Using Predictive Analytics to Battle Churn and Increase Profits

Churn remains one of the most compelling challenges in the telecom industry, and by all indications the problem is likely to get worse before it improves. With providers annually losing as many as one-third of their customers, strategies to acquire and retain customers have become critical. At the same time, operators are losing significant revenue because of the high cost of retaining unprofitable customers.

Churn in the mobile industry costs operators collectively about $9 billion in lost revenue and $7.5 billion in lost acquisition costs every year. In addition, research firm In-Stat/MDR expects about 46 percent of U.S. customers will switch to a rival operator in the year after wireless local number portability takes effect, compared to the current rate of 30 percent.

The figure is even higher for ISPs—pegged at between 36 and 48 percent—adding up to a loss of one-third to one-half of all new subscribers acquired in a year. Cable broadband providers lose on average 2 percent of their customers every month, a hefty loss when one considers the $300-$600 average cost to initially acquire those customers.

To combat this problem, service providers are attempting to make better use of their existing data to prevent customers from defecting and make existing customers more profitable. However, when it comes to customer data, collecting everything means nothing. According to research firm Gartner, analytics have increased in importance as enterprises recognize their potential for alleviating the paralyzing condition known as ‘infoglut’—an overwhelming information and data overload. Enterprises may pay for their failure to invest in analytics with decreased productivity and inferior decision making.

As next-generation mobile networks and bundled voice, video and Internet services create more usage and customer data, the need for business intelligence solutions also will grow. Gartner predicts that by 2012, providers will process 30 times more data than in 2002. The key is to pinpoint the crucial details from this sea of information.

Using the Past to Predict the Future

Predictive analytics—using models to predict future events from today’s data—enables service providers to gain some benefit from the volume of data it collects on an ongoing basis.

Pioneered by the credit card industry to predict the credit risk of customers or the likelihood of fraud, predictive analytics solutions have hit the mainstream in the telecommunications industry. For example, when Yankee Group recently asked operators in North America to identify their top business intelligence initiatives for the next 12 months, customer and product profitability analysis ranked the highest priority. Customer churn analysis and predictions around market segmentation followed.

Predictive analytics tools must not only predict which customers are at greatest risk of churning or most likely to purchase new services, but must also recommend specific actions to help save a customer or increase ARPU.

For this to be successful, analytics tools must be used on a one-to-one basis, targeting the individual customer, not the customer base as a whole.

Predictive analytics solutions use algorithms and machine-learning techniques to make connections between bits of information about a provider’s customer base to predict future behavior. Machine-learning techniques are ideal for analyzing large data sets like those found in the telecommunications industry where relationships within that data are complex and frequently changing.

Predictive analytics deploys numerous types of modeling methodologies, including the use of neural networks—artificial intelligence that attempts to mimic the way a human brain works. By correlating subtle similarities in hundreds of data categories that would otherwise go undetected in a conventional model, neural networks—like the brain—can connect the dots and extrapolate important information about an operators’ customer base.

These trends are specific to each individual operator, which is where the benefits of machine learning come in. These predictive models detect the trends in the operator’s own data, and thus give a service provider enough time to prevent a customer departure.

For this to work effectively, the entire modeling process must be operationalized, i.e. made to work reliably in the high-volume, mission-critical environments of the billing and customer care systems. On a frequent basis—for instance, weekly—subscriber data must flow reliably from the billing and customer care systems, through a data warehouse suitably structured for analytic processing, through a modeling and prediction engine, and finally be updated back to the operational systems to impact the business. This tight integration to operational systems is one of the major stumbling blocks in implementing a predictive analytics system.

Analytic data models and human analysis combine to return recommendations on upselling and retaining individual customers. To accomplish these goals, one of the first distinctions to understand is the difference between customer value and customer profitability.

Defining Customer Value Versus Customer Profitability

Customer value takes into account the entire relationship, including factors such as customer referrals, customer lifetime, and ongoing support costs. Profitability is purely financial—how much revenue is the customer bringing in versus the costs of delivering the service month-by-month.

To gain a more accurate view of customer profitability, the comparison of revenues against costs takes place at the individual subscriber level. The answers about customer profitability exist within the information that providers already collect from billing engines, data warehouses, data mediation applications, CRM applications, self-care and EBPP tools. The precise factors to include in a profitability calculation vary widely from operator to operator, though it is generally straightforward to determine monthly profitability once the necessary data has been gathered together.

Some typical costs that impact customer profitability include call center support, churn and package churn (a subscriber that drops a particular package of services) and network capacity used to provide the service.

For example, a text message may have a 90 percent profit margin for the carrier. Voice may yield a 50 to 60 percent margin for the carrier. To increase customer profitability, service providers can determine which users are willing to pay a higher price to send a text message, resulting in more profit for the provider and shorter usage of the network.

Identifying the Right Data to Make Better Business Decisions

How does a provider determine which existing data will help address specific business problems? Identify the correct data to get to the “why” of the problem, then determine if the data that will provide the answer is being captured. For example, if a provider wants to know why customers in a certain ZIP code are churning at a higher rate than other customers, is location-specific data such as a record of service outages available?

Using data modeling to define what data is relevant and important can help service providers resolve these questions. The combination of automated modeling and human analysis can categorize the data gathered to give a more detailed focus into the critical indicators of “why” customers do what they do.

To take action on existing data, service providers must be able to make the information actionable by CSRs during live customer interactions. Even the most insightful data reports can’t save an individual customer from churning if the CSR can’t access the data in an easy-to-use format and put the information to immediate use.

False Predictors – Predictive Analytics’ Unique Data Problem

False predictors are unique to predictive modeling and are a by-product of the nature of modern data warehouses. Predictive models rely on past patterns to predict future patterns, yet most data sources provide only a snapshot in time. A false predictor arises when fields in the source databases falsely appear to predict the model target, when that relationship is actually an artifact of the data source.

For example, most telecommunications databases will have a list of equipment items in the subscriber’s possession such as mobile phones and cable boxes. When a subscriber churns, most companies delete or reassign all equipment records belonging to the customer. This makes the count of subscriber equipment records a false predictor, because if the subscriber has no equipment, they have already churned.

The false predictor actually throws the model off; predictions made on customers with equipment records greater than zero will be worse than they should be. However, simply discarding this field from the model is less than ideal. For instance, the count of equipment records before a subscriber churns is almost always useful for predicting churn. In practice, false predictors are very common, and eliminating them can require a combination of both human and software intelligence.

Return on Investment

What results can be achieved by service providers experimenting with predictive analytics as a course of daily operations? One CSG Systems analytics customer has seen significant results almost immediately after training CSRs on its use.

Of the customers identified at highest risk of churning, the provider saved more than 80 percent of customers contacted by offering successful retention strategies. Similarly, of customers contacted with upsell offers, 95 percent accepted the offers.

As a result of higher upsell and cross selling to receptive customers, the provider increased its average ARPU by $287 for every customer that purchased new products or services when offered them.

Although not all churn problems can be addressed through call center agents, many service providers don’t want to wait for a customer to call and complain before making an outreach. Fortunately, using predictive analytics to increase proactive outreach to customers most likely to churn can produce a compelling return on investment.

Take, for example, an operator with 1 million subscribers that experiences a 1.5 percent churn rate. This operator generates $40 on average in revenue per subscriber and pays $300 to acquire each new customer (assume the operator spends only $40 to retain a customer).

Using conventional outreach methods to contact 2 percent of its subscribers, the operator is able to reach only 5 percent of the customers at highest risk of churning. Through this “guesswork” method, the operator experiences net annual savings of $728,000 in churn reductions and upsell opportunities.

Now assume the same operator uses a customer analytics engine. By contacting the most appropriate 2 percent of its customer base, that operator has reached 23.5 percent of customers at greatest risk of flight.

In this case, the operator realizes net annual savings of $3.76 million, even when taking into account the cost of implementing predictive analytic tools.

Such examples of significant cost savings show that analytics can play an important role within telecom networks.

With the application of these tools, telecommunications service providers will improve their customer relationships and ultimately build a stronger, more profitable business, one customer at a time.



Dr. Richard Wolniewicz holds a Ph.D. in Computer Science, Database Systems and specializes in the processing and transformation of very large data sets for analysis. He oversees the Advanced Technology Group, part of the CSG Systems’ Analytics team, including experts in machine learning and predictive technologies. He can be reached at richard_wolniewicz@csgsystems.com.