requires having historical data on the types of products existing
customers have purchased and various attributes of those customers
such as demographics, behaviors, etc.
Once we have identified the customers who are likely to purchase,
we can further assess which of these customers are likely to be profit-
able. This also requires historical data containing the profiles of
customers deemed to be profitable and unprofitable. Certain classifi-
cation techniques, such as decision trees, can produce a set of profiles,
or rules , highlighting the characteristics of profitable customers. This
information can then be used to select profitable customers.
Now that we know which customers are likely to purchase, which
of those are likely to be loyal, and which will be the most profitable,
we can perform response modeling to determine who is likely to
respond to a campaign. Then, we may even go one step further to
determine which channel is best for contacting such customers.
Once a business has customers, one of the next problems is how to
keep those customers. Customer retention, or answering the question
“how do I keep my current customers?” is a problem faced by busi-
nesses in most every industry. For example, in financial services, a
customer who leaves is called an “attriter” and the problem is
referred to as “attrition.” In telecommunications, a customer who
leaves is called a “churner” and the problem is referred to as
“churn.” Regardless of the terminology, the basic problem is the
same: which customers are likely to leave and why?
Customers may leave for many reasons, for example, poor service,
moving out of the area, or the availability of more competitive offers.
However, these reasons are not always obvious until after the fact.
An effective customer retention effort often requires identifying cus-
tomers before they leave so that some action can be taken, if war-
ranted, to retain those customers. We say “if warranted” because
some customers may not be worth retaining. Customers who have
low value or represent a net loss to the business when considering
support and maintenance costs fall into this category.
Data mining can be applied to identify characteristics of individu-
als and their past and current behavior to determine much more
subtle indicators of attrition or churn. For example, in wireless phone
service, a customer whose minutes of usage drop from a four-week
moving average of 500 minutes per week to 50 minutes per week