<<

. 92
( 137 .)



>>



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 applica­
tions where the goal is forecasting the number of customers who are going to
open accounts.


100%

90%

80%
Proportion Responded




70%
60%

50%

40%
30%

20%

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.
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 min­
ing. A good place to begin is with an outline of a typical acquisition strategy.
Companies that use direct mail or outbound telemarketing purchase lists.
Some lists are historically very good, so they would be used in their entirety.
For names from less expensive lists, one set of models is based on appended
demographics, when such demographics are available at the household level.
When such demographics are not available, neighborhood demographics are
used instead in a different set of models.
One of the challenges in direct marketing is the echo effect”prospects may
be reached by one channel but come in through another. For instance, a com­
pany might send a group of prospects an email message. Instead of respond­
ing to the email on the Web, some respondents might call a call center. Or
customers may receive an advertising message or direct mail, yet respond
through the Web site. Or an advertising campaign may encourage responses
through several different channels at the same time. Figure 14.7 shows an
example of the echo effect, as shown by the correlation between two channels,
inbound calls and direct mail. Another challenge is the funneling effect during
customer activation described in the next section.

WA R N I N G The echo effect may artificially under- or overestimate the
performance of channels, because customers inspired by one channel may be
attributed to another.




Customer Activation
Once a prospect has exhibited an interest, there is some sort of activation
process. This may be as simple as a customer filling out a registration form on
a Web site. Or, it might involve a more lengthy approval process, such as a
credit check. Or, it could be a bit more onerous, as in the example of life insur­
ance companies who often want to perform an underwriting exam that might
Data Mining throughout the Customer Life Cycle 465


include taking blood samples before setting rates. In general, activation is an
operational process, more focused on business needs than analytic needs.
As an operational process, customer activation may seem to have little to do
with data mining. There are two very important interactions, though. The first
is that activation provides a view of new customers at the point when they
join. This is a very important perspective on the customer, and, as a data
source, it needs to be preserved. Both the initial conditions and subsequent
changes are of interest.

T I P Customer activation provides the initial conditions of the customer
relationship. Such initial conditions are often useful predictors of long term

customer behavior.


Activation is also important because it narrows it further refines the cus­
tomer base. This is a funneling effect, as shown in Figure 14.8. This process is
for a newspaper subscription, a familiar process analogous to many similar
processes. It basically has the following steps:
The Sale. A prospect shows interest in getting a subscription, by providing
address and payment information, either on the Web, on a call, or on a
mail-in response card.
The Order. An account is created, which includes a preliminary verifica­
tion on the address and payment information.
The Subscription. The paper is actually physically delivered, requiring
further verification of the address and special delivery instructions.
The Paid Subscription. The customer pays for the paper.




Peaks and troughs often occur at about
the same time for these two channels.
Number of Starts
0

10

20

30

40

50

60

70

80

90

100

110

120

130

140

150

160

170

180

190

200

210

220




Week Number
Figure 14.7 Correlation between two channels over time suggests that one channel may
be leaking into another or something external is affecting both channels.
466 Chapter 14

Ne
w Ne
Sa w le
New sales come in le Sa Sa
le
ew
through many N
le le
Ne
channels. Sa Sa
w N
w w
Sa
Ne ew Ne
le le
Sa Sa
le
ew
N


Only sales with verifiable
O
addresses and credit r
rd Or Or de
er
er de de Or
d
Or
cards become orders. r r




Only orders with routable n
io
pt Su n
ri t io
bs
sc
addresses become ip
cr
ub cr
ip t
bs
S io n
subscriptions. Su




Only some subscriptions Paid
Subscription
are paid.


Figure 14.8 The customer activation process funnel eliminates responders at each step of
the activation process.


Each of these steps loses some customers, perhaps only a few percent per­
haps more. For instance, credit cards may be invalid, have improper expiration
dates, or not match the delivery address. The customer may live outside the
delivery region. The deliverers may not understand special delivery instruc­
tions. The address may be in an apartment building that does not allow access,
or the customer may simply not pay. Most of these are operational considera­
tions (the exception is whether or not the customer pays), and they illustrate the
kinds of operational concerns and processes involved with customer activation.
Data mining can play a role in understanding when customers are not mov­
ing through the process the way they should be”or what characteristics cause
a customer to fail during the activation stage. These results are best used to
improve the operational processes. They can also provide guidance during
acquisition, by highlighting strategies that are bringing in sales that are not
converted to paid subscriptions.
For Web-related businesses, customer activation is usually, although not
always, an automatic process that takes little time. When it works well, there is
no problem. Although it can take a short amount of time, it is a critical part of

<<

. 92
( 137 .)



>>