Putting the Brakes on Churn

Posted: 01/1999

Putting the Brakes on Churn
By John Andrews

If you run a growing telecommunications company, imagine 10 of your most valued
subscribers sitting in your office. Then picture four of those customers standing up,
slamming the door and walking away from your company forever.

Not a pretty thought. But with churn rates now exceeding 3 percent a month in many
telecom markets, most operators can and will lose up to 40 percent of their current
customers within a year. As many as five providers now compete for subscribers in some
markets, and with the biggest players now offering single-rate packages as low as 10 cents
a minute, both average revenue per users (ARPUs) and telecom profits are on the decline.
With customer-acquisition costs now approaching $400 per subscriber, churn alone costs
U.S. and European operators an estimated $4 billion each year.

Telecom managers are cutting costs, boosting their marketing efforts–and searching
aggressively for ways to reduce the rate and cost of subscriber churn.

Today more and more companies are turning to a powerful new customer-retention
technology called predictive-churn modeling. Predic-tive modeling analyzes historical
subscriber records to identify and model prechurn behavior, and then applies that profile
to current customer lists to identify those who are most likely to churn. For companies
hoping to keep customers on board–and that’s just about every telecom provider these
days–predictive modeling can be used to spot, target and retain at-risk subscribers.

Scientific Forecasting

Some progressive organizations recognized the benefits of predictive modeling as early
as 1996. A few of them built their own in-house modeling systems, or purchased fairly
simple off-the-shelf software packages. Those solutions may have worked for a time, but in
today’s fast-paced telecommunications business, it soon became apparent that more
sophisticated systems were needed. Software that relied on a single, hard-to-update model
provided very limited benefits in markets in which competitive conditions and customer
preferences changed almost monthly.

Fortunately, a new and far more powerful generation of predictive- churn modeling
technology now has emerged. These newer systems make use of information gathered from a
broad range of company sources, including customer or subscriber data, usage histories,
rate plan and feature data, network information, customer credit and billing and payment
histories. Advanced techniques have been developed and refined for the creation and
application of the predictive model.

Image: Building A Predictive-Churn Model

Gathering Data

The process begins with the collection of a wide range of historical subscriber
information. This data is evaluated, streamlined and, in some cases, recoded for use in
the modeling process. Next, churn data is gathered for the period immediately following
the test-model time frame. This churn data includes details on whether a customer has
defected or not, deactivation data and, if available, a reason code explaining why the
subscriber left the company. This initial step is critical and must be completed
thoroughly and diligently. In many cases, this data-validation step can represent up to
one-third of the total time and cost of a predictive-churn project.

Modeling Databases

As shown in the flow chart above, the resultant data file is split in two sections on a
nonbiased basis. One data set, called the "learning" database, is used to create
the predictive churn models. Advanced churn modeling software will analyze this
"learning" database to evaluate subscriber behavior in the weeks and months
prior to the defection. This information automatically will be converted into
sophisticated algorithms that form the basis for the predictive model. The remaining data
will be saved as a "testing" database for use in a blind test that is conducted
later to verify the accuracy of the models.

Creating the Model

This resulting model, or profile, essentially is a complex statistical formula which,
when applied to available data on any set of customers, generates a "churn
score" for each individual. The churn score measures how likely a given customer is
to defect: the higher the score, the more likely a subscriber is to churn. By ranking
subscribers according to their churn score, a company then can target its retention
program to these high-risk prospects. This same technique can be used to create
"lifetime value" numbers to quantify the long-term revenue potential of each
subscriber.

Testing the Model

After it is refined and adjusted, the model is evaluated using the portion of the
original data set retained for this blind test. The model’s algorithms are applied to the
"testing" database, from which churn scores and a ranked churn list are created.
Because we already know the real-world outcomes for these customers, the model’s
performance now can be closely analyzed. The closer the match between the modeling
prediction and actual outcomes, the better the model. If needed, modeling predictions can
be improved by increasing the volume, sources or quality of raw data or by adjusting other
modeling metrics.

Applying the Model

Now that there’s a workable churn model, it can be applied to current customer data to
predict future churn activity. For each forecasting period, a new set of churn scores will
be calculated and a new ranked churn list will be created.

Evolving the Model

Existing models can be refined and new models can be built as the market changes over
time. The frequency with which this is done depends on various factors including the
competitive dynamics within a particular market. In some telecommunications markets, for
example, more than five wireless providers compete for customer loyalty. The dynamic
nature of these crowded markets virtually assures that subscriber behavior will shift
dramatically over time–and predictive models must be adjusted frequently to reflect these
key market changes.

Because market dynamics change so quickly, many companies that previously relied on
very simple software packages, or on systems developed in-house, now find it is extremely
difficult and time-consuming to keep their models up-to-date. More advanced predictive
modeling software–which employs high-performance data mining engines and constantly
updated algorithms, and automatically determines the best combination of modeling
methods–gives telecom providers access to more precise and up-to-the-moment
churn-forecasting results.

Of course, predictive modeling alone cannot solve the problem of subscriber churn. To
succeed, this method must be incorporated into an aggressive and comprehensive retention
program. Most successful programs also include competitive rate plans, network
improvements, proactive loyalty programs, more accurate reporting systems and
well-organized save/win-back efforts.

Adjusting the Model

Regardless of which solution a telecommunications company selects, to enjoy positive
results, predictive modeling must be applied on a consistent and regular basis. This
approach requires the extraction and importation of subscriber data into the modeling
application, or the creation of a direct link between warehoused data and the modeling
process. It also may be desirable to rescore customer records on a weekly basis to ensure
that changes in activity, payments or service translate to an updated churn score. In this
situation, a working model is used to score the full database on an ongoing basis, a
process that can be performed at the server/warehouse level or at the application/server
level, depending on a given provider’s needs.

Using subscriber data as a key element in a predictive-churn modeling system also may
require some adjustments in how companies design and manage their databases. It may be
necessary to create new data fields in accounting, customer contract or other systems. In
some legacy-based computer systems, nonvalidated data, such as handset identification or
feature types entered into a comment field, may not produce reliable results when used for
predictive analysis. Data cleansing may be required to use data that is valuable but not
formatted correctly.

Some adjustments also may need to be made in organizations that use subscriber
information to provide online support for customer service representatives (CSRs). If the
company wants its CSRs to use churn-based information on a proactive basis, this would
require changes in how the information is presented on the customer service system screen.
For example, a churn score field might be added to show the CSR how likely the customer is
to churn.

Bottom-Line Benefits

Predictive-churn modeling already has started to deliver solid, bottom-line customer
retention results for companies the world over.

In one real-world application, a European wireless operator determined that 40 percent
of its churners could be found within a 2 percent subset of its subscriber base. Upon
further investigation, the company also learned that a particular handset type was
associated with this troublesome segment. Armed with this key forecasting information, the
company launched a highly targeted intervention program designed to solve the handset
difficulties and prevent further significant customer losses.

In another recent example, a U.S.-based service provider used predictive churn models
to analyze selected billing data and reduce churn by 40 percent. Those are the kinds of
numbers that make customer-retention managers sit up and take notice.

Predictive-churn modeling can be a key element in a comprehensive and aggressive
customer-retention strategy. By applying the most advanced and cost-effective modeling
techniques, telecom providers can reduce both the rate and cost of churn. That means
keeping more subscribers, and more profits, right where they want them.

John Andrews is a senior account executive with SLP InfoWare Inc., Paris, a provider
of churn-management and customer-retention software for the telecommunications industry.
He can be reached in SLP’s Chicago office by telephone at +1 312 407 6597 or via e-mail at
[email protected].