem is determining the 20% that are the most profitable.

With many industries hampered by churn issues, identifying profitable customers has become a competitive necessity. If a company is to spend its limited marketing budget on customer retention, such expenditures should be directed at those with the greatest potential lifetime value.

There’s no magic formula to determine who makes up “best customers,” though data mining can unlock some knowledge. Data mining uses statistical modeling techniques to predict a customer’s response or purchase based on the combinations of transactional and demographic variables known about that individual. These techniques can also be used to estimate measures of value, such as revenue obtained or profit derived from a customer.

If a series of models can be constructed, beginning with one to identify those customers most likely to churn and another specifying those with the greatest potential lifetime value, a business can start to get an idea of where to target its retention expenditures.

When this paradigm is extended further to develop a series of models to predict other products or services in which a customer is likely to be interested, specific communications can be targeted for these individuals and crafted. The ultra-competitive wireless carrier industry can be used to demonstrate the real-world application of these concepts.

In a recent survey, it was estimated that 34 million wireless customers worldwide would transfer their business from one carrier to another this year. By 2004, this number is expected to reach 77 million. The average wireless carrier in the United States loses about 2.5% of its customers each month.

Consequently, wireless carriers must reduce the number of customers who churn, identify high-value customers, decide what other products/services may interest them, and create personalized communications to targeted segments of their customer base.

Using data obtained from one such carrier, the methodology that could be applied to construct churn and customer value models and identify cross-selling opportunities can be illustrated.

The objective of a churn model is to look at a set of data with a dependent variable, such as likelihood to churn, which includes records incorporating both positive and negative outcomes. In the churn model, the dependent variable is “did not renew a contract.” To balance this, we need a sample of customers who had contracts during the same period and decided to renew. Attached to this sample data will be all demographic and transactional variables.

Using a range of statistical techniques, we can produce a series of predictive models. If we randomly select 50% of the validation set, we would expect to find 50% of the known churners in this portion of the file. But with the different models generated, we can identify nearly 80% of known churners.

The variables should be ranked in order of importance. In this example, the number of mailings a customer has received is the most important variable, followed closely by the number of months since the service was first purchased.

A look at the customer profile of this variable shows that customers who purchased in the past 11 months had a 15.3% chance of churning, compared with those who first purchased more than 65 months ago, who had only a 2% chance of churning.

Does a business really want to keep all potential churners? Probably not. What a company does want to do is retain those customers likely to contribute the greatest revenue or profit to the organization. Therefore the next stage is to build a model that predicts future revenue expected from these customers.

Again, we can apply these models to a validation set to assess its performance. The model has predicted what percentage of the customer base will generate the greatest revenue. The actual annual revenue from these customers is about $388, compared with the average across all customers, which is about $150. The bottom 10% of the file averages around $50.

By applying the churn model to the customer database, we can allocate a probability to each record that the customer will churn. Then by applying the value model, we can estimate each customer’s revenue. The combination of the two identifies the highest-value customers likely to churn. These individuals are now the principal targets for the portion of the marketing budget allocated to churn reduction.

The final stage of the analysis is to identify which products or services each customer is likely to buy. This can be achieved using a cross-sell model, which will produce a series of models for each individual product or service together with a cluster map outlining products offered. For our sample telecommunications company, such analysis showed that flat cell rate and voicemail are the most strongly related. Therefore, they could be packaged together for positive results.

The final stage is to use this analysis to create personalized marketing messages to be included on customers’ statements. Customers would receive these targeted offers if they were identified as potential churners, had a high potential value, and did not own the product that the model showed they would be interested in purchasing.

The effectiveness of this approach can be gauged by assuming the company has 1 million customers. With a 20% churn rate, this would equate to 200,000 potential churners each year. Based on our analysis in building the revenue model, we found that the average revenue for a customer in the top 30% by value was $267.

Andrew Greenyer is director of customer relationship solutions for Group 1 Software, Lanham, MD.