In other cases, the ongoing relationship is just a beginning. A credit card may send a bill every month; however, nothing charged, nothing owed. A long-distance provider may charge a customer every month, but it may only be for the monthly minimum. A cataloger sends catalogs to customers, but most will not make a purchase. In such cases, usage stimulation is an important part of the relationship.
Subscription-based relationships have two key events—the beginning and end of the relationship. When these events are well defined, then survival analysis (Chapter 12) is a good candidate for understanding the duration of the relationship. However, sometimes defining the end of the relationship is difficult:
■■ A credit card relationship may end when a customer has no balance and has made no transactions for a specified period of time (such as 3
months or 6 months).
■■
A catalog relationship may end when a customer has not purchased from the catalog in a specified period of time (such as 18 months).
■■
An affinity card relationship may end when a customer has not used the card for a specified period of time (such as 12 months).
470643 c14.qxd 3/8/04 11:18 AM Page 460
460 Chapter 14
Even when the relationship is quite well understood, there may be some tricky situations. Should the end date of the relationship be the date of customer contact or the date the account is closed? Should customers who fail to pay their last bill be considered the same as customers who were stopped for nonpayment?
These situations are meant as guidelines for understanding the customer relationship. It is worthwhile to map out the different stages of customer interactions. Figure 14.4 shows different elements of customer experience for newspaper subscription customers. These customers basically have the following types of interactions:
■■
Starting the subscription via some channel
■■
Changing the product (weekday to 7-day, weekend to 7-day, 7-day to weekday, 7-day to weekend)
■■
Suspending delivery (typically for a vacation)
■■
Complaining
■■
Stopping the subscription (either voluntarily or forced)
In a subscription-based relationship, it is possible to understand the customer over time, gathering all these disparate types of events into a single picture of the customer relationship.
Stop
Complain
Temporarily
Some Channel
Respond from
SUBSCRIBER
Voluntary
paying
Stop for
Churn
Other
Bill
Pay
Reason
Not P
SALE
ORDER
START
Create
Deliver
y Bill a
ay
Account
Paper
P
Not Pay
SUBSCRIBER
Forced
late paying
Stop Paying
Churn
Stop
Complain
Temporarily
Figure 14.4 (Simplified) customer experience for newspaper subscribers includes several different types of interactions.
470643 c14.qxd 3/8/04 11:18 AM Page 461
Data Mining throughout the Customer Life Cycle 461
Business Processes Are Organized around the
Customer Life Cycle
The customer life cycle describes customers in terms of the length and depth of their relationship. Business processes move customers from one phase of the life cycle to the next, as shown in Figure 14.5. Looking at these business processes is valuable, because this is precisely what businesses want to do: make customers more valuable over time. In this section, we look at these different processes and the role that data mining plays in them.
Customer Acquisition
Customer acquisition is the process of attracting prospects and turning them into customers. This is often done by advertising and word of mouth, as well as by targeted marketing. Data mining can and does play an important role in acquisition. Chapter 5, for instance, has an interesting example of using expected values derived from chi-square to highlight differences in acquisition among different regions. Such descriptive analyses can suggest best practices to spread through different regions.
There are three important questions with regards to acquisition, which are investigated in this section: Who are the prospects? When is a customer acquired? What is the role of data mining?
Acquisition
Activation
Relationship Management
Retention
Former
Customers
High
Value
Voluntary
Churn
Target
New
High
Responder
Customer
Market
Customer
Potential
Forced
Rest of
Low Value
Churn
World
Winback
Figure 14.5 Business processes are organized around the customer life cycle.
470643 c14.qxd 3/8/04 11:18 AM Page 462
462 Chapter 14
Who Are the Prospects?
Understanding who prospects are is quite important because messages should be targeted to an audience of prospects. From the perspective of data mining, one of the challenges is using historical data when the prospect base changes.
Here are three typical reasons why care must be used when doing prospecting:
■■
Geographic expansion brings in prospects, who may or may not be similar to customers in the original areas.
■■
Changes to products, services, and pricing may bring in different target audiences.
■■
Competition may change the prospecting mix.
These are the types of situations that bring up the question: Will the past be a good predictor of the future? In most cases, the answer is “yes,” but the past has to be used intelligently.
The following story is an example of the care that needs to be taken. One company in the New York area had a large customer base in Manhattan and was looking to expand into the suburbs. They had done direct mail campaigns focused on Manhattan, and built a model set derived from responders to these campaigns. What is important for this story is that Manhattan has a high concentration of very expensive neighborhoods, so the model set was biased TEAMFLY
toward the wealthy. That is, both the responders and nonresponders were much wealthier than the average inhabitant of the New York area.
When the model was extended to areas outside Manhattan, what areas did the model choose? It chose a handful of the wealthiest neighborhoods in the surrounding areas, because these areas looked most like the historical responders in Manhattan. Although there were good prospects in these areas, the model missed many other pockets of potential customers. By the way, these other pockets were discovered through the use of control groups in the mailing—essentially a random sampling of names from surrounding areas.
Some areas in the control groups had quite high response rates; these were wealthy areas, but not as wealthy as the Manhattan neighborhoods used to build the model.
WA R N I N G Be careful when extending response models from one geographic area to another. The results may tell you more about similar geographies than about response.
When Is a Customer Acquired?
There is usually an underlying process in the acquisition of customers; the details of the process depend on the particular industry, but there are some general steps:
Team-Fly®
470643 c14.qxd 3/8/04 11:18 AM Page 463
Data Mining throughout the Customer Life Cycle 463
■■
Customers respond in some way and on some date. This is the “sale”
date.
■■
In an account-based relationship, the account is created. This is the
“account open date.”
■■
The account is used in some fashion.
Sometimes, all these things happen at the same time. However, there are invariably complications—bad credit card numbers, misspelled addresses, buyer’s remorse, and so on. The result is that there may be several dates that correspond to the acquisition date.
Assuming that all relevant dates are available, which is the best to use? That depends on the particular purpose. For instance, after a direct mail drop or an email drop, it might be interesting to see the response curve to know when responses are expected to come in, as shown in Figure 14.6. For this purpose, the sale date is most important date, because it indicates customer behavior and the question is about customer behavior. Whatever might cause the account open date to be delayed is not of interest.
A different question would have a different answer. For comparing the response of different groups, for instance, the account open date might be more important. Prospects who register a “sale” but whose account never opens should be excluded from such an analysis. This is also true in applications where the goal is forecasting the number of customers who are going to open accounts.
100%
90%
80%
70%
60%
50%
40%
tion Responded
30%
opor
20%
Pr
10%
0%
0
7
14
21
28
35
42
49
56
63
70
77
84
91
98 105 112
119
Days after First Response
Figure 14.6 These response curves for three direct mail campaigns show that 80 percent of the responses came within 5 to 6 weeks.
470643 c14.qxd 3/8/04 11:18 AM Page 464
464 Chapter 14
What Is the Role of Data Mining?
Available data limits the role that predictive modeling can play. Predictive modeling is used for channels such as direct mail and telemarketing, where the cost of contact is relatively high. The goal is to limit the contacts to prospects that are more likely respond and become good customers. Data available for such endeavors falls into three categories:
■■
Source of prospect
■■
Appended individual/household data
■■
Appended demographic data at a geographic level (typical census
block or census block group)
The purpose here is to discuss prospecting from the perspective of data mining. A good place to begin is with an outline of a typical acquisition strategy.