106 / Journal of Marketing, October 2004

Journal of Marketing
Vol. 68 (October 2004), 106–125
Rajkumar Venkatesan & V. Kumar
A Customer Lifetime Value
Framework for Customer Selection
and Resource Allocation Strategy
The authors evaluate the usefulness of customer lifetime value (CLV) as a metric for customer selection and mar-
keting resource allocation by developing a dynamic framework that enables managers to maintain or improve cus-
tomer relationships proactively through marketing contacts across various channels and to maximize CLV simulta-
neously. The authors show that marketing contacts across various channels influence CLV nonlinearly. Customers
who are selected on the basis of their lifetime value provide higher profits in future periods than do customers
selected on the basis of several other customer-based metrics. The analyses suggest that there is potential for
improved profits when managers design resource allocation rules that maximize CLV. Managers can use the
authors’ framework to allocate marketing resources efficiently across customers and channels of communication.
Rajkumar Venkatesan is Assistant Professor in Marketing (e-mail:
RVenkatesan@sba.uconn.edu), and V. Kumar is ING Chair Professor in
Marketing and Executive Director, ING Center for Financial Services (e-
mail: vk@sba.uconn.edu), School of Business, University of Connecticut.
This research study was supported partially by a grant from the Teradata
Center for Customer Relationship Management at Duke University, the
Marketing Science Institute, and the Institute for Study of Business Mar-
kets, and the authors owe special thanks to them. The study greatly ben-
efited from audience discussions at the National Center for Database
Marketing conference in Philadelphia and the Marketing Science Institute
conference at INSEAD, as well as at Michigan State University, Curtin
University, Tilburg University, State University of New York–Buffalo, and
the 2003 Marketing Science Conference. The authors thank Munshi Mah-
fuddin and Rajendra Ladda for research assistance. Special thanks are
due to Don Lehmann, Rick Staelin, Greg Allenby, and John Lynch for their
comments on the proposal version of this study. The authors thank a
business-to-business firm for providing the data for this study. They also
thank the anonymous JM reviewers for providing suggestions to enhance
the contribution of this study. They thank Renu and Andrew Peterson for
copyediting the manuscript.
Maximizing CLV
Some researchers have recommended CLV as a metric for
selecting customers and designing marketing programs
(Reinartz and Kumar 2003; Rust, Zeithaml, and Lemon
2004). However, there is no empirical evidence as to the
usefulness of CLV compared with that of other customer-
based metrics. Table 1 compares the CLV framework pro-
posed in this study with the existing literature on CLV and
database marketing. A comparison of the studies listed in
Table 1 shows that most of the previous studies provide
guidelines for calculating CLV and return on investment at
the aggregate level. Some recent studies (Reinartz and
Kumar 2003; Rust, Zeithaml, and Lemon 2004) provide
empirical evidence for the existence of a relationship
between marketing actions and CLV at the aggregate level.
However, as Berger and colleagues (2002) state, none of the
studies has proposed and tested a framework that provides
rules for resource allocation across various channels of
communication for each individual customer and across
customers.
On the basis of the comparisons in Table 1, we summa-
rize the significant contributions of our study as follows:
We provide a framework for measuring CLV that links the
influence of communications across various channels on
CLV. We also evaluate the usefulness of CLV as a metric for
customer selection and develop a framework for marketing
resource allocation that maximizes CLV. Given the assumed
link between CLV and firm profitability (Hogan et al.
2002), these are important issues.
In this study, we use customer data from a large
business-to-business (B2B) manufacturer to illustrate the
proposed framework empirically. The customer database of
the organization focuses on B2B customers. Our analyses
show that marketing communications across various chan-
nels influence CLV nonlinearly. The results from our analy-
ses suggest that customers selected on the basis of CLV pro-
vide higher profits than do customers selected on the basis
C
ustomer lifetime value (CLV) is rapidly gaining
acceptance as a metric to acquire, grow, and retain
the “right” customers in customer relationship man-
agement (CRM). However, many companies do not use
CLV measurements judiciously. Either they work with
undesirable customers to begin with, or they do not know
how to customize the customer’s experience to create the
highest value (Thompson 2001). The challenge that most
marketing managers currently face is to achieve conver-
gence between marketing actions (e.g., contacts across vari-
ous channels) and CRM. Specifically, they need to take all
the data they have collected about customers and integrate
them with how the firm interacts with its customers. In the
academic literature, Berger and colleagues (2002) support
the allocation of resources to maximize the value of the cus-
tomer base, and they strongly argue that such resource allo-
cation models are needed.
C
u
s
t
o
m
e
r

L
i
f
e
t
i
m
e

V
a
l
u
e

F
r
a
m
e
w
o
r
k

/

1
0
7
TABLE 1
Comparing the Proposed CLV Framework with Existing Models
Return on
Investment How Market Resource Resource Comparison
Modeled ing Communi- CLV-Based Allocation Allocation of Customer-
Type of Representa- and CLV cation Affects Resource for Each Across Based Statistical
Model tive Research Calculated Calculation CLV Allocation Customer Channels Metrics Details
CLV Berger and No Yes Yes No No No No Yes
Nasr (1998)
Berger et al. Yes Yes Yes Yes No Yes No No
(2002)
Customer Blattberg and Yes Yes No Yes No No No Yes
equity Deighton
(1996)
Libai, Yes Yes Yes No No No No No
Narayandas,
and Humby
(2002)
Database Reinartz and Yes Yes No No No No No Yes
marketing Kumar (2000)
Bolton, Lemon, Yes Yes Yes No No No No Yes
and Verhoef
(2004)
CLV Reinartz and Yes Yes Yes No No No Yes Yes
antecedents Kumar (2003)
Rust, Zeithaml, Yes Yes Yes No No No Yes Yes
and Lemon
(2004)
CLV-based Berger and Yes Yes Yes Yes No No No No
resource Bechwati
allocation (2001)
Present Study Yes Yes Yes Yes Yes Yes Yes Yes
108 / Journal of Marketing, October 2004
of other widely used CRM metrics. In addition, there is the
potential for substantial improvement in profits when man-
agers design resource allocation rules that maximize CLV.
In the next section, we develop the framework for the
measurement and maximization of CLV. We then propose
hypotheses about the influence of supplier-specific factors
and customer characteristics on the various CLV compo-
nents. In the subsequent section, we explain the models and
data we used to estimate CLV. We then discuss the results
from our analyses and explain the comparison of CLV with
other metrics for customer selection. Specifically, we com-
pare the aggregate profits provided by high-CLV customers
with those of customers who score high on several other
customer-based metrics. In the section “Resource Alloca-
tion Strategy,” we provide details on allocating resources
that maximize CLV. Our objective there is to evaluate the
extent to which CLV, and thus profits, can be increased by
allocating marketing resources across channels of contact
for each customer so as to maximize his or her respective
CLVs. Finally, we derive implications based on the results,
discuss the limitations of our study, and identify areas for
further research.
CLV Measurement and
Maximization
The various components of CLV include purchase fre-
quency, contribution margin, and marketing costs (however,
the various CLV components can vary depending on the
industry). Some of the antecedents of purchase frequency
and contribution margin (e.g., marketing communications)
are under management’s control and affect the variable
costs of managing customers. We use these antecedents to
maximize CLV.
Objective Function: CLV
Typically, CLV is a function of the predicted contribution
margin, the propensity for a customer to continue in a rela-
tionship (customer retention), and the marketing resources
allocated to the customer. In general, CLV can be calculated
as follows:
where
i = customer index,
t = time index,
n = forecast horizon, and
r = discount rate.
In contractual settings, managers are interested in pre-
dicting customer retention, or the likelihood of a customer
staying in or terminating a relationship. However, in non-
contractual settings, the focus is more on predicting future
customer activity because there is always a chance that the
customer will purchase in the future. Therefore, managers
who calculate CLV in noncontractual settings are interested
in predicting future customer activity and the predicted con-
( )
(
1 CLV =
Future contribution margin
i
it
− Future cost
it
)
( )
,
1
1
+


r
t
t
n
tribution margin from each customer. Previous researchers
have used the variable P(Alive), which represents the prob-
ability that a customer is alive (and thus exhibits purchase
activity) given his or her previous purchase behavior
(Reinartz and Kumar 2000), to predict future customer
activity in noncontractual settings. However, the measure
assumes that when a customer terminates a relationship, he
or she does not return to the supplier. This is also called the
“lost-for-good” scenario (Rust, Zeithaml, and Lemon
2004). If a customer is won back after termination, the com-
pany treats the customer as a new customer and ignores its
history with the customer.
Another method for predicting future customer activity
is to predict the frequency of a customer’s purchases given
his or her previous purchases. The assumption underlying
this framework is that customers are most likely to reduce
their frequency of purchase before terminating a relation-
ship. This assumption is in accordance with theories about
the different phases in a relationship and relationship life
cycles (Dwyer, Schurr, and Oh 1987; Jap 2001). In addition,
such a methodology enables a customer to return to the sup-
plier after a temporary dormancy in a relationship. Thus, in
this framework, we measure CLV by predicting the pur-
chase pattern (purchase frequency or interpurchase times)
over a reasonable period. This is also called the “always-a-
share” scenario. The lost-for-good approach is questionable
because it systematically understates CLV (Rust, Zeithaml,
and Lemon 2004). Thus, we use the always-a-share
approach in this study. Given predictions of contribution
margin, purchase frequency, and variable costs, the CLV
function we use can be represented as follows:
where
CLV
i
= lifetime value of customer i;
CM
i,y
= predicted contribution margin from cus-
tomer i (computed from a contribution
margin model) in purchase occasion y,
measured in dollars;
r = discount rate for money (set at 15%
annual rate in our study);
c
i,m,l
= unit marketing cost for customer i in chan-
nel m in year l (the formulation of CLV
does not change if l is used to represent
periods other than one year);
x
i,m,l
= number of contacts to customer i in chan-
nel m in year l;
frequency
i
= predicted purchase frequency for
customer i;
n = number of years to forecast; and
T
i
= predicted number of purchases made by
customer i until the end of the planning
period.
In addition to accurate measurement of CLV for each
customer, our objective is to allocate resources so as to
maximize CLV. Thus, we model the purchase frequency and
( )
( )
2
1
1
CLV
r
i
y
T
i
CM
i,y
y
frequency
i

+


c x
m
i,m, i,m,
1
l
l l
l




×
+
1
1
n
r ( )
,
Customer Lifetime Value Framework / 109
1
In this study, we do not use a budget constraint on the total
resources available for contacting customers. Therefore, we are
interested only in allocating resources across channels for each
individual customer (i.e., within each customer across channels).
However, our framework can be applied to allocate resources
across channels with each individual customer and across cus-
tomers in the presence of a budget constraint.
contribution margin of customers as a function of marketing
resource variables such as channel contact. We then use the
customer responsiveness to marketing actions, obtained
from the purchase frequency and contribution margin mod-
els, to develop resource allocation strategies that maximize
CLV. In summary, the objective is to identify the resource
allocation rules across various channels of communication
for each individual customer such that the respective CLVs
(as provided in Equation 2) are maximized.
1
Our objective
function is subject to the following constraints: frequency >
0 ∀ i, t, and x
i,m,l
≥ 0 ∀ i, m, l.
Discounting contribution margin. We first focus on the
discounting of contribution margin over a period of time.
Assume that it is currently year l = 1 and that we need to
forecast the contribution margin from each customer for the
next n years (i.e., until l + n). It is possible that a customer
makes several purchases in a given year. Berger and Nasr
(1998, Equation 2) and Rust, Zeithaml, and Lemon (2004)
provide guidelines for discounting contribution margin
from customers when there is more than one purchase occa-
sion (y) per year. In this approach, the discount rate from a
customer is scaled according to his or her frequency of pur-
chase (as is shown in Equation 2). For example, consider
when the planning horizon is one year and the frequency of
purchases is two times (frequency = 2). The first purchase
occasion (y = 1) occurs after 6 months; therefore,
y/frequency = .5 (in other words, we use the square root of
the discount rate). The second purchase occasion (y = 2)
occurs after 12 months; therefore, y/frequency = 1.
Discounting cost allocations. The discounting of cost
allocations is straightforward if we assume that there is a
yearly allocation of resources (as is the case in most organi-
zations) and that the cost allocation occurs at the beginning
of the year (the present period). Thus, the cost allocation in
the first year need not be discounted, the cost allocation in
the second year needs to be discounted for one year, and so
on. Thus, we raise the denominator in the cost function cal-
culation to current year – 1 (i.e., l – 1).
Discussion of model constraints. The constraints ensure
the nonnegativity of the predicted purchase frequency and
communication levels for each customer i during period l.
Antecedents of Purchase
Frequency and Contribution Margin
Purchase Frequency
An objective of relationship marketing is to ensure future
purchase activity. Purchase frequency is also a component
of our CLV calculation. Therefore, as the basis for selecting
antecedents to predict purchase frequency, we use the
commitment–trust theory of relationship marketing (Mor-
gan and Hunt 1994) as well as previous research in cus-
tomer equity and CLV (Bolton, Lemon, and Verhoef 2004;
Bowman and Narayandas 2001; Reinartz and Kumar 2003;
Rust, Zeithaml, and Lemon 2004) and channel communica-
tions (Grewal, Corner, and Mehta 2001; Mohr and Nevin
1990; Morgan and Hunt 1994; Rindfleisch and Heide
1997).
The overall theoretical framework that we used is pro-
vided in Figure 1. We summarize the antecedents of pur-
chase frequency, their operationalization, expected effects,
and the rationale for our expectations in Table 2. Next, we
provide a detailed discussion for a few of the hypotheses
that are unique to our study.
Supplier-Specific Factors: Channel
Communications
In this study, we classify channels of communication into
the following contact modes: rich (e.g., face-to-face, trading
event meetings), standardized (e.g., direct mail, telephone),
and Web based (Mohr and Nevin 1990). Although we
expect the relationships between different channels of com-
munication and predicted customer activity to be similar,
we need to analyze customer responses separately across
different channels because the costs of serving customers
across different channels are different, and customers might
exhibit different responsiveness across the various channels.
The costs of communication in each channel can influence
managers’ frequency of communication in each channel.
Frequency of rich and standardized modes. Face-to-face
communications and trading event meetings are the richest
and most direct mode of communication possible among
channel members (Mohr and Nevin 1990). Relational cus-
tomers tend to have high commitment and trust with their
suppliers, which results in less uncertainty, more coopera-
tion, and less complexity in their relationships than in those
of transactional customers (Morgan and Hunt 1994). Rich
modes of communication are preferred to standardized
modes when issues in the channel structure are complex
and when there is a high degree of uncertainty in the rela-
tionship. Rich modes of communication are also effective in
converting transactional customers to relational ones (Gane-
san 1994).
Direct mail and telephone communication are the most
standardized and cost-effective modes of individual-level
communication available to an organization. Standardized
modes are also the most cost-effective method for identify-
ing customers who are interested in an organization’s cur-
rent promotion (Shepard 2001). For transactional cus-
tomers, direct mail can be used in combination with
telephone sales to generate interest in products while simul-
taneously improving the return on investment (Nash 1993).
For relational customers, direct mail serves to maintain
commitment and trust by communicating relationship bene-
fits (Morgan and Hunt 1994) and to inform the best cus-
tomers about new product offerings. Therefore, although
the purpose of standardized communication may be differ-
ent for transactional customers than for relational ones, we
expect that the marginal response for increased frequency is
the same across segments.
110 / Journal of Marketing, October 2004
FIGURE 1
A Conceptual Framework for Measuring and Using CLV
Customer Characteristics:
Switching Costs
•Upgrading
•Cross-buying
Customer Characteristics
•Lagged contribution margin
•Establishment size
•Industry category
•Total quantity of purchases
Predicted
Purchase
Frequency
Contribution
Margin
Total
Profit
Net Present Value
(Future Profits)
= CLV
Marketing
Costs
Discount
Rate
Allocate Appropriate
Resources
Generate
allocation rules
Supplier-Specific Factors:
Channel Communication
•Level of rich modes
•Level of standardized modes
•Intercontact time
Supplier-Specific Factors
•Total marketing communication
Customer Characteristics:
Involvement
•Bidirectional communication
•Number of returns
•Number of Web-based contacts
Customer Characteristics: Previous
Behavior
•Product category purchased
However, it has been proposed that too much communi-
cation causes a relationship to be dysfunctional (Fournier,
Dobscha, and Mick 1997). In addition, the marginal
response to a higher level of rich modes of communication
need not always be higher; sometimes, it even can be nega-
tive. Although the utility of marketing contacts is not ques-
tioned, too much contact can overload buyers and have dys-
functional consequences (e.g., ubiquitous junk mail). Thus:
H
1
: An inverted U-shaped relationship exists between the fre-
quency of rich and standardized modes of communication
and a customer’s predicted purchase frequency.
Intercontact time. Following the theory that leads to H
1
,
we expect that there exists an optimal level of intercontact
time between suppliers and buyers. Higher levels of previ-
ous communications lead to trust with the supplier and act
as glue that holds together a communication channel (Mor-
gan and Hunt 1994). However, too much communication
may be dysfunctional. In addition, the marginal utility of an
additional piece of information from a supplier firm in a
short period is low. To maximize the effect of each contact,
supplier firms need to pace their communication schedule
to suit customer needs. Thus:
H
2
: An inverted U-shaped relationship exists between inter-
contact time and a customer’s predicted purchase
frequency.
Customer Characteristics: Customer Involvement
Bidirectional communication. Research on channel
communications shows that highly relational channel struc-
tures are associated with large bidirectional communica-
tions among channel members (Mohr and Nevin 1990).
Although customer-initiated contacts are associated primar-
ily with complaints in business-to-consumer settings, the
same is not necessarily true for B2B settings. In a B2B set-
ting, customers can initiate contacts with suppliers for sev-
eral reasons, such as if they have new needs that the sup-
plier may be able to fulfill, if they want the supplier to
conduct training programs at the customer’s site (Cannon
and Homburg 2001), or if the supplier invites the customer
to participate in new product development sessions. On
most occasions, bidirectional communication in channels
strengthens a relationship, indicates customer involvement,
Customer Lifetime Value Framework / 111
Variable
Upgrading
Cross-buying
Bidirectional
communica-
tion
Returns
Frequency of
Web-based
contacts
Relationship
benefits
Frequency of
rich modes
of communi-
cation
Frequency of
standardized
modes of
communica-
tion
Intercontact
time
Operationalization
Number of product purchase upgrades
until an observed purchase
Number of different product categories
a customer has purchased
Ratio of number of customer-initiated
contacts to total number of customer
contacts (customer initiated and
supplier initiated) between two observed
purchases
Number of products the customer
returns between two observed
purchases
Number of times in a month the
customer contacts the supplier through
the Internet between two observed
purchases
Indicator variable of whether a customer
is a premium service member (based
on revenue contribution in the previous
year)
Number of customer contacts by the
supplier in a month (through sales
personnel) between two observed
purchases
Number of customer contacts by the
supplier in a month (through telephone
or direct mail) between two observed
purchases
Average time between two customer
contacts by the supplier across all
channels of communication between
two observed purchases
Expected
Effect
+
+
+
ʝ
+
+
ʝ
ʝ
ʝ
Rationale
Customers who upgrade have higher
switching costs with each upgrade,
which can lead to lower propensity to
leave and higher recurrent needs
(Bolton, Lemon, and Verhoef 2004).
Customers who purchase across
several product categories have higher
switching costs and recurrent needs
(Bowman and Narayandas 2001;
Reinartz and Kumar 2003).
Two-way communication between
parties strengthens the relationship and
ensures that the focal firm is recalled
when a need arises (Morgan and Hunt
1994).
Returns provide an opportunity for firms
to satisfy customers and ensure repeat
purchases (Reinartz and Kumar 2003),
but too many purchases can be
detrimental to the relationship and can
indicate that the firm has not used the
return opportunities appropriately.
Customers who use online
communication want transaction
efficiencies, and customers who want to
create efficiencies are highly relational
and have recurring needs (Grewal,
Corner, and Mehta 2001; Rindfleisch
and Heide 1997).
Acknowledgment of customers with
relationship benefits reduces the
propensity of customers to quit and
increases the probability that the focal
firm is recalled when a need arises
(Morgan and Hunt 1994).
Timely communication between parties
reduces the propensity of a customer to
quit a relationship (Mohr and Nevin
1990; Morgan and Hunt 1994), but too
much communication can be detrimental
to the relationship (Fournier, Dobscha,
and Mick 1997; Nash 1993); thus, there
is an optimal communication level.
A long time between contacts can lead
to forgetfulness, but contacts that are
too soon can cause dysfunction.
TABLE 2
Antecedents, Covariates, and Expected Effects
Purchase Frequency Model
Antecedents
112 / Journal of Marketing, October 2004
Variable Operationalization
Expected
Effect Rationale
TABLE 2
Continued
Product
category
Lagged
contribution
margin
Total marketing
efforts
Total quantity
purchased
Size of firm
Industry
category
Two indicator variables: one indicates a
hardware purchase; the other indicates
a software purchase
Customer’s contribution margin from the
previous year
Total number of customer contacts
across all channels
Total quantity of products the customer
purchased across all product categories
Number of employees in the customer
firm
Standard industrial classification–based
industry category to which the customer
firm belongs
+
+
+
+
A customer’s purchase patterns may
depend on the product category
purchased.
Previous revenue is a good predictor of
current revenue and accounts for any
model misspecification (Niraj, Gupta,
and Narasimhan 2001).
Previous marketing communications
and depth (quantity) of purchases
positively affect contribution margin
(Gupta 1988; Tellis and Zufryden 1995).
Control variables that accommodate for
customer heterogeneity (Niraj, Gupta,
and Narasimhan 2001).
Contribution Margin Model
Antecedents
Covariate
Covariates
Notes: For the purchase frequency model, the dependent variable is purchase frequency; for the contribution margin model, the dependent
variable is contribution margin.
and increases interdependence among channel members
(Ganesan 1994; Mohr and Nevin 1990). Thus:
H
3
: The higher the level of bidirectional communication, the
higher is a customer’s predicted purchase frequency.
Frequency of Web-based contacts. In this study, we ana-
lyze Web-based contacts separately from other channels of
communication because Web-based communication is cus-
tomer initiated (i.e., a passive mode of operation for the
supplier). However, there are several advantages to tracking
Web-based contacts in a B2B setting. First, Web-based
communication between buyers and suppliers is the most
cost-effective method of communication. Second, Web-
based contacts from the buyers provide some important sig-
nals to the supplier about the buyer’s relationship orienta-
tion. Grewal, Corner, and Mehta (2001) find that
organizations enter and actively participate in electronic
markets if their motivation is to improve efficiency in trans-
actions. In addition, participation in electronic markets (or
use of Web-based initiatives) improves transaction effec-
tiveness and efficiency (Rindfleisch and Heide 1997). Effi-
ciency of communication and transactions among channel
members is associated with a relational structure and higher
customer involvement (Mohr and Nevin 1990; Sheth and
Parvatiyar 1995). Thus:
H
4
: The higher the number of Web-based contacts from a cus-
tomer, the higher is the customer’s predicted frequency of
purchase.
Contribution Margin
The antecedents that we adopt to predict contribution mar-
gin are based on findings from previous research on
antecedents of customer revenue (Niraj, Gupta, and
Narasimhan 2001) and purchase quantity (Gupta 1988; Tel-
lis and Zufryden 1995). As with purchase frequency, we
classify the antecedents of contribution margin as supplier-
specific factors (total marketing efforts) and customer char-
acteristics (lagged contribution margin and purchase quan-
tity). We use size of an establishment and industry category
as covariates in our model. In Table 2, we provide a descrip-
tion of the antecedents that we propose influence customer
purchase frequency and contribution margin. In the data-
base, we also provide the operationalization of the
antecedents, their expected effects, and the rationale for our
expectations based on previous research. Because all the
antecedents we use in our contribution margin model are
based on previous research and findings, we do not discuss
the hypotheses in detail. In Table 2, we provide a descrip-
tion of the covariates we use as control variables in the pur-
chase frequency and contribution margin models.
Modeling CLV and Data
Model Development
To predict CLV, we need a stochastic model to predict each
customer’s purchase frequency and a panel-data model that
Customer Lifetime Value Framework / 113
predicts contribution margin. In this study, we assume that
the amount a customer spends is independent of purchase
timing. This is a rather restrictive assumption for frequently
bought consumer goods (Tellis and Zufryden 1995). How-
ever, in our product category, we find that the correlation
between purchase frequency and contribution margin is not
significant.
Purchase Frequency
We model a customer’s purchase frequency using the gener-
alized gamma model of interpurchase timing that Allenby,
Leone, and Jen (1999) developed. The generalized gamma
model also accommodates the commonly used exponential
distribution for interpurchase times (Reinartz and Kumar
2003). The likelihood function for the purchase frequency
model is given as follows:
where
f(t
ij
|α, λ
i
, γ) = the density function for the generalized
gamma distribution (i.e., the probability
of the jth purchase for customer i occur-
ring at period t, given α, λ
i
, γ);
S(t
ij
|α, λ
i
, γ) = the survival function for the generalized
gamma distribution (i.e., the probability
of the jth purchase for customer i occur-
ring at a given period is greater than t,
given that the jth purchase has not
occurred until time t, given α, λ
i
, γ);
c
ij
= the censoring indicator, where c
ij
= 1 if
the jth interpurchase time for the ith cus-
tomer is not right-censored, and c
ij
= 0 if
the jth interpurchase time for the ith cus-
tomer is right-censored;
Φ
ijk
= the probability of observation j for the ith
customer belonging to subgroup k; and
α, λ
i
, γ = the parameters of the generalized gamma
distribution.
Because we use a generalized gamma distribution to model
interpurchase time and the likelihood function in Equation
3, the expected time until next purchase is given as follows:
The ratio of 12 (because we use months as the unit of
analysis) to the expected time until next purchase (which
we obtain by modeling a generalized gamma distribution on
the interpurchase times, as is shown in the work of Allenby,
Leone, and Jen [1999]) gives the predicted purchase fre-
quency. The parameters α and γ establish the shape of the
interpurchase time distribution, and λ
i
is the individual-
specific purchase rate parameter. We assume that the popu-
lation consists of k subgroups, and Φ
ik
provides the mass
point (i.e., weight) for each subgroup. We model the proba-
( ) 4
1
Φ Γ
Γ
ik
k
k
k k

× +
|
(
'
`
J
J

]
]
]
( )
|
¦
|
|
|
|
α
γ α
||
|
|
|
× λ
ik
.
( ) | , , 3
1 1
L f t
ijk
k
K
k ij k ik k
c
j
ij

( )


Φ α λ γ
JJ
i
n
k ij k ik k
c
i
ij
S t
∏ ∏


1
1
( | , , ) ,
( )
α λ γ
2
We also used lagged interpurchase time instead of the log of
the lagged interpurchase time, and we did not find any difference
in the substantive conclusions. We used log of the lagged interpur-
chase times because lagged interpurchase times can have a thresh-
old effect on the influence of current interpurchase times (Allenby,
Leone, and Jen 1999). The log of the lagged interpurchase time
achieved this objective in scaling the tail of the lagged interpur-
chase time distribution.
bility of a customer belonging to each subgroup Φ
ik
as a
probit function of the antecedents and covariates of pur-
chase frequency. Specifically, we represent the link function
as Φ
ik
= f(x
ij
β
i
), where x
ij
represents the antecedents and
covariates of purchase frequency for customer i in purchase
occasion j, and β
i
represents the customer-specific response
coefficients.
Our model framework, as presented in Equation 3,
resembles a hierarchical Bayes formulation of the concomi-
tant continuous mixture model. To address the issue of
endogeneity, we use the one-period lagged value for all the
antecedents and covariates in our analysis (Villas-Boas and
Winer 1999). However, to account for any extraneous fac-
tors, we also use the log of the lagged interpurchase.
2
The
specification of the model enables us to estimate individual
customer-level coefficients for the influence of the covari-
ates on the probability of a customer belonging to a particu-
lar subgroup and thus interpurchase times.
Contribution Margin
We model the contribution margin from a customer using
panel-data regression methodologies. We needed to address
endogeneity issues while using lagged contribution margin
as an independent variable in our model. In panel-data mod-
els with lagged dependent variables, the endogeneity in for-
mulation can be alleviated with a one-period difference in
the dependent variable and a two-period lagged dependent
variable as an independent variable (Baltagi 1998). We use
the growth in contribution margin from period t – 1 to t as
the dependent variable and the contribution margin in
period t – 2 as an independent variable. The other indepen-
dent variables we used are specific to period t – 1 (this also
accommodates the issue of endogeneity). The independent
variables in the contribution margin model are thus lagged
contribution margin, lagged total quantity purchased,
lagged firm size, industry category, and lagged total market-
ing efforts. Thus, the contribution margin model is
(5) ∆CM
i,t
= β
0
+ β
1
CM
i,t – 2
+ β
2
Quantity
i,t – 1
+ β
3
Size
i,t – 1
+ Σ
j
β
j
Industry
j
+ β
4
Totmark
i,t – 1
+ e
i,t
,
where
∆CM
i,t
= difference in contribution margin from
period t – 1 to period t for customer i,
measured in dollars;
Size
i,t – 1
= firm size for customer i in period t – 1,
measured as number of employees;
Industry = indicator variable for industry category
of the customer firm;
Totmark
i,t – 1
= total number of contacts made to cus-
tomer i in period t – 1;
114 / Journal of Marketing, October 2004
Quantity
i,t – 1
= total quantity of products bought by
customer i in period t – 1;
e
i,t
= error term;
i = index for the customer; and
t = index for time.
Data
We used data from a large multinational computer hardware
(servers, workstations, and personal computers) and soft-
ware (integration and application) manufacturer for the
empirical application of our framework. The company’s
database focuses on business customers. The product cate-
gories in the database represent different areas of high-
technology products. In addition, for these product cate-
gories, it is the choice of the buyer and seller to develop
their relationships, and there are significant benefits for
both parties to maintain a long-standing relationship. The
choice of vendors for the products is normally made after
much deliberation by the buyer firm. Even though the firm’s
products are durable goods, they require constant mainte-
nance and frequent upgrades, which provides the variance
required in modeling customer response. For our analyses,
we used two cohorts of customers: Cohorts 1 and 2. We
assigned customers to Cohort 1 (Cohort 2) if their first pur-
chase with the manufacturer occurred in the first quarter of
1997 (first quarter of 1998). In our samples, we removed
customers who had missing values for either rich or stan-
dardized modes of communication. We also restricted our
sample to customers who had made at least five purchases.
Overall, we removed 20% of the original cohort of cus-
tomers for our analyses, which resulted in an effective sam-
ple size of 1316 and 873 observations for customers in
Cohorts 1 and 2, respectively. The interpurchase time for
customers in Cohort 1 ranges from 1.5 to 23 months; for
Cohort 2, it ranges from 1 month to 20 months.
Purchase frequency model. We used each observed pur-
chase for a customer as an observation in the purchase fre-
quency model. For Cohort 1, we selected customers who
made their first purchase in the first quarter of 1997. For
each customer, we omitted the first observed purchase in
our analysis sample because the first observed purchase is
restricted to be within three months for all customers in the
cohort, and theoretically the customer retention phase
begins after the first purchase. The antecedents and covari-
ates we used can be classified as cumulative and current
effects. The cumulative effects antecedents include cross-
buying and upgrading, and their values represent the total
number of different products (for cross-buying) or upgrades
that the customer has purchased since the first purchase
until the current observed purchase.
The current-effects antecedents include bidirectional
communication; returns; relationship benefits; frequency of
rich, standardized, and Web-based contacts; and intercon-
tact time. The covariates in the purchase frequency model
(type of product purchased) can be classified as current
effects. We calculated the current-effects antecedents and
covariates on the basis of the activities of the customer or
the supplier (in the case of channel communications)
between the previous observed purchase (j – 1) and the cur-
rent observed purchase (j). To assess the inverted U-shaped
relationships, we used a quadratic conversion (including the
square covariate term in Equation 3) of the respective
antecedent. For all customers, we used the interpurchase
times until the end of 2000 as our calibration sample. We
used the 2001 data as a holdout sample and to compare
strategies. All the antecedents and covariates we used in our
analyses are lagged variables. Specifically, for observed
purchase j, the cumulative-effects antecedents represent the
customer’s activity since relationship initiation until
observed purchase j – 1. Similarly, for observed purchase j,
the current effects antecedents and covariates represent the
customer’s (or supplier’s) activity between observed pur-
chases j – 2 and j – 1.
Contribution margin model. To model contribution mar-
gin from a customer, we used the annual sales from various
purchases of each customer. For customers in Cohort 1, we
used the annual sales from each customer from 1997 to
2000. Given our model structure in Equation 5, there are
two observations per customer in our analysis sample.
Specifically, for each customer, for Observation 1 the
dependent variable is the difference in contribution margin
between 2000 and 1999, and the independent variables
include the contribution in 1998, the firm size in 1999, the
industry category of the customer, the total number of con-
tacts made to the customer in 1999, and the total quantity of
products purchased in 1999. Similarly, for Observation 2,
the dependent variable is the difference in contribution mar-
gin between 1999 and 1998, and the independent variables
include the contribution margin in 1997. As we stated previ-
ously, we used the 2001 data as a holdout sample and to
compare strategies. The descriptive statistics for the data
and the correlation matrix of the antecedents are provided in
Table 3.
Results from Estimation of CLV
Purchase Frequency Model
As we discussed previously, we used an effective sample
size of 1316 and 873 observations that belong to Cohort 1
(first purchase in 1997) and Cohort 2 (first purchase in
1998), respectively, for our analyses. We discuss the results
from Cohort 1 in detail in the text. The results from Cohort
2 are quite similar to those of Cohort 1 and are provided
along with the results for Cohort 1 in the corresponding
tables. We censored our data set in 2000 and used the 2001
data as our validation (or holdout) sample in both cohorts.
We estimated the purchase frequency model in Equation 3
using Markov chain Monte Carlo (MCMC) methods. The
results from the purchase frequency model for Cohorts 1
and 2 are provided in Table 4 (for details on the model esti-
mation, model selection, model performance, and support
for hypotheses, see Appendix A). The coefficients of the
antecedents reported in Table 4 are the means from the pos-
terior samples of β
i
, and the signs for the coefficients repre-
sent their influence on a customer’s purchase frequency.
There are several insights that we derive from Table 4,
which we discuss subsequently.
Model fit. The results show that the generalized gamma
model with two subgroups provides a good fit to the data
C
u
s
t
o
m
e
r

L
i
f
e
t
i
m
e

V
a
l
u
e

F
r
a
m
e
w
o
r
k

/

1
1
5
Mean
5
(4.5)
1.32
(1.16)
2.58
(3.13)
.84
(.62)
.91
(.86)
3.88
(4.37)
.09
(.12)
1.79
(1.52)
20.74
(22.81)
15.3
(16.1)
Standard
Deviation
8.4
(6.8)
.89
(.91)
1.7
(1.5)
2.41
(3.52)
3.7
(2.81)
25.81
(24.94)
.29
(.25)
5.69
(5.84)
47.75
(45.81)
13.8
(14.2)
Purchase
Frequency
1
.62***
.53***
.68***
.36***
.40***
.41***
.45***
.44***
.51***
Upgrading
1
.39*
.51*
.09
.21
.31**
.15**
.22*
–.08
Cross-
Buying
1
.48*
.13
.15
.36**
.29**
.34*
.06
Bidirec-
tional
Communi-
cation
1
.11*
.17**
.22
.32*
.24**
–.01
Returns
1
.06
–.21*
.05
.07
–.01
Frequency
of Web-
Based
Contacts
1
.25
.34*
.32*
–.04
Relation-
ship
Benefit–
Premium
Service
Level
1
.07
.19*
–.03
Frequency
of Rich
Modes
1
.30**
–.11
Frequency
of Stan-
dardized
Modes
1
.06
Intercon-
tact Time
(Days)
1
TABLE 3
Descriptive Statistics and Correlation Matrix
Variable
Purchase
frequency
Upgrading
Cross-buying
Bidirectional
communi-
cation
Returns
Frequency of
Web-
based
contacts
Relationship
benefit–
premium
service
level
Frequency of
rich modes
Frequency of
standard-
ized
modes
Intercontact
time
(days)
Purchase Frequency Model
1
1
6
/

J
o
u
r
n
a
l

o
f

M
a
r
k
e
t
i
n
g
,
O
c
t
o
b
e
r

2
0
0
4
TABLE 3
Continued
*Significant at α = 10%.
**Significant at α = 5%.
***Significant at α < 1%.
Notes: The frequency of rich, standardized, and Web-based contacts represents the frequency of the respective mode of communication between two consecutive purchases. Values in paren-
theses for the descriptive statistics represent Cohort 2. The correlation matrix for Cohort 2 is similar to that of Cohort 1 and is available on request from the authors. We do not report the
correlation matrix of the covariates because it does not have any substantive interpretation.
Growth in Lagged Total Total
Variable Mean Standard Deviation Contribution Margin Contribution Margin Marketing Efforts QuantityPurchase
Contribution Margin Model
Growth in contribution margin 4955 417,270 1.00***
(4827) (381,297)
Lagged contribution margin 31,143 323,139 –.78*** 1.00**
(32,825) (318,867)
Total marketing efforts 3.49 17.26 .61*** .08** 1.00
(4.5) (16.24) .00
Total quantity purchased 1.89 13.40 .71*** .05** .07 1
(2.01) (14.02)
Customer Lifetime Value Framework / 117
TABLE 4
Coefficients for the Generalized Gamma Purchase Frequency Models
Model 2 Model 3
Model 1 (Generalized Gamma (Generalized Gamma
(Generalized Gamma with Mixture, Without with Mixture and
Variable Without Mixture) Temporal Variation) Temporal Variation)
Component 1
α
1
3.05 (4.12)** 4.1 (4.2)** 3.27 (3.59)**
υ
1
26.18 (25.08)** 42.97 (43.82)** 47.05 (48.90)**
θ
1
.007 (.002)** .01 (.08)** .04 (.02)**
γ
1
1.2 1.2 (1.3) 1.2 (1.5)
Mass point .54 (.55) .54 (.56)
Component 2
α
2
.58 (1.57)** 1.48 (3.61)**
υ
2
38.21 (34.28)** 32.07 (49.02)**
θ
2
.008 (.012)** .005 (.001)**
γ
2
.9 (.9) .9 (.9)
Mass point .46 (.45) .46 (.44)
Coefficients
β
01
.89 (–1.52)** –2.05 (–2.51)*
Lagged log interpurchase
time –2.09 (–2.85)**
Antecedents: Customer Characteristics
Upgrading 5.01 (5.62)**
Cross-buying 6.08 (6.92)**
Bidirectional communication 1.49 (1.01)**
Returns 10.52 (9.98)**
(Returns)
2
–3.84 (–4.01)*
Frequency of Web-based
contacts 3.52 (2.38)**
Antecedents: Supplier-Specific Factors
Relationship benefits 9.78 (7.65)**
Frequency of rich modes of
communication 4.50 (5.65)**
(Frequency of rich modes of
communication)
2
–1.30 (–1.28)**
Frequency of standardized
modes of communication 6.53 (7.02)**
(Frequency of standardized
modes of communication)
2
–.28 (–.53)**
Intercontact time 9.64 (8.56)**
(Intercontact time)
2
–3.21 (–4.05)**
Log-likelihood –3298.91 (–3627.08) –2908.66 (–3297.04) –2237.82 (–2437.91)
AIC –6606 (–7262) –5825 (–6602) –4484 (–4884)
BIC –6639 (–7295) –5910 (–6687) –4683 (–5083)
Relative absolute error .95 (.91) .84 (.80) .51 (.53)
*Posterior sample values between the 2.5th and 97.5th percentile do not contain zero.
**Posterior sample values between the .5th percentile and 99.5th percentile do not contain zero.
Notes: Values in parentheses represent Cohort 2. The product category variable was not significant in our analysis, and thus we do not include
it here. Relative absolute error is with respect to a moving average model. The significance levels apply to the coefficients of Cohorts 1
and 2.
and is better than other models for modeling purchase fre-
quency (log-likelihood for Model 3 = –2237.82, Akaike
information criterion [AIC] = –4484, and Bayesian infor-
mation criterion [BIC] = –4683). We also used a hazard
model with the finite mixture framework (Kamakura and
Russell 1989) to model purchase frequency, and we found
that our proposed model fits the data better and has better
predictive capabilities.
3
The parameter estimates of the purchase frequency model are
based on mean values from 50 repeats with random starting values
for each repetition. We adopted such a procedure to ensure that the
parameter estimates are global optimal values and are not affected
by any local maxima.
Distribution parameters.
3
The mean expected purchase
frequency for Subgroup 1 is 4.2 purchases in a year, and the
118 / Journal of Marketing, October 2004
4
We also estimated the contribution margin model at the
monthly, quarterly, and semiannual levels, but we did not find any
mean expected purchase frequency for Subgroup 2 is 1.01
purchases in a year. Given the variation in expected pur-
chase frequencies in each subgroup, we term Subgroup 1
the “active state” and Subgroup 2 the “inactive state.” The
component masses for Subgroup 1 (ϕ
1
) and Subgroup 2

2
) are .54 and .46, respectively. This implies that we
expect 54% of the customers to be active in the prediction
window and 46% of the customers to be inactive in the pre-
diction window.
Supplier-specific factors. Our analyses indicate that a
supplier’s contact strategy and provision of relationship
benefits significantly affect a customer’s predicted purchase
frequency. We find that the frequency of contacts affects
purchase frequency nonlinearly. Specifically, we find an
inverted U-shaped relationship. This leads us to believe that
there is an optimal level of marketing communication for
each customer. A firm’s increasing communication beyond
a certain threshold may result in diminishing returns in
terms of customer purchase frequency. This finding also
provides the reasoning to determine the optimal level of
resources that needs to be allocated across channels to max-
imize CLV in Phase 2.
The coefficients of the marketing contacts reveal a dif-
ference in the influence of various channels on customer
purchase frequency. The coefficient of the quadratic term
for rich modes of communication (–1.30 for Cohort 1) is
higher than the coefficient of the quadratic term for stan-
dardized modes of communication (–.28 for Cohort 1).
Thus, we can infer that the rate of diminishing returns (after
exceeding the threshold) is much higher for the rich mode
of communications than for standardized modes. Therefore,
although the rich mode of communication is interactive and
effective, firms should use it with great caution.
Customer characteristics. The results indicate that
upgrading and cross-buying positively influence a cus-
tomer’s purchase frequency. This is in line with the findings
of Reinartz and Kumar (2003), who also find that breadth of
purchase positively affects a customer’s duration in a prof-
itable relationship. We also find that the higher the bidirec-
tional communication between the customer and supplier,
the higher is the customer’s purchase frequency. Thus, in
addition to timely communication from the supplier to the
customer (Morgan and Hunt 1994), communication from a
customer can be a good indicator of a customer’s activity.
With respect to returns from a customer, our analysis
suggests that managers need to exercise caution. We find
support for an inverted U-shaped relationship between pur-
chase frequency and returns. This indicates that customers
who return products within a threshold are a good asset to
the firm. The results highlight the importance of firms’ rec-
ognizing the customers who establish contact through the
online channel in their CRM strategies.
Contribution Margin Model
We estimated the contribution margin model with annual
data from 1997 (t – 4) to 2000 (t) for Cohort 1 and from
1998 to 2000 for Cohort 2.
4
The revenue in 2001 (t + 1) acts
significant differences in model performance. Thus, to maintain
simplicity, we used the contribution margin model with the annual
data.
as a holdout sample for Cohorts 1 and 2. The coefficients of
the contribution margin model are provided in Table 5. The
main insight from Table 5 is that the contribution margin
model provides an adjusted R
2
of .68 and thus can explain
significant variation in contribution margin from customers.
We derive several other insights from Table 5, which we
discuss subsequently.
Supplier-specific factors. Total lagged marketing efforts
contribute significantly to an explanation of variation in
current contribution margin, which implies that a supplier’s
contact strategy affects both purchase frequency and contri-
bution margin.
Customer characteristics. The two-period lagged contri-
bution margin provides the highest contribution to an expla-
nation of current-period contribution margin. In addition,
lagged quantity of goods is significant in explaining varia-
tion in contribution margin. Firm size and industry category
explain variation in current-period contribution margin.
Among the various industry categories, firms in the finan-
cial services, technology, consumer packaged goods, and
government industry categories provide, on average, a
higher contribution margin than do firms in other industry
categories. However, firms in the education industry pro-
vide a lower contribution margin than do firms in other
industries.
Customer Selection Strategy
In this section, we compare the customer selection capabili-
ties of the following: CLV, our proposed metric; previous-
period customer revenue (PCR), a simple metric; past cus-
tomer value (PCV), which is widely considered a good
TABLE 5
Regression Results from the Contribution Margin
Model
Parameter
Independent Variable Estimate
Intercept N.S. (N.S.)
Contribution in t – 2 .83*** (.85***)
Lagged total quantity purchased .02** (.03**)
Size .02** (.02**)
Aerospace N.S. (N.S.)
Financial services .02* (.02**)
Manufacturing N.S. (N.S.)
Technology .03** (.02**)
Consumer packaged goods .03*** (.03***)
Education, K–12 –.03*** (–.02***)
Travel N.S. (N.S.)
Government .02* (.02*)
Lagged total level of marketing effort .04*** (.06***)
*Significant at α = .10.
**Significant at α = .05.
***Significant at α < .01.
Notes: The reported coefficients are standardized estimates; the
values in parentheses represent Cohort 2. N.S. = not
significant.
Customer Lifetime Value Framework / 119
5
We also compared the metrics with a censoring time at 18
months and prediction window of 30 months. The substantive
results of the study hold even for a 30-month prediction window.
The results are available on request from the authors.
predictor of future customer value; and customer lifetime
duration (CLD), a forward-looking metric that is also used
as a proxy for loyalty. (We also compared the customer
selection capabilities of CLV with other customer-based
metrics, such as share of wallet and recency, frequency, and
monetary value. The results were similar to the comparison
with PCR, PCV, and CLD.) In general, organizations in
direct marketing situations rank-order their customers on
the basis of a particular metric and prioritize their resources
from best customers to worst customers on the basis of the
rank order (Roberts and Berger 1999). Descriptions of the
various customer-based selection metrics used in our analy-
ses are provided in Appendix B.
Performance of the Customer Selection Metrics
To compare the performance of the four metrics, we rank-
ordered customers from best to worst according to each
metric and then compared the sales, costs, and profits from
the top 5%, 10%, and 15% of customers. We used the data
from the first 30 months to score and sort the customers on
each metric, as do Reinartz and Kumar (2003). We then
compared the actual sales, variable costs of communication,
and profits for the top 5%, 10%, and 15% of customers
from the censoring period (30 months) until the end of the
observation window (48 months).
5
To select customers for
contact, in general organizations choose the top 5% to 15%
of their customers, rank-ordered on the basis of a scoring
metric. Selection of more customers to contact may not be
feasible because of limited time and resources. Thus, to
reflect industry practice, we compared the performance of
our metric among the top 5%, 10%, and 15% of customers.
The results are provided in Table 6, and the reported values
are cell medians. We subsequently summarize the results
from our comparison.
Overall, Table 6 shows that the proposed metric better
identifies profitable customers than do other metrics we
compared in the study, such as PCR, PCV, and CLD. On the
basis of the 18-month prediction window, we expect the
average net profits of customers selected from the top 5%
using the proposed CLV metric to be $143,295 (after
accounting for cost of goods sold [70%] and variable costs
of communication), whereas the average net profits are
$70,929, $130,785, and $106,389 for the top 5% of the cus-
tomers selected on the basis of PCR, PCV, and CLD,
respectively. These findings hold across all the percentage
subgroups. The results provide substantial support for
incorporating the responsiveness of each individual cus-
tomer across various communication channels and for the
usefulness of CLV as a metric for customer scoring and cus-
tomer selection. Although the difference in profits from use
of PCR, PCV, CLD, and CLV is, on average, approximately
$40,000 for a customer in the top 5% sample, the difference
in total profit across the top 5% to 15% of the entire cus-
tomer base can easily yield more than $1 million.
Resource Allocation Strategy
Having evaluated the usefulness of using CLV for customer
selection, we now describe our methodology for designing
resource allocation strategies that maximize CLV. The
framework also provides managers a tool for assessing
return on marketing investments by identifying avenues for
optimal resource allocation across channels of communica-
tion for each individual customer (and possibly across cus-
tomers), so as to maximize CLV. The marketing literature
has provided guidelines for optimal resource allocation in
acquisition and retention decisions (Blattberg and Deighton
1996; Blattberg, Getz, and Thomas 2001), promotion
expenditures (Berger and Bechwati 2001; Berger and Nasr
1998), and marketing actions when future brand switching
TABLE 6
Comparisons of CRM Metrics for Customer Selection
Percentage of Cohort
(Selected from Top) CLV PCR PCV CLD
5%
Gross profit ($) 144,883 71,908 131,735 107,719
Variable costs ($) 1,588 00,979 000,950 000,790
Net profit ($) 143,295 70,929 130,785 106,389
10%
Gross profit ($) 78,401 27,981 72,686 55,837
Variable costs ($) 1,245 00,943 000,794 00,610
Net profit ($) 77,156 27,038 71,892 55,227
15%
Gross profit ($) 56,147 15,114 52,591 44,963
Variable costs ($) 00,807 00,944 000,809 00,738
Net profit ($) 55,340 14,170 51,782 44,225
Notes: All metrics are evaluated at 30 months, with an 18-month prediction window. Cohort 2 provides similar results. The reported values are
cell medians. Gross profit is residual revenue after removing cost of goods sold. In general, for the firm that provided the database, the
cost of goods sold is approximately 70%; thus, gross profit = revenue × .3.
120 / Journal of Marketing, October 2004
6
In Equation 2, the level of contacts in each channel for each
customer is varied each year. However, to simplify our optimiza-
tion routine, we assumed that the level of contacts is equal across
the prediction period. Our assumption can be viewed as taking the
average level of contacts in the prediction periods.
is considered (Rust, Zeithaml, and Lemon 2004). These
guidelines represent a significant step toward incorporation
of long-term customer profitability effects into firm-level
managerial decision making. However, the models provide
less insight into decisions about how to manage individual
customers in a way that accounts for the heterogeneity, and
they do not provide a mechanism for dynamic updating of
profitability assessment (Libai, Narayandas, and Humby
2002).
Our resource allocation algorithm uses Equation 2 as
the objective function, and the purpose of the optimization
is to find the level of contacts across various channels with
each individual customer that would maximize CLV. Equa-
tion 2 is a function of predicted purchase frequency (based
on Equation 4), predicted contribution margin (based on
Equation 5), and marketing costs. We first estimated the
responsiveness of customers (coefficients, β
s
) to marketing
contacts from the purchase frequency model and the contri-
bution margin model. To design a CLV-based resource allo-
cation strategy, we held the coefficients constant and identi-
fied the level of covariates (i.e., level of channel contacts)
for each customer that would maximize CLV. In summary,
in Equation 2, the contacts made to a customer across vari-
ous channels are under the supplier’s control and thus can
be used to maximize CLV, depending on the cost of each
mode of communication and the responsiveness of the cus-
tomer (in terms of both purchase frequency and contribu-
tion margin) to each channel of communication.
6
In other
words, with respect to marketing resources for a firm, the
customer contact levels across different channels appear in
the revenue and cost sides of Equation 2 and thus avoid the
scope of corner solutions.
We used a genetic algorithm to derive the levels of con-
tact desired for each individual customer that maximize
CLV. Genetic algorithms (Goldberg 1989) are simulation-
based, parallel-search algorithms that have been used in
econometrics (Dorsey and Mayer 1995; Liang and Wong
2001) to obtain optimal solutions when the complexity of
the optimization function tends to be intractable and multi-
dimensional. In our study, support for a purchase frequency
distribution with two subgroups led us to believe that the
optimization surface is multimodal. In addition, we
intended to allocate resources for each customer on the
basis of individual responsiveness. These issues made our
optimization problem extremely complex and intractable
with traditional analytical methods. Thus, we resorted to a
search algorithm to find the optimal resource allocation lev-
els. In addition, the multimodal nature of the optimization
surface (given the support for a mixture distribution for pur-
chase frequency) motivated us to use a parallel-search tech-
nique, which is not susceptible to local minima (common in
multimodal distributions) (Venkatesan, Krishnan, and
Kumar 2004). Appendix C explains how we used a parallel-
search technique for resource allocation purposes.
Aggregate Results
The total net present value of future profits from a resource
allocation strategy that maximizes CLV (with the predicted
contribution margin) is approximately $44 million. We also
computed the total net present value of future profits when
the organization uses its current resource allocation strat-
egy. Specifically, for each customer, we maintained the
resource allocation levels for the most recent year and cal-
culated CLV over a three-year period. We find that, based
on this status quo resource allocation strategy, the total net
present value of future profits is approximately $24 million.
Therefore, we find that a resource allocation strategy that
maximizes CLV results in an increase in profits by approxi-
mately 83%. The total cost of communication (over three
years), based on the resource allocation strategy that maxi-
mizes CLV, is approximately $1 million. The total cost of
communication in the organization’s current strategy is
approximately $716,188. We find that the organization
improves profits by increasing costs of serving customers
(rather than cost of communication in the previous year) by
48%. The return on marketing communication to the orga-
nization, based on its current strategy, is approximately $34
million ($24 million/$716,188). With a communication
strategy that maximizes CLV, the return on marketing com-
munication to the organization is approximately $44 million
($44 million/$1 million). Thus, it is possible to improve
profits and return on marketing communication by appro-
priately identifying customers for target communications
and by matching the channel of communication with cus-
tomer preferences. The aggregate results suggest that given
the improvement of approximately $20 million among a
sample of 216 customers, there is a potential for the firm to
increase its revenue by at least $1 billion across its entire
customer base. However, such benefits may not be realized
immediately because the firm also needs to incur costs to
move toward a customer-centric view and to train its
employees to manage customers on the basis of CLV.
Implications, Limitations, and
Further Research
The objective of our study was to analyze the usefulness of
CLV as a metric for customer selection and resource alloca-
tion strategy. First, we developed and estimated an individ-
ual customer-level objective function, the goal of which is
to measure CLV. Second, we demonstrated the superiority
of selecting customers for contact on the basis of CLV com-
pared with commonly used metrics such as PCR, PCV, and
CLD. Third, we evaluated the benefits of designing market-
ing communications that maximize CLV. We now discuss
the implications of our study and how managers can use
this knowledge to design efficient marketing programs. We
also provide an outline for future researchers to build on the
framework proposed herein.
Implications
Antecedents of purchase frequency and contribution
margin. The theoretical implications of the purchase fre-
quency model are also related to the CUSAMS customer
Customer Lifetime Value Framework / 121
asset management framework (Bolton, Lemon, and Verhoef
2004). In this study, we tested parts of the CUSAMS frame-
work and found empirical support for the parts we tested.
Specifically, the CUSAMS framework proposes that mar-
keting instruments (e.g., direct mailings, reward programs)
affect a customer’s price perceptions, satisfaction, and com-
mitment. In turn, these affect the length, depth, and breadth
of a relationship, which then ultimately influence CLV. We
find empirical support for marketing instruments’ effects on
purchase frequency (rich and standardized modes of com-
munication and relationship benefits) and contribution mar-
gin (total marketing efforts), both of which ultimately influ-
ence CLV. We also find that breadth of purchases
(cross-buying) and depth of buying (upgrading) affect pur-
chase frequency, which ultimately influences CLV. In addi-
tion, we find support for a nonlinear relationship between
supplier communications and purchase frequency. This sup-
ports Fournier, Dobscha, and Mick’s (1997) expectations
that too much communication between suppliers and cus-
tomers can be disruptive. Thus, our results indicate that
managers need to be cautious when designing marketing
communication strategies across different channels and
need to be wary of contacting customers too many times,
especially through rich modes of communication.
We find an inverted U-shaped relationship between
returns and purchase frequency. A possible benefit from
customers who return products could be the opportunity to
understand the reasons for dissatisfaction. In addition, cus-
tomers who return products within a certain period may do
so because they have inherent trust in the supplier and
because they expect future benefits, such as improvements
in the quality of the product. However, a customer’s return-
ing too many times may indicate erosion of trust with the
firm or a lower level of future activity. We also find that
customers that establish contact through the online channel
of communication exhibit high frequency of purchases and
have high involvement. Therefore, the online channel can
provide an ideal setting for B2B firms to enhance their cus-
tomer relationships. We find that in addition to influencing
purchase frequency, marketing communications influence
the expected contribution margin from a customer. Also in
the B2B scenario, industry category and size seem to be
important factors that influence the magnitude of contribu-
tion margin.
Enhancement of marketing productivity. Rust, Zeithaml,
and Lemon (2004) propose that firm strategies and tactical
marketing actions affect the marketing productivity chain.
Our analyses of customer selection investigate how firms
can enhance strategies, and our analysis of optimal resource
allocation investigates how firms can improve tactical mar-
keting actions.
Customer selection. A CLV metric better identifies cus-
tomers that provide higher future profits than do PCR, PCV,
and CLD. Our analyses indicate that CLV is preferred to
incorporate the dynamics of customer purchase behavior
into the customer selection process. Managers can substan-
tially improve their return on marketing investments by
using a dynamic, customer-level measure of CLV for scor-
ing rather than using the other metrics suggested in the lit-
erature and by prioritizing contact programs.
Resource allocation strategy. The results from our study
highlight the importance of firms’ considering individual
customers’ responsiveness to marketing communication as
well as the costs involved across various channels of com-
munication when making resource allocation decisions. Our
analyses suggest that there is a potential for substantial
improvement in CLV through appropriate design of market-
ing contacts across various channels. When firms design
resource allocation rules, they can realize the increase in
profits by incorporating the differences in individual cus-
tomer responsiveness to various channels of communication
and the potential value provided by the customer. The pro-
posed resource allocation strategy can be a basis for evalu-
ating the potential benefits of CRM implementations in
organizations, and it provides accountability for strategies
geared toward managing customer assets.
To summarize, the major conclusions that we derive
from our study are the following:
•Marketing communication across various channels affects
CLV nonlinearly;
•CLV performs better than other commonly used customer-
based metrics for customer selection such as PCR, PCV, and
CLD; and
•Managers can improve profits by designing marketing com-
munications that maximize CLV.
Limitations and Further Research
The study has limitations that further studies can address.
The results of this study are from a customer database in the
high-technology industry. Further studies need to investi-
gate whether the results are generalizable to other industries
and settings. In addition, further research needs to develop
models that combine forecasts of aggregate competitive
responses to marketing actions and customer brand switch-
ing with individual-level models of direct marketing. We
also consider only the average levels of optimal communi-
cation strategy in a channel. However, firms can further
improve the efficiency of communication strategy by appro-
priately sequencing their customer contacts across different
channels. In addition, in our study, we provide a framework
for maximizing CLV with marketing communications.
However, note that we do not compare our proposed
resource allocation strategy (that focuses on maximizing
CLV) with a strategy that focuses on allocating resources to
customers for which the increment in CLV from appropriate
design of marketing communication is highest. For exam-
ple, with an appropriately designed marketing strategy, it is
possible that customers that previously had high CLV con-
tinue to have high CLV in the future, irrespective of the
level of marketing communications, and that some
customers that have had low CLV transform to high-CLV
customers. Further research can investigate whether the
customers selected for high levels of marketing communi-
cations are the same when the resource allocation strategy
focuses on maximizing CLV or on maximizing incremental
CLV. Finally, the sum of optimal CLVs for each individual
customer need not lead to the optimal customer equity in
122 / Journal of Marketing, October 2004
the case of a budget constraint. We performed our optimiza-
tion with a budget constraint, and there were no changes in
the substantive results of the study. In addition, our
optimization algorithm is flexible enough to allow for inclu-
sion of the budget constraint without any substantial
adjustments.
The customer- and supplier-specific antecedents used in
the customer response model also can directly affect costs
and thus margins. However, because we estimated the cus-
tomer response model in a single step, which we then
included in a net present value function (Equation 1) that
includes both costs and margins, we assessed the indirect
effect of the covariates on both costs and margins. Further
studies can develop and test hypotheses that directly relate
CRM efforts to costs and margins. In addition, it can be
expected that margins change over time. In this case, the
value that a customer provides to a firm is a function of
both the expected time frame until the next purchase and
the contribution margin at that particular period. We treated
several antecedents used in our framework (e.g., upgrading,
cross-buying, bidirectional communication, number of
returns, number of Web-based contacts) as exogenous vari-
ables in our analysis. We used the lagged variables of these
to account for potential endogeneity. Further research can
investigate more sophisticated techniques that explicitly
treat these variables as endogenous. Finally, a notable issue
that arises from our analyses is whether the recommenda-
tions from an optimization framework pan out when imple-
mented in the real market. Although our study is a step in
the right direction to assess the accountability of marketing
actions, a field experiment that tests the recommendations
of such a framework on a test group against a control group
that is managed according to existing norms would provide
a stronger justification for CRM-based efforts.
Appendix A
Results from Data Analyses
Purchase Frequency Model
Model estimation. For Cohorts 1 and 2, the results are
based on 50,000 samples of the MCMC algorithm (for
details of the algorithm, see Allenby, Leone, and Jen 1999,
Appendix). We simulated the posterior distribution using
five parallel chains with overdispersed starting values. In
each chain, we used the initial 40,000 iterations as burn-in
and used the last 10,000 iterations to obtain posterior statis-
tics. We used “slice-sampling” (Neal 2000) to obtain ran-
dom samples. We assessed the autocorrelations of the poste-
rior samples to perform thinning. The autocorrelation
functions revealed that every fifth sample is unrelated to
every other fifth sample. Thus, the posterior statistics are
based on 2000 samples (using every fifth sample from
10,000 samples). We assessed the convergence of the algo-
rithm from the line plots of the posterior sample and by
evaluating Gelman and Rubin’s (1992) √ statistic. Values
of √ that are closer to 1 reveal that all the chains have con-
verged to true posterior distribution. In our analyses, we
found that the value of √ ranged from 1.2 to .9 (in previ-

R

R

R
ous research [Cowles and Carlin 1996], these values have
been found to indicate convergence of the MCMC chains).
Our approach to investigating within-chain autocorrela-
tions, computing Gelman and Rubin’s statistic, and visually
inspecting the sampling plots has been recommended
widely for convergence diagnostics (Cowles and Carlin
1996). We implemented the estimation algorithm using the
GAUSS software package, and we performed the conver-
gence diagnostics using CODA. We also used multiple prior
values for the estimates and did not find a significant impact
on the posterior samples; this is because we used diffuse
prior values and because of the large sample size of our data
set.
Model selection. Table 4 shows that a simple general-
ized gamma model without the probit link function or time-
varying covariates (Model 1) provides a log-likelihood of
–3298.91 (we calculated the log-likelihood using Newton
and Raferty’s [1994] log marginal density measure). A gen-
eralized gamma model with the probit link function (two
subgroups) but no time-varying covariates (Model 2) pro-
vides a log-likelihood of –2908.66. Finally, the log-
likelihood for the generalized gamma model with the probit
link function (with two subgroups) and time-varying covari-
ates (Model 3) is –2237.82. Increasing the number of sub-
groups in Models 2 or 3 did not result in significantly
higher log-likelihoods. We also computed the AIC and BIC
for model selection. For both measures, a higher value indi-
cates a better model. We find that Model 3 has the highest
value for both AIC and BIC (AIC = –4484 for Model 3,
–5825 for Model 2, and –6606 for Model 1; BIC = –4683
for Model 3, –5910 for Model 2, and –6639 for Model 1).
In addition, a latent-class finite mixture model with two
segments provides a likelihood of –2938. Overall, the
results show that the generalized gamma model with two
subgroups provides a good fit to the data.
Distribution parameters. The values of γ in the general-
ized gamma model are γ
1
= 1.2 and γ
2
= .9 for Subgroups 1
and 2, respectively. The component masses for Subgroup 1

1
) and Subgroup 2 (ϕ
2
) are .54 and .46, respectively. All
the distribution parameters (α
k
, ν
k
, and θ
k
) have more than
99% of their samples different from zero. The mean
expected interpurchase time for Subgroup 1 is 4.2 pur-
chases in a year, and the mean expected frequency for Sub-
group 2 is 1.01 purchases in a year. Given the variation in
expected frequencies in each subgroup, we term Subgroup
1 the “active state” and Subgroup 2 the “inactive state.”
Influence of antecedents and covariates. Our empirical
analyses support all the proposed effects of the antecedents
on purchase frequency.
Out-of-sample forecasting accuracy. We used relative
absolute error (RAE) to evaluate the forecasting accuracy of
the generalized gamma model compared with that of a
naive moving average model (the moving average model is
implemented as an updated average of every consecutive
interpurchase time for a customer). We used all but the last
observation for each customer (calibration sample) to esti-
mate the parameters of each model. We then used the mean
posterior values to compute the expected time until next
Customer Lifetime Value Framework / 123
purchase from Equation 3. We then compared the forecast
time until the next purchase with the holdout sample
(formed from the last observation for each customer) to
assess the mean absolute deviation (MAD). The RAE is
given by the ratio of the model MAD to the MAD based on
the moving average measure. Based on the RAE measure,
the generalized gamma model with time-varying covariates
(Model 3) has an RAE of .51, compared with that of a naive
moving average technique. The MAD from Model 3 is 5.21
months, compared with 10.24 months for the moving aver-
age measure. Model 3 provides the best improvement in
forecasting accuracy compared with Model 2 (RAE = .84)
and Model 1(RAE = .95). We also assessed the predictive
capability of the purchase frequency model using hit rate. In
other words, we assessed the number of purchases in the
holdout sample that the model also correctly predicted as a
purchase. We observe that among customers who bought
within 12 months, the model currently identifies 89% of
them, and among customers who did not buy within 12
months, the model currently identifies 90% of them.
Contribution Margin Model
Estimation and influence of antecedents and covariates.
We estimated the contribution margin model on annual data
from 1997 (t – 4) to 2000 (t) for Cohort 1 and from 1998 to
2000 for Cohort 2. The revenue in 2001 (t + 1) is a holdout
sample for Cohorts 1 and 2. The coefficients of the contri-
bution margin model are provided in Table 5. The contribu-
tion margin model provides an adjusted R
2
of .68. Overall,
the results of the analyses support all the hypothesized
relationships.
Out-of-sample forecasting accuracy. As with typical
time-series models, we advanced the independent variables
by one period to forecast the dependent variable in period
t + 1. Specifically, when the contribution margin model is
used to predict period t + 3, the contribution margin in
period t + 1 is an independent variable and is obtained from
the prediction in period t + 1. The contribution margin
model predicts the growth in revenue from period t to t + 1.
We obtained the magnitude of revenue in period t + 1 by
adding the predicted value from the contribution margin
model to the base revenue in period t. We evaluated the per-
formance of the contribution margin model in the holdout
sample on the basis of our estimates from the calibration
sample. Table A1 provides the descriptive statistics of the
observed contribution margin and the predicted contribution
margin in the holdout sample (period t + 1). In the holdout
sample, the mean predicted contribution margin in period
t + 1 is approximately $67,729, and the mean observed con-
tribution margin is approximately $64,396.
Appendix B
Description of Customer Selection
Metrics
CLV
We entered predictions from the purchase frequency (Equa-
tion 3) and contribution margin (Equation 4) models into
Equation 2 to obtain the net present value of future profits
(period t + 1) from each customer. The purchase frequency
model predicts the expected time in months until next pur-
chase for each customer. We assigned a 30% margin after
accounting for cost of goods sold (the managers who pro-
vided the database informed us that 30% was a nominal
margin for most of their products), and we computed the
variable costs using costs of communication. The mean unit
cost of standardized modes of communication is approxi-
mately $3, and the mean unit cost of rich modes of commu-
nication is $60. We computed the unit cost of communica-
tion for each customer as the ratio of the total contacts for a
given channel in a given year to the total cost of contact for
a given mode in a given year. Finally, we used an annual
discount rate of 15% for each customer, which is based on
the lending rate that is appropriate for the time of the study.
PCR and PCV
We define PCR as the revenue provided by the customer in
the most recent observed purchase. We define PCV as the
cumulative profits obtained from a customer until the cur-
rent period. The cumulative profits are calculated annually
from a customer’s initiation until the current period. We
projected the profit in each year to current terms using a
discount factor. The PCV calculation is as follows:
where CM
i,t
is the contribution margin for customer i in
period t; MC
i,t
denotes marketing costs for customer i in
period t; t is an index for time period (t = 0 for the period of
customer initiation; for example, t = 0 for 1997 for Cohort 1
customers, and t = 0 for 1998 for Cohort 2 customers); T is
( ) ( ) B PCV CM MC r
i it it
T t
t
1 1 − ( ) × +

00
T

,
TABLE A1
Comparison of Descriptive Statistics Between Observed Contribution Margin and Predicted Contribution
Margin in the Holdout Sample
Mean Standard Deviation Minimum Maximum
Observed 50,199 23,850 –171 1,010,881
(49,229) (24,836) (–181) (1,420,981)
Predicted 57,729 24,538 –178 1,897,257
(74,283) (22,598) (–159) (1,938,458)
Notes: All reported values are in dollars and are rounded to the nearest integer. Values in parentheses represent Cohort 2.
124 / Journal of Marketing, October 2004
the current period; and r is the discount rate, which we set at
15%.
CLD
In our analysis, we evaluated the probability that a customer
is alive or dead in the planning window using the P(Alive)
measure that Schmittlein and Peterson (1987) and Reinartz
and Kumar (2002) recommend. The P(Alive) measure uses
the previous purchase pattern to predict the probability that
a customer is still alive at each period in the prediction win-
dow. Higher values of P(Alive) indicate longer lifetime
duration.
Appendix C
Genetic Algorithms to Develop
Resource Allocation Strategies
Researchers in marketing have only recently begun to rec-
ognize the potential benefits of using genetic algorithms
(Balakrishnan and Jacob 1996; Midley, Marks, and Cooper
1997; Naik, Mantrala, and Sawyer 1998; Venkatesan,
Krishnan, and Kumar 2004) in deriving optimal strategies
for complex marketing problems. The genetic algorithm
proceeds by searching for the optimal level of contact for
each customer that maximizes CLV. The sum of the optimal
CLVs from each individual customer provides the optimal
customer equity of the analysis sample.
7
Specifically, we
varied the contact levels for each customer and then calcu-
lated the sum of CLV of all customers in the sample. Our
objective was to calculate the maximum value for this sum
of the CLVs. In this case, our optimization algorithm maxi-
mized the objective function by varying 432 parameters in
Cohort 1 (216 customers and 2 parameters for each cus-
tomer [levels of rich and standardized modes]). Following
research in customer equity (Rust, Zeithaml, and Lemon
2004), we set the time frame for our optimization frame-
work as three years. Our database provides information on
the approximate unit cost of communication through rich
modes and standardized modes to each customer. On aver-
age, the unit marketing cost through standardized modes
(average of direct mail and telephone sales) is $3, and the
unit marketing cost through rich modes (salesperson con-
tacts) is $60. Thus, the cost of communication through rich
modes is approximately 20 times the cost of communica-
tion through standardized modes. Such a cost index is com-
monly encountered. We set the parameters in the genetic
algorithm as follows: population size = 200, probability of
crossover = .8, probability of mutation = .25, and conver-
gence criteria = difference in solution in the last 10,000 iter-
ations should be less than .01%. We ran the genetic algo-
rithm at least 50 times and used the mean of the resource
levels corresponding to the maximum CLV from each run
as the resource reallocation rule for each customer.
REFERENCES
7
When forecasting contribution margin for three years ahead,
note that we have data on marketing activities from 1997 to 2001.
Our last prediction for the long-term analysis is in period 2003, for
which we need marketing variables in 2002, which we do not
have. For forecasting the contribution margin in 2003, we used the
average of marketing-mix variables from 1997 to 2001 for imput-
ing the values in 2002. However, for the optimization phase, the
optimized marketing-mix variables are substituted for the 2002
values. We thank a reviewer for bringing this to our attention.
Allenby, Greg M., Robert P. Leone, and Lichung Jen (1999), “A
Dynamic Model of Purchase Timing with Application to Direct
Marketing,” Journal of American Statistical Association, 94
(June), 365–74.
Balakrishnan, P.V. and Varghese A. Jacob (1996), “Genetic Algo-
rithms of Product Design,” Management Science, 42 (8),
1105–1118.
Baltagi, Badi H. (1998), Econometric Analysis of Panel Data.
New York: John Wiley & Sons.
Bell, David, John Deighton, Werner J. Reinartz, Roland T. Rust,
and Gordon Swartz (2002), “Seven Barriers to Customer
Equity Management,” Journal of Service Research, 5 (1),
77–85.
Berger, Paul D. and Nada Nasr Bechwati (2001), “The Allocation
of Promotion Budget to Maximize Customer Equity,” OMEGA,
29 (1), 49–61.
———, Ruth N. Bolton, Douglas Bowman, Elten Briggs, V.
Kumar, A. Parasuraman, and Creed Terry (2002), “Marketing
Actions and the Value of Customer Assets: A Framework for
Customer Asset Management,” Journal of Service Research, 5
(1), 39–54.
——— and Nada I. Nasr (1998), “Customer Lifetime Value: Mar-
keting Models and Applications,” Journal of Interactive Mar-
keting, 12 (Winter), 17–30.
Blattberg, Robert C. and John Deighton (1996), “Manage Market-
ing by the Customer Equity Test,” Harvard Business Review,
74 (July–August), 136–44.
———, Gary Getz, and Jacquelyn S. Thomas (2001), Customer
Equity: Building and Managing Relationships as Assets.
Boston: Harvard Business School Press.
——— and Edward Malthouse (2003), “Evaluating the Predictive
Accuracy of Long-Term Value Models for Customer Relation-
ship Management Strategies,” working paper, Department of
Marketing, Northwestern University.
Bolton, Ruth and Katherine N. Lemon (1999), “A Dynamic Model
of Customers’ Usage of Services: Usage as an Antecedent and
Consequence of Satisfaction,” Journal of Marketing Research,
36 (May), 171–86.
———, ———, and Peter C. Verhoef (2004), “The Theoretical
Underpinnings of Customer Asset Management: A Framework
and Propositions for Future Research,” Journal of the Academy
of Marketing Science, 32 (3), 271–92.
Bowman, Douglas and Das Narayandas (2001), “Managing
Customer-Initiated Contacts with Manufacturers: The Impact
on Share of Category Requirements and Word-of-Mouth
Behavior,” Journal of Marketing Research, 38 (August),
281–97.
Cannon, Joseph P. and Christian Homburg (2001), “Buyer–Seller
Relationships and Customer Firm Costs,” Journal of Market-
ing, 65 (January), 29–43.
Cowles, Mary K. and Bradley P. Carlin (1996), “Markov Chain
Monte Carlo Convergence Diagnostics: A Comparative
Review,” Journal of American Statistical Association, 91
(June), 883–904.
Del Moral, P. and L. Miclo (2001), “Asymptotic Results for
Genetic Algorithms with Applications to Nonlinear Estima-
tion,” in Theoretical Aspects of Evolutionary Computation, L.
Kallel, B. Naudts, and A. Rogers, eds. Berlin: Springer-Verlag,
439–94.
Dhar, Ravi and Rashi Glazer (2003), “Hedging Customers,” Har-
vard Business Review, 81 (May), 3–8.
Dorsey, Robert E. and Walter J. Mayer (1995), “Genetic Algo-
rithms for Estimation Problems with Multiple Optima,” Jour-
nal of Business and Economic Statistics, 13 (1), 53–67.
Dwyer, Robert F., Paul H. Schurr, and Sejo Oh (1987), “Develop-
ing Buyer–Seller Relationships,” Journal of Marketing, 51
(April), 11–27.
Fournier, Susan, Susan Dobscha, and David Glen Mick (1997),
“Preventing the Premature Death of Relationship Marketing,”
Harvard Business Review, 75 (January–February), 2–8.
Ganesan, Shankar (1994), “Determinants of Long-Term Orienta-
tion in Buyer–Seller Relationships,” Journal of Marketing, 58
(April), 1–19.
Gelman, A. and D.B. Rubin (1992), “Inference from Iterative Sim-
ulation Using Multiple Sequences,” Statistical Science, 7 (3),
457–511.
Goldberg, David E. (1989), “A Simple Genetic Algorithm,” in
Genetic Algorithms in Search Optimization and Machine
Learning. Boston: Addison-Wesley Longman, 10–14.
Grewal, Rajdeep, James M. Corner, and Raj Mehta (2001), “An
Investigation into the Antecedents of Organizational Participa-
tion in Business-to-Business Electronic Markets,” Journal of
Marketing, 65 (July), 17–34.
Gupta, Sunil (1988), “The Impact of Sales Promotion on When,
What, and How Much to Buy,” Journal of Marketing Research,
25 (November), 342–55.
Hogan, John E., Donald R. Lehmann, Mario Merino, Rajendra K.
Srivastava, Jacquelyn S. Thomas, and Peter C. Verhoef (2002),
“Linking Customer Assets to Financial Performance,” Journal
of Service Research, 5 (August), 26–38.
Jap, Sandy D. (2001), “The Strategic Role of Salesforce in Devel-
oping Customer Satisfaction Across the Relationship Lifecy-
cle,” Journal of Personal Selling & Sales Management, 21 (2),
95–108.
Kamakura, Wagner and Gary Russell (1989), “A Probabilistic
Choice Model for Market Segmentation and Elasticity Struc-
ture,” Journal of Marketing Research, 26 (November), 379–90.
Liang, Faming and Wing Hung Wong (2001), “Real-Parameter
Evolutionary Monte Carlo with Application to Bayesian Mix-
ture Models,” Journal of the American Statistical Association,
96 (454), 653–66.
Libai, Barak, Das Narayandas, and Clive Humby (2002), “Toward
an Individual Customer Profitability Model: A Segment Based
Approach,” Journal of Service Research, 5 (1), 69–76.
Midley, David F., Robert E. Marks, and Lee G. Cooper (1997),
“Breeding Competitive Strategies,” Management Science, 43
(3), 257–75.
Mohr, Jakki and John R. Nevin (1990), “Communication Strate-
gies in Marketing Channels: A Theoretical Perspective,” Jour-
nal of Marketing, 54 (October), 36–51.
Morgan, Robert M. and Shelby D. Hunt (1994), “The
Commitment–Trust Theory of Relationship Marketing,” Jour-
nal of Marketing, 58 (July), 20–38.
Customer Lifetime Value Framework / 125
Naik, Prasad A., Murali K. Mantrala, and Alan G. Sawyer (1998),
“Planning Media Schedules in the Presence of Dynamic Adver-
tising Quality,” Marketing Science, 17 (3), 214–35.
Nash, Edward L. (1993), Database Marketing: The Ultimate Mar-
keting Tool. New York: McGraw-Hill.
Neal, Radford M. (2000), “Slice-Sampling,” Technical Report No.
2005, Department of Statistics, University of Toronto.
Newton, M.A. and A.E. Raferty (1994), “Approximate Bayesian
Inference with the Weighted Likelihood Bootstrap,” Journal of
the Royal Statistical Society, (Series B, 56), 3–48.
Niraj, Rakesh, Mahendra Gupta, and Chakravarthi Narasimhan
(2001), “Customer Profitability in a Supply Chain,” Journal of
Marketing, 65 (July), 1–16.
Reinartz, Werner J. and V. Kumar (2000), “On the Profitability of
Long-Life Customers in a Noncontractual Setting: An Empiri-
cal Investigation and Implications for Marketing,” Journal of
Marketing, 64 (October), 17–35.
——— and ——— (2002), “The Mismanagement of Customer
Loyalty,” Harvard Business Review, 80 (7), 86–94.
——— and ——— (2003), “The Impact of Customer Relation-
ship Characteristics on Profitable Lifetime Duration,” Journal
of Marketing, 67 (January), 77–99.
Rindfleisch, Aric and Jan B. Heide (1997), “Transaction Cost
Analysis: Past, Present, and Future Applications,” Journal of
Marketing, 61 (October), 30–55.
Roberts, Mary Lou and Paul Berger (1999), Direct Marketing
Management, 2d ed. Upper Saddle River, NJ: Prentice Hall.
Rust, Roland T., Valarie A. Zeithaml, and Katherine N. Lemon
(2004), “Return on Marketing: Using Customer Equity to
Focus Marketing Strategy,” Journal of Marketing, 68 (Janu-
ary), 23–53.
Sawhney, Mohanbir (2001), “Don’t Homogenize, Synchronize,”
Harvard Business Review, 81 (July–August), 101–108.
Schmittlein, David and Donald G. Morrison (1985), “Is the Cus-
tomer Alive?” The American Statistician, 39 (4), 291–95.
——— and Robert A. Peterson (1994), “Customer Base Analysis:
An Industrial Purchase Process Application,” Marketing Sci-
ence, 13 (1), 41–67.
Shepard, David (2001), The New Direct Marketing: How to Build
a Profit-Driven Database Marketing Strategy, 3d ed. New
York: McGraw-Hill.
Sheth, Jagdish N. and Atul Parvatiyar (1995), “Relationship in
Consumer Markets: Antecedents and Consequences,” Journal
of the Academy of Marketing Science, 23 (4), 255–71.
Tellis, Gerard and Fred Zufryden (1995), “Tackling the Retailer
Decision Maze: Which Brands to Discount, How Much, When,
and Why?” Marketing Science, 14 (3), 271–300.
Thompson, Bob (2001), “CRM and the Internet: Leading Edge
Strategies for a Multi-Channel World,” (accessed May 19,
2004), [available at http://www.businessweek.com/adsections/
care/crm/crm_index.htm].
Venkatesan, Rajkumar, Trichy V. Krishnan, and V. Kumar (2004),
“Evolutionary Estimation of Macro-Level Diffusion Models,”
Marketing Science, forthcoming.
Villas-Boas, Miguel J. and Russell S. Winer (1999), “Endogeneity
in Brand Choice Models,” Management Science, 45 (10),
1324–38.

TABLE 1 Comparing the Proposed CLV Framework with Existing Models
Return on Investment Modeled and Calculated No Yes Yes How Market ing Communication Affects CLV Yes Yes No Resource Allocation for Each Customer No No No Resource Allocation Across Channels No Yes No Comparison of CustomerBased Metrics No No No

Type of Model CLV

Representative Research Berger and Nasr (1998) Berger et al. (2002)

CLV Calculation Yes Yes Yes

CLV-Based Resource Allocation No Yes Yes

Statistical Details Yes No Yes

Customer equity

Blattberg and Deighton (1996) Libai, Narayandas, and Humby (2002)

Yes

Yes

Yes

No

No

No

No

No

Database marketing

Reinartz and Kumar (2000) Bolton, Lemon, and Verhoef (2004)

Yes Yes

Yes Yes

No Yes

No No

No No

No No

No No

Yes Yes

Customer Lifetime Value Framework / 107

CLV antecedents

Reinartz and Kumar (2003) Rust, Zeithaml, and Lemon (2004)

Yes Yes

Yes Yes

Yes Yes

No No

No No

No No

Yes Yes

Yes Yes

CLV-based resource allocation

Berger and Bechwati (2001) Present Study

Yes

Yes

Yes

Yes

No

No

No

No

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

frequencyi = predicted purchase frequency for customer i. Our objective there is to evaluate the extent to which CLV. the measure assumes that when a customer terminates a relationship. Thus. and the marketing resources allocated to the customer. Schurr. In addition. Thus. our objective is to allocate resources so as to maximize CLV. managers who calculate CLV in noncontractual settings are interested in predicting future customer activity and the predicted con- 108 / Journal of Marketing. Thus. In addition.l = number of contacts to customer i in channel m in year l. in noncontractual settings.y y − i + r) frequency l =1 ∑ n ∑ m c i. In the section “Resource Allocation Strategy. Jap 2001). discuss the limitations of our study. Specifically. ci. CLV is a function of the predicted contribution margin. the various CLV components can vary depending on the industry). This is also called the “lost-for-good” scenario (Rust. and identify areas for further research. and Ti = predicted number of purchases made by customer i until the end of the planning period. CLV Measurement and Maximization The various components of CLV include purchase frequency. t = time index. The lost-for-good approach is questionable because it systematically understates CLV (Rust.l (1 + r )l − 1 . October 2004 . we measure CLV by predicting the purchase pattern (purchase frequency or interpurchase times) over a reasonable period.m. r = discount rate for money (set at 15% annual rate in our study). Objective Function: CLV Typically. Some of the antecedents of purchase frequency and contribution margin (e. to predict future customer activity in noncontractual settings. and r = discount rate. This assumption is in accordance with theories about the different phases in a relationship and relationship life cycles (Dwyer. Finally. Previous researchers have used the variable P(Alive). In contractual settings. we use the always-a-share approach in this study. Given predictions of contribution margin. or the likelihood of a customer staying in or terminating a relationship.of other widely used CRM metrics. We then propose hypotheses about the influence of supplier-specific factors and customer characteristics on the various CLV components. However. and variable costs.. In addition to accurate measurement of CLV for each customer. contribution margin. (1 + r ) t where i = customer index. In the next section.” we provide details on allocating resources that maximize CLV. and Lemon 2004). can be increased by allocating marketing resources across channels of contact for each customer so as to maximize his or her respective CLVs. the CLV function we use can be represented as follows: (2) CLVi = y = 1 (1 ∑ Ti CM i. and thus profits. where CLVi = lifetime value of customer i. there is the potential for substantial improvement in profits when managers design resource allocation rules that maximize CLV. n = forecast horizon. CLV can be calculated as follows: (1) CLVi = tribution margin from each customer. managers are interested in predicting customer retention. the focus is more on predicting future customer activity because there is always a chance that the customer will purchase in the future. the propensity for a customer to continue in a relationship (customer retention).y = predicted contribution margin from customer i (computed from a contribution margin model) in purchase occasion y. In general. purchase frequency. the company treats the customer as a new customer and ignores its history with the customer. we develop the framework for the measurement and maximization of CLV. measured in dollars. This is also called the “always-ashare” scenario. Zeithaml. marketing communications) are under management’s control and affect the variable costs of managing customers. we derive implications based on the results. In the subsequent section. we compare the aggregate profits provided by high-CLV customers with those of customers who score high on several other customer-based metrics. which represents the probability that a customer is alive (and thus exhibits purchase activity) given his or her previous purchase behavior (Reinartz and Kumar 2000).l = unit marketing cost for customer i in channel m in year l (the formulation of CLV does not change if l is used to represent periods other than one year). we model the purchase frequency and ∑ n t=1 (Future contribution margin it − Future cost it ) .g. CMi. However. in this framework. and Lemon 2004). xi. we explain the models and data we used to estimate CLV. We then discuss the results from our analyses and explain the comparison of CLV with other metrics for customer selection.m. he or she does not return to the supplier. n = number of years to forecast. Another method for predicting future customer activity is to predict the frequency of a customer’s purchases given his or her previous purchases. Therefore. The assumption underlying this framework is that customers are most likely to reduce their frequency of purchase before terminating a relationship. and marketing costs (however. such a methodology enables a customer to return to the supplier after a temporary dormancy in a relationship.m.m.l × x i. Zeithaml. and Oh 1987. If a customer is won back after termination. We use these antecedents to maximize CLV.

We then use the customer responsiveness to marketing actions. Customer Lifetime Value Framework / 109 . although the purpose of standardized communication may be different for transactional customers than for relational ones. The overall theoretical framework that we used is provided in Figure 1. It is possible that a customer makes several purchases in a given year. Zeithaml. expected effects. and so on. direct mail serves to maintain commitment and trust by communicating relationship benefits (Morgan and Hunt 1994) and to inform the best customers about new product offerings. Discounting contribution margin. more cooperation. we raise the denominator in the cost function calculation to current year – 1 (i. to develop resource allocation strategies that maximize CLV. direct mail can be used in combination with telephone sales to generate interest in products while simultaneously improving the return on investment (Nash 1993). standardized (e. Lemon. as the basis for selecting antecedents to predict purchase frequency.m.g. Standardized modes are also the most cost-effective method for identifying customers who are interested in an organization’s current promotion (Shepard 2001). within each customer across channels). Discounting cost allocations. the cost allocation in the first year need not be discounted. and Lemon 2004) and channel communications (Grewal. and xi. Assume that it is currently year l = 1 and that we need to forecast the contribution margin from each customer for the next n years (i. we use the square root of the discount rate).. we classify channels of communication into the following contact modes: rich (e. Morgan and Hunt 1994. and Lemon (2004) provide guidelines for discounting contribution margin from customers when there is more than one purchase occasion (y) per year. the objective is to identify the resource allocation rules across various channels of communication for each individual customer such that the respective CLVs (as provided in Equation 2) are maximized. and less complexity in their relationships than in those of transactional customers (Morgan and Hunt 1994). In this approach. Supplier-Specific Factors: Channel Communications In this study. Therefore. therefore. Relational customers tend to have high commitment and trust with their suppliers. Frequency of rich and standardized modes. we are interested only in allocating resources across channels for each individual customer (i. Face-to-face communications and trading event meetings are the richest and most direct mode of communication possible among channel members (Mohr and Nevin 1990). therefore. In summary. Next. Berger and Nasr (1998..g. and Verhoef 2004. Therefore.. The discounting of cost allocations is straightforward if we assume that there is a yearly allocation of resources (as is the case in most organizations) and that the cost allocation occurs at the beginning of the year (the present period). However. Discussion of model constraints. The constraints ensure the nonnegativity of the predicted purchase frequency and communication levels for each customer i during period l. which results in less uncertainty. Rich modes of communication are also effective in converting transactional customers to relational ones (Ganesan 1994). For example.e. Mohr and Nevin 1990. telephone). Thus. gan and Hunt 1994) as well as previous research in customer equity and CLV (Bolton. our framework can be applied to allocate resources across channels with each individual customer and across customers in the presence of a budget constraint. m.1 Our objective function is subject to the following constraints: frequency > 0 ∀ i. The costs of communication in each channel can influence managers’ frequency of communication in each channel. Zeithaml. Rindfleisch and Heide 1997). We summarize the antecedents of purchase frequency. Although we expect the relationships between different channels of communication and predicted customer activity to be similar. l. we need to analyze customer responses separately across different channels because the costs of serving customers across different channels are different. and customers might exhibit different responsiveness across the various channels.5 (in other words. We first focus on the discounting of contribution margin over a period of time. y/frequency = . Bowman and Narayandas 2001. we do not use a budget constraint on the total resources available for contacting customers. and the rationale for our expectations in Table 2. we expect that the marginal response for increased frequency is the same across segments. l – 1)... Antecedents of Purchase Frequency and Contribution Margin Purchase Frequency An objective of relationship marketing is to ensure future purchase activity. The first purchase occasion (y = 1) occurs after 6 months. we provide a detailed discussion for a few of the hypotheses that are unique to our study. their operationalization. Equation 2) and Rust. consider when the planning horizon is one year and the frequency of purchases is two times (frequency = 2). Corner. Thus. obtained from the purchase frequency and contribution margin models. Direct mail and telephone communication are the most standardized and cost-effective modes of individual-level communication available to an organization. face-to-face. we use the commitment–trust theory of relationship marketing (Mor1In this study. the cost allocation in the second year needs to be discounted for one year. and Mehta 2001. Reinartz and Kumar 2003. Rich modes of communication are preferred to standardized modes when issues in the channel structure are complex and when there is a high degree of uncertainty in the relationship. t. For transactional customers. Therefore. until l + n). The second purchase occasion (y = 2) occurs after 12 months. Purchase frequency is also a component of our CLV calculation. Rust. y/frequency = 1.l ≥ 0 ∀ i. trading event meetings).e. and Web based (Mohr and Nevin 1990). For relational customers.e. direct mail. the discount rate from a customer is scaled according to his or her frequency of purchase (as is shown in Equation 2).contribution margin of customers as a function of marketing resource variables such as channel contact.

too much contact can overload buyers and have dysfunctional consequences (e. if they want the supplier to conduct training programs at the customer’s site (Cannon and Homburg 2001). To maximize the effect of each contact. too much communication may be dysfunctional. the marginal response to a higher level of rich modes of communication need not always be higher.. In addition. Although the utility of marketing contacts is not questioned. However. sometimes. In a B2B setting. On most occasions. ubiquitous junk mail). October 2004 . Thus: H1: An inverted U-shaped relationship exists between the frequency of rich and standardized modes of communication and a customer’s predicted purchase frequency. In addition. we expect that there exists an optimal level of intercontact time between suppliers and buyers. customers can initiate contacts with suppliers for several reasons. Customer Characteristics: Customer Involvement Bidirectional communication. the same is not necessarily true for B2B settings. Thus: H2: An inverted U-shaped relationship exists between intercontact time and a customer’s predicted purchase frequency. and Mick 1997). Research on channel communications shows that highly relational channel structures are associated with large bidirectional communications among channel members (Mohr and Nevin 1990).g. it even can be negative. Dobscha. indicates customer involvement. or if the supplier invites the customer to participate in new product development sessions. the marginal utility of an additional piece of information from a supplier firm in a short period is low. supplier firms need to pace their communication schedule to suit customer needs. Intercontact time. such as if they have new needs that the supplier may be able to fulfill. bidirectional communication in channels strengthens a relationship. Higher levels of previous communications lead to trust with the supplier and act as glue that holds together a communication channel (Morgan and Hunt 1994). Following the theory that leads to H1. it has been proposed that too much communication causes a relationship to be dysfunctional (Fournier. Although customer-initiated contacts are associated primarily with complaints in business-to-consumer settings. 110 / Journal of Marketing.FIGURE 1 A Conceptual Framework for Measuring and Using CLV Customer Characteristics: Switching Costs •Upgrading •Cross-buying Customer Characteristics: Involvement •Bidirectional communication •Number of returns •Number of Web-based contacts Customer Characteristics: Previous Behavior •Product category purchased Supplier-Specific Factors: Channel Communication •Level of rich modes •Level of standardized modes •Intercontact time Discount Rate Predicted Purchase Frequency Total Profit Net Present Value (Future Profits) = CLV Marketing Costs Generate allocation rules Customer Characteristics •Lagged contribution margin •Establishment size •Industry category •Total quantity of purchases Contribution Margin Allocate Appropriate Resources Supplier-Specific Factors •Total marketing communication However.

Morgan and Hunt 1994). there is an optimal communication level. but contacts that are too soon can cause dysfunction. and Expected Effects Variable Purchase Frequency Model Antecedents Upgrading Number of product purchase upgrades until an observed purchase + Customers who upgrade have higher switching costs with each upgrade. Corner. Reinartz and Kumar 2003). thus. which can lead to lower propensity to leave and higher recurrent needs (Bolton. and Mehta 2001. Covariates. and customers who want to create efficiencies are highly relational and have recurring needs (Grewal. + Customers who use online communication want transaction efficiencies. Two-way communication between parties strengthens the relationship and ensures that the focal firm is recalled when a need arises (Morgan and Hunt 1994). Customers who purchase across several product categories have higher switching costs and recurrent needs (Bowman and Narayandas 2001. Lemon. and Verhoef 2004). Acknowledgment of customers with relationship benefits reduces the propensity of customers to quit and increases the probability that the focal firm is recalled when a need arises (Morgan and Hunt 1994). Rindfleisch and Heide 1997). Nash 1993). but too many purchases can be detrimental to the relationship and can indicate that the firm has not used the return opportunities appropriately.TABLE 2 Antecedents. Timely communication between parties reduces the propensity of a customer to quit a relationship (Mohr and Nevin 1990. but too much communication can be detrimental to the relationship (Fournier. Returns provide an opportunity for firms to satisfy customers and ensure repeat purchases (Reinartz and Kumar 2003). Dobscha. Customer Lifetime Value Framework / 111 . Operationalization Expected Effect Rationale Cross-buying Number of different product categories a customer has purchased + Bidirectional communication Ratio of number of customer-initiated contacts to total number of customer contacts (customer initiated and supplier initiated) between two observed purchases Number of products the customer returns between two observed purchases + Returns Frequency of Web-based contacts Number of times in a month the customer contacts the supplier through the Internet between two observed purchases Relationship benefits Indicator variable of whether a customer is a premium service member (based on revenue contribution in the previous year) + Frequency of rich modes of communication Frequency of standardized modes of communication Intercontact time Number of customer contacts by the supplier in a month (through sales personnel) between two observed purchases Number of customer contacts by the supplier in a month (through telephone or direct mail) between two observed purchases Average time between two customer contacts by the supplier across all channels of communication between two observed purchases A long time between contacts can lead to forgetfulness. and Mick 1997.

We use size of an establishment and industry category as covariates in our model. we do not discuss the hypotheses in detail. we classify the antecedents of contribution margin as supplierspecific factors (total marketing efforts) and customer characteristics (lagged contribution margin and purchase quantity). we also provide the operationalization of the antecedents. Gupta. Modeling CLV and Data Model Development To predict CLV. we analyze Web-based contacts separately from other channels of communication because Web-based communication is customer initiated (i. Grewal. As with purchase frequency. and Narasimhan 2001). we need a stochastic model to predict each customer’s purchase frequency and a panel-data model that 112 / Journal of Marketing. In Table 2. In addition. there are several advantages to tracking Web-based contacts in a B2B setting. Corner. Notes: For the purchase frequency model. their expected effects. Contribution Margin The antecedents that we adopt to predict contribution margin are based on findings from previous research on antecedents of customer revenue (Niraj. Total number of customer contacts across all channels Total quantity of products the customer purchased across all product categories Number of employees in the customer firm Standard industrial classification–based industry category to which the customer firm belongs + + + Control variables that accommodate for customer heterogeneity (Niraj. we provide a description of the covariates we use as control variables in the purchase frequency and contribution margin models. the other indicates a software purchase Expected Effect Rationale A customer’s purchase patterns may depend on the product category purchased. Second. Thus: H3: The higher the level of bidirectional communication. and the rationale for our expectations based on previous research. Mohr and Nevin 1990). Sheth and Parvatiyar 1995). October 2004 . for the contribution margin model. Previous marketing communications and depth (quantity) of purchases positively affect contribution margin (Gupta 1988. the higher is a customer’s predicted purchase frequency. Webbased contacts from the buyers provide some important signals to the supplier about the buyer’s relationship orientation. the dependent variable is contribution margin. we provide a description of the antecedents that we propose influence customer purchase frequency and contribution margin. Thus: H4: The higher the number of Web-based contacts from a customer. Contribution Margin Model Antecedents Lagged contribution margin Total marketing efforts Total quantity purchased Covariates Size of firm Industry category Customer’s contribution margin from the previous year + Previous revenue is a good predictor of current revenue and accounts for any model misspecification (Niraj. Tellis and Zufryden 1995). a passive mode of operation for the supplier). In Table 2. First. participation in electronic markets (or use of Web-based initiatives) improves transaction effectiveness and efficiency (Rindfleisch and Heide 1997). Efficiency of communication and transactions among channel members is associated with a relational structure and higher customer involvement (Mohr and Nevin 1990. Because all the antecedents we use in our contribution margin model are based on previous research and findings. Frequency of Web-based contacts.. and Narasimhan 2001) and purchase quantity (Gupta 1988. In the database. However. Web-based communication between buyers and suppliers is the most cost-effective method of communication. the higher is the customer’s predicted frequency of purchase. and Mehta (2001) find that organizations enter and actively participate in electronic markets if their motivation is to improve efficiency in transactions.e. the dependent variable is purchase frequency.TABLE 2 Continued Variable Covariate Product category Operationalization Two indicator variables: one indicates a hardware purchase. and increases interdependence among channel members (Ganesan 1994. Gupta. In this study. and Narasimhan 2001). Tellis and Zufryden 1995). Gupta.

t – 1 = total number of contacts made to customer i in period t – 1.t – 1 = firm size for customer i in period t – 1. the probability of the jth purchase for customer i occurring at period t. Sk ( t ij |α k . the expected time until next purchase is given as follows: (4) where ∆CMi.e. Γ (α k )   The ratio of 12 (because we use months as the unit of analysis) to the expected time until next purchase (which we obtain by modeling a generalized gamma distribution on the interpurchase times.t – 1 + β3Sizei. λ ik . Φijk = the probability of observation j for the ith customer belonging to subgroup k.e.t – 1 + ei. Purchase Frequency We model a customer’s purchase frequency using the generalized gamma model of interpurchase timing that Allenby. γ k ) c ij bility of a customer belonging to each subgroup Φik as a probit function of the antecedents and covariates of purchase frequency. The parameters α and γ establish the shape of the interpurchase time distribution. where xij represents the antecedents and covariates of purchase frequency for customer i in purchase occasion j. To address the issue of endogeneity. γ) = the survival function for the generalized gamma distribution (i. λi. and we did not find any difference in the substantive conclusions. given α. we also use the log of the lagged interpurchase. given α. Specifically. However. given that the jth purchase has not occurred until time t. We use the growth in contribution margin from period t – 1 to t as the dependent variable and the contribution margin in period t – 2 as an independent variable. to account for any extraneous factors. S(tij|α. Leone. in our product category. λi.t – 1 + ΣjβjIndustryj + β4Totmarki. the endogeneity in formulation can be alleviated with a one-period difference in the dependent variable and a two-period lagged dependent variable as an independent variable (Baltagi 1998). Thus. λi.. We assume that the population consists of k subgroups. Sizei. as is shown in the work of Allenby. and Jen (1999) developed. and Jen [1999]) gives the predicted purchase frequency. and cij = 0 if the jth interpurchase time for the ith customer is right-censored. γ = the parameters of the generalized gamma distribution.. γ k ) (1 − c ij ) . λi. cij = the censoring indicator. Because we use a generalized gamma distribution to model interpurchase time and the likelihood function in Equation 3.t. we find that the correlation between purchase frequency and contribution margin is not significant. resembles a hierarchical Bayes formulation of the concomitant continuous mixture model. we represent the link function as Φik = f(xijβi). lagged total quantity purchased. The log of the lagged interpurchase time achieved this objective in scaling the tail of the lagged interpurchase time distribution.t = difference in contribution margin from period t – 1 to period t for customer i. and Jen 1999). In this study. and lagged total marketing efforts. This is a rather restrictive assumption for frequently bought consumer goods (Tellis and Zufryden 1995). the probability of the jth purchase for customer i occurring at a given period is greater than t. Our model framework. Leone. and Φik provides the mass point (i. γ) = the density function for the generalized gamma distribution (i.e. where f(tij|α. The likelihood function for the purchase frequency model is given as follows: (3) L= ∏ ∏ ∑ Φ f ( t |α .t = β0 + β1CMi. λ ijk k ij k i = 1 j = 1k = 1 n Ji K ik . The generalized gamma model also accommodates the commonly used exponential distribution for interpurchase times (Reinartz and Kumar 2003). In panel-data models with lagged dependent variables. Leone. measured as number of employees.. measured in dollars. The independent variables in the contribution margin model are thus lagged contribution margin. The other independent variables we used are specific to period t – 1 (this also accommodates the issue of endogeneity). Customer Lifetime Value Framework / 113 . γ). and λi is the individualspecific purchase rate parameter.predicts contribution margin. We model the proba- 2We also used lagged interpurchase time instead of the log of the lagged interpurchase time. we use the one-period lagged value for all the antecedents and covariates in our analysis (Villas-Boas and Winer 1999). Industry = indicator variable for industry category of the customer firm. We needed to address endogeneity issues while using lagged contribution margin as an independent variable in our model. and α. We used log of the lagged interpurchase times because lagged interpurchase times can have a threshold effect on the influence of current interpurchase times (Allenby. industry category. as presented in Equation 3. the contribution margin model is (5) ∆CMi. ∑Φ k ik   1   × Γ  α k +  γk       × λ ik . γ). we assume that the amount a customer spends is independent of purchase timing. Totmarki. and βi represents the customer-specific response coefficients. weight) for each subgroup.t – 2 + β2Quantityi. lagged firm size. However. Contribution Margin We model the contribution margin from a customer using panel-data regression methodologies. where cij = 1 if the jth interpurchase time for the ith customer is not right-censored.2 The specification of the model enables us to estimate individual customer-level coefficients for the influence of the covariates on the probability of a customer belonging to a particular subgroup and thus interpurchase times. λi.

for observed purchase j. the total number of contacts made to the customer in 1999. for Observation 1 the dependent variable is the difference in contribution margin between 2000 and 1999. and personal computers) and software (integration and application) manufacturer for the empirical application of our framework. The results from the purchase frequency model for Cohorts 1 and 2 are provided in Table 4 (for details on the model estimation. We also restricted our sample to customers who had made at least five purchases. We used the 2001 data as a holdout sample and to compare strategies. for these product categories. the current effects antecedents and covariates represent the customer’s (or supplier’s) activity between observed purchases j – 2 and j – 1. We calculated the current-effects antecedents and covariates on the basis of the activities of the customer or the supplier (in the case of channel communications) between the previous observed purchase (j – 1) and the current observed purchase (j). workstations. relationship benefits. All the antecedents and covariates we used in our analyses are lagged variables. the industry category of the customer. the dependent variable is the difference in contribution margin between 1999 and 1998. We assigned customers to Cohort 1 (Cohort 2) if their first purchase with the manufacturer occurred in the first quarter of 1997 (first quarter of 1998). and theoretically the customer retention phase begins after the first purchase. respectively. Even though the firm’s products are durable goods. We used each observed purchase for a customer as an observation in the purchase frequency model. The current-effects antecedents include bidirectional communication. There are several insights that we derive from Table 4. To model contribution margin from a customer. it is the choice of the buyer and seller to develop their relationships. The coefficients of the antecedents reported in Table 4 are the means from the posterior samples of βi. we used two cohorts of customers: Cohorts 1 and 2. Contribution margin model. for each customer. and t = index for time. which we discuss subsequently. Model fit. which provides the variance required in modeling customer response.5 to 23 months. To assess the inverted U-shaped 114 / Journal of Marketing. For Cohort 1. for Cohort 2. i = index for the customer. there are two observations per customer in our analysis sample. and Web-based contacts. the cumulative-effects antecedents represent the customer’s activity since relationship initiation until observed purchase j – 1.t – 1 = total quantity of products bought by customer i in period t – 1. Similarly. The results show that the generalized gamma model with two subgroups provides a good fit to the data . Results from Estimation of CLV Purchase Frequency Model As we discussed previously. The descriptive statistics for the data and the correlation matrix of the antecedents are provided in Table 3. we removed customers who had missing values for either rich or standardized modes of communication. we used a quadratic conversion (including the square covariate term in Equation 3) of the respective antecedent. and there are significant benefits for both parties to maintain a long-standing relationship. The company’s database focuses on business customers. Specifically. model performance. we used the annual sales from various purchases of each customer. we removed 20% of the original cohort of customers for our analyses. and intercontact time. The covariates in the purchase frequency model (type of product purchased) can be classified as current effects. and their values represent the total number of different products (for cross-buying) or upgrades that the customer has purchased since the first purchase until the current observed purchase. model selection. As we stated previously. standardized. and support for hypotheses. we used the 2001 data as a holdout sample and to compare strategies. and the total quantity of products purchased in 1999. we used the interpurchase times until the end of 2000 as our calibration sample. The cumulative effects antecedents include crossbuying and upgrading. and the independent variables include the contribution margin in 1997. Data We used data from a large multinational computer hardware (servers. The interpurchase time for customers in Cohort 1 ranges from 1. we selected customers who made their first purchase in the first quarter of 1997. which resulted in an effective sample size of 1316 and 873 observations for customers in Cohorts 1 and 2. frequency of rich. Given our model structure in Equation 5. ei. Specifically. October 2004 relationships. the firm size in 1999. for our analyses. for observed purchase j.Quantityi.t = error term. The antecedents and covariates we used can be classified as cumulative and current effects. they require constant maintenance and frequent upgrades. We discuss the results from Cohort 1 in detail in the text. it ranges from 1 month to 20 months. For all customers. We censored our data set in 2000 and used the 2001 data as our validation (or holdout) sample in both cohorts. we used the annual sales from each customer from 1997 to 2000. for Observation 2. we omitted the first observed purchase in our analysis sample because the first observed purchase is restricted to be within three months for all customers in the cohort. The product categories in the database represent different areas of hightechnology products. For our analyses. In addition. The results from Cohort 2 are quite similar to those of Cohort 1 and are provided along with the results for Cohort 1 in the corresponding tables. Similarly. Purchase frequency model. We estimated the purchase frequency model in Equation 3 using Markov chain Monte Carlo (MCMC) methods. and the independent variables include the contribution in 1998. respectively. we used an effective sample size of 1316 and 873 observations that belong to Cohort 1 (first purchase in 1997) and Cohort 2 (first purchase in 1998). and the signs for the coefficients represent their influence on a customer’s purchase frequency. returns. For each customer. Overall. see Appendix A). The choice of vendors for the products is normally made after much deliberation by the buyer firm. In our samples. For customers in Cohort 1.

8) .34* .7 (2.58 (3.15** .79 (1.06 1 .19* 1 .22 –.88 (4.8 (14.15 .08 .29 (.05 .52) 20.68*** 1 .40*** .7 (1.25 1 Customer Lifetime Value Framework / 115 1.12) .11* .74 (22.13) .37) 1 .16) 2.06 –.69 (5.2) .34* .TABLE 3 Descriptive Statistics and Correlation Matrix Relationship Benefit– Premium Service Level Variable Mean Standard Deviation 8.91 (.5) Upgrading Cross-buying Bidirectional communication Returns Frequency of Webbased contacts Relationship benefit– premium service level Frequency of rich modes Frequency of standardized modes Intercontact time (days) 1.84) 47.1) 13.75 (45.01 –.81) .30** 1 15.84 (.36** .09 .81) 5.52) 3.5) 2.44*** .48* 1 .41*** .21 .09 (.45*** .03 –.81) 25.06 1 .32 (1.86) 3.39* .22* .11 .13 .04 –.4 (6.53*** .36*** .07 .94) Purchase Frequency Upgrading 1 .32* .89 (.01 –.62*** .25) .3 (16.51*** –.17** 1 .31** .81 (24.51* CrossBuying Bidirectional Communication Returns Frequency of WebBased Contacts Frequency of Rich Modes Frequency of Standardized Modes Intercontact Time (Days) Purchase Frequency Model Purchase 5 frequency (4.91) 1.07 .62) .41 (3.29** .32* .21* .24** .

and Web-based contacts represents the frequency of the respective mode of communication between two consecutive purchases.5) 1.297) 323.08** . October 2004 TABLE 3 Continued Variable Contribution Margin Model Growth in contribution margin Lagged contribution margin Total marketing efforts Total quantity purchased Mean 4955 (4827) 31.00 .00** . Notes: The frequency of rich.78*** .00*** –.867) 17.89 (2. Values in parentheses for the descriptive statistics represent Cohort 2. .00 .143 (32.139 (318. ***Significant at α < 1%.02) Growth in Contribution Margin 1.825) 3.61*** . The correlation matrix for Cohort 2 is similar to that of Cohort 1 and is available on request from the authors.05** 1.40 (14.270 (381. We do not report the correlation matrix of the covariates because it does not have any substantive interpretation. standardized.07 Lagged Contribution Margin Total Marketing Efforts Total QuantityPurchase 1 *Significant at α = 10%.49 (4. **Significant at α = 5%.24) 13.116 / Journal of Marketing.26 (16.71*** 1.01) Standard Deviation 417.

Distribution parameters.01)** 10.012)** .53)** 9.30 (–1.002)** 1.49 (1.52 (9. Relative absolute error is with respect to a moving average model.05)** –2237.02)** –.48 (3.46 (.85)** 5.66 (–3297.007 (.55) .02)** 1.53) *Posterior sample values between the 2.2 (1.62)** 6.91) –4484 (–4884) –4683 (–5083) .98)** –3.5) .05 (48.TABLE 4 Coefficients for the Generalized Gamma Purchase Frequency Models Model 1 (Generalized Gamma Without Mixture) 3. and thus we do not include it here.82.9) . and is better than other models for modeling purchase frequency (log-likelihood for Model 3 = –2237.89 (–1.5th and 97.28)** .90)** .1 (4.5th percentile do not contain zero.01 (5.56) 1.95 (.08 (6. **Posterior sample values between the .21 (–4.65)** Variable Component 1 α1 υ1 θ1 γ1 Mass point Component 2 α2 υ2 θ2 γ2 Mass point Coefficients β01 Lagged log interpurchase time Antecedents: Customer Characteristics Upgrading Cross-buying Bidirectional communication Returns (Returns)2 Frequency of Web-based contacts Antecedents: Supplier-Specific Factors Relationship benefits Frequency of rich modes of communication (Frequency of rich modes of communication)2 Frequency of standardized modes of communication (Frequency of standardized modes of communication)2 Intercontact time (Intercontact time)2 Log-likelihood AIC BIC Relative absolute error –3298.91 (–3627.58 (1.45) .82)** .52)** Model 3 (Generalized Gamma with Mixture and Temporal Variation) 3. The significance levels apply to the coefficients of Cohorts 1 and 2.5th percentile do not contain zero.2 purchases in a year.80) –1.5th percentile and 99.18 (25. We also used a hazard model with the finite mixture framework (Kamakura and Russell 1989) to model purchase frequency.005 (.08)** .05 (4.91) –2908.38)** (7.2 (1.54 (.04 (.51 (.92)** 1.46 (.84 (–4.59)** 47.3 The mean expected purchase frequency for Subgroup 1 is 4.97 (43. Akaike information criterion [AIC] = –4484.07 (49. Customer Lifetime Value Framework / 117 .05 (–2.9 (.64 (8.50 (2. Notes: Values in parentheses represent Cohort 2.04) –5825 (–6602) –5910 (–6687) .53 (7.08)** 1.28)** 6.28 (–.65)** (5.08) –6606 (–7262) –6639 (–7295) .01)* 3.02)** .54 (.52 9.84 (. and Bayesian information criterion [BIC] = –4683).44) –2.61)** 32.008 (.51)* –2.21 (34.9) . and we found that our proposed model fits the data better and has better predictive capabilities.2)** 42. We adopted such a procedure to ensure that the parameter estimates are global optimal values and are not affected by any local maxima.2 Model 2 (Generalized Gamma with Mixture.12)** 26.56)** –3.3) . Without Temporal Variation) 4.57)** 38. The product category variable was not significant in our analysis.09 (–2. and the 3The parameter estimates of the purchase frequency model are based on mean values from 50 repeats with random starting values for each repetition.78 4.001)** .27 (3.9 (.82 (–2437.01 (.

) (. and semiannual levels. communication from a customer can be a good indicator of a customer’s activity.S.S. we can infer that the rate of diminishing returns (after exceeding the threshold) is much higher for the rich mode of communications than for standardized modes. Customer Selection Strategy In this section. we used the contribution margin model with the annual data.S.) (. but we did not find any as a holdout sample for Cohorts 1 and 2. a simple metric. we compare the customer selection capabilities of the following: CLV.S. Notes: The reported coefficients are standardized estimates. 118 / Journal of Marketing. Thus.54 and . We find that the frequency of contacts affects purchase frequency nonlinearly. Customer characteristics. Given the variation in expected purchase frequencies in each subgroup. firms should use it with great caution. = not significant. In addition. respectively. We derive several other insights from Table 5.03** . A firm’s increasing communication beyond a certain threshold may result in diminishing returns in terms of customer purchase frequency. The two-period lagged contribution margin provides the highest contribution to an explanation of current-period contribution margin.03***) (–. However. The coefficient of the quadratic term for rich modes of communication (–1.30 for Cohort 1) is higher than the coefficient of the quadratic term for standardized modes of communication (–. who also find that breadth of purchase positively affects a customer’s duration in a profitable relationship.02**) (. which we discuss subsequently. Contribution Margin Model We estimated the contribution margin model with annual data from 1997 (t – 4) to 2000 (t) for Cohort 1 and from 1998 to 2000 for Cohort 2. consumer packaged goods.01 purchases in a year. the higher is the customer’s purchase frequency.03*** N. We find support for an inverted U-shaped relationship between purchase frequency and returns. to maintain simplicity.46. firms in the education industry provide a lower contribution margin than do firms in other industries. and government industry categories provide. previousperiod customer revenue (PCR). We also find that the higher the bidirectional communication between the customer and supplier.S. Thus. our proposed metric. Thus. **Significant at α = .05. This leads us to believe that there is an optimal level of marketing communication for each customer.mean expected purchase frequency for Subgroup 2 is 1.02* .68 and thus can explain significant variation in contribution margin from customers. which is widely considered a good significant differences in model performance. quarterly. Specifically. Therefore. The coefficients of the contribution margin model are provided in Table 5.02* N.) (.03*** –. The results indicate that upgrading and cross-buying positively influence a customer’s purchase frequency. lagged quantity of goods is significant in explaining variation in contribution margin. Supplier-specific factors. our analysis suggests that managers need to exercise caution. Firm size and industry category explain variation in current-period contribution margin.83*** . Our analyses indicate that a supplier’s contact strategy and provision of relationship benefits significantly affect a customer’s predicted purchase frequency. TABLE 5 Regression Results from the Contribution Margin Model Independent Variable Intercept Contribution in t – 2 Lagged total quantity purchased Size Aerospace Financial services Manufacturing Technology Consumer packaged goods Education. . With respect to returns from a customer. ***Significant at α < . October 2004 .S. This finding also provides the reasoning to determine the optimal level of resources that needs to be allocated across channels to maximize CLV in Phase 2.” The component masses for Subgroup 1 (ϕ1) and Subgroup 2 (ϕ2) are . Total lagged marketing efforts contribute significantly to an explanation of variation in current contribution margin.10. The main insight from Table 5 is that the contribution margin model provides an adjusted R2 of . The results highlight the importance of firms’ recognizing the customers who establish contact through the online channel in their CRM strategies. we find an inverted U-shaped relationship.02*) (.) (.02** N. Customer characteristics. although the rich mode of communication is interactive and effective.S. we term Subgroup 1 the “active state” and Subgroup 2 the “inactive state. The coefficients of the marketing contacts reveal a difference in the influence of various channels on customer purchase frequency.4 The revenue in 2001 (t + 1) acts 4We also estimated the contribution margin model at the monthly.02**) (N. . This is in line with the findings of Reinartz and Kumar (2003).03**) (. N.28 for Cohort 1).02** . firms in the financial services. This indicates that customers who return products within a threshold are a good asset to the firm.85***) (.02***) (N. Supplier-specific factors. past customer value (PCV). a higher contribution margin than do firms in other industry categories.S. . technology. in addition to timely communication from the supplier to the customer (Morgan and Hunt 1994). the values in parentheses represent Cohort 2.S. This implies that we expect 54% of the customers to be active in the prediction window and 46% of the customers to be inactive in the prediction window.06***) *Significant at α = . . Among the various industry categories. which implies that a supplier’s contact strategy affects both purchase frequency and contribution margin.01. on average.02**) (N.04*** (N. K–12 Travel Government Lagged total level of marketing effort Parameter Estimate N.

038 15.782 CLD 107. and profits for the top 5%. Cohort 2 provides similar results. to reflect industry practice.929. CLD. gross profit = revenue × .735 000. whereas the average net profits are $70. Table 6 shows that the proposed metric better identifies profitable customers than do other metrics we compared in the study. Selection of more customers to contact may not be feasible because of limited time and resources. as do Reinartz and Kumar (2003). and CLD. PCV. and CLD. Gross profit is residual revenue after removing cost of goods sold.785 72.610 55. rank-ordered on the basis of a scoring metric.929 27. and the reported values are cell medians. The results were similar to the comparison with PCR.295 (after accounting for cost of goods sold [70%] and variable costs of communication).389 for the top 5% of the customers selected on the basis of PCR.979 70.908 00.686 000.588 143. approximately $40. The results are provided in Table 6.225 Notes: All metrics are evaluated at 30 months.389 55. we now describe our methodology for designing resource allocation strategies that maximize CLV. and CLD. The reported values are cell medians.719 000. respectively. We subsequently summarize the results from our comparison.predictor of future customer value. On the basis of the 18-month prediction window. and CLV is.) In general. $130. Performance of the Customer Selection Metrics To compare the performance of the four metrics.147 00. 10%. We used the data from the first 30 months to score and sort the customers on each metric.591 000. The marketing literature has provided guidelines for optimal resource allocation in acquisition and retention decisions (Blattberg and Deighton 1996. PCV. for the firm that provided the database. the cost of goods sold is approximately 70%.5 To select customers for contact. and marketing actions when future brand switching 5We also compared the metrics with a censoring time at 18 months and prediction window of 30 months. and 15% of customers. Descriptions of the various customer-based selection metrics used in our analyses are provided in Appendix B. with an 18-month prediction window.950 130. frequency. and profits from the top 5%.883 1. we expect the average net profits of customers selected from the top 5% using the proposed CLV metric to be $143. such as PCR. costs.3. Blattberg.943 27. we rankordered customers from best to worst according to each metric and then compared the sales. such as share of wallet and recency.227 44.000 for a customer in the top 5% sample. Customer Lifetime Value Framework / 119 .170 PCV 131. Although the difference in profits from use of PCR. and $106. Thus. and customer lifetime duration (CLD). Overall. thus. 10%. The results are available on request from the authors. on average. the difference in total profit across the top 5% to 15% of the entire customer base can easily yield more than $1 million. so as to maximize CLV. PCV. in general organizations choose the top 5% to 15% of their customers. Resource Allocation Strategy Having evaluated the usefulness of using CLV for customer selection. and Thomas 2001).245 77.295 78. (We also compared the customer selection capabilities of CLV with other customer-based metrics. In general.785. The results provide substantial support for incorporating the responsiveness of each individual customer across various communication channels and for the usefulness of CLV as a metric for customer scoring and customer selection. TABLE 6 Comparisons of CRM Metrics for Customer Selection Percentage of Cohort (Selected from Top) 5% Gross profit ($) Variable costs ($) Net profit ($) 10% Gross profit ($) Variable costs ($) Net profit ($) 15% Gross profit ($) Variable costs ($) Net profit ($) CLV 144. Getz.401 1. 10%. promotion expenditures (Berger and Bechwati 2001. The substantive results of the study hold even for a 30-month prediction window. and 15% of customers. and monetary value.340 PCR 71.963 00.738 44. a forward-looking metric that is also used as a proxy for loyalty. and 15% of customers from the censoring period (30 months) until the end of the observation window (48 months). Berger and Nasr 1998).114 00. PCV.809 51.156 56.981 00. These findings hold across all the percentage subgroups.790 106.892 52.837 00.794 71. We then compared the actual sales.944 14. we compared the performance of our metric among the top 5%. The framework also provides managers a tool for assessing return on marketing investments by identifying avenues for optimal resource allocation across channels of communication for each individual customer (and possibly across customers). variable costs of communication. organizations in direct marketing situations rank-order their customers on the basis of a particular metric and prioritize their resources from best customers to worst customers on the basis of the rank order (Roberts and Berger 1999).807 55.

we held the coefficients constant and identified the level of covariates (i. Equation 2 is a function of predicted purchase frequency (based on Equation 4). The theoretical implications of the purchase frequency model are also related to the CUSAMS customer 120 / Journal of Marketing. it is possible to improve profits and return on marketing communication by appropriately identifying customers for target communications and by matching the channel of communication with customer preferences. Thus.188. Implications Antecedents of purchase frequency and contribution margin. We first estimated the responsiveness of customers (coefficients. Therefore. based on this status quo resource allocation strategy. These issues made our optimization problem extremely complex and intractable with traditional analytical methods. In addition. we intended to allocate resources for each customer on the basis of individual responsiveness. such benefits may not be realized immediately because the firm also needs to incur costs to move toward a customer-centric view and to train its employees to manage customers on the basis of CLV. based on the resource allocation strategy that maximizes CLV. predicted contribution margin (based on Equation 5). and Lemon 2004). βs) to marketing contacts from the purchase frequency model and the contribution margin model. the contacts made to a customer across various channels are under the supplier’s control and thus can be used to maximize CLV. we demonstrated the superiority of selecting customers for contact on the basis of CLV compared with commonly used metrics such as PCR. These guidelines represent a significant step toward incorporation of long-term customer profitability effects into firm-level managerial decision making. and they do not provide a mechanism for dynamic updating of profitability assessment (Libai. First. there is a potential for the firm to increase its revenue by at least $1 billion across its entire customer base. We also computed the total net present value of future profits when the organization uses its current resource allocation strategy. Aggregate Results The total net present value of future profits from a resource allocation strategy that maximizes CLV (with the predicted contribution margin) is approximately $44 million. the level of contacts in each channel for each customer is varied each year. is approximately $1 million. 6In Equation 2. support for a purchase frequency distribution with two subgroups led us to believe that the optimization surface is multimodal. for each customer. based on its current strategy.. and Humby 2002). the total net present value of future profits is approximately $24 million. Thus. However. Our assumption can be viewed as taking the average level of contacts in the prediction periods. The total cost of communication in the organization’s current strategy is approximately $716. in Equation 2. In addition. we developed and estimated an individual customer-level objective function. Genetic algorithms (Goldberg 1989) are simulationbased. We find that the organization improves profits by increasing costs of serving customers (rather than cost of communication in the previous year) by 48%. However. The aggregate results suggest that given the improvement of approximately $20 million among a sample of 216 customers. With a communication strategy that maximizes CLV. Appendix C explains how we used a parallelsearch technique for resource allocation purposes. However. We also provide an outline for future researchers to build on the framework proposed herein. Second. which is not susceptible to local minima (common in multimodal distributions) (Venkatesan.6 In other words. We used a genetic algorithm to derive the levels of contact desired for each individual customer that maximize CLV. Krishnan.is considered (Rust. Limitations. is approximately $34 million ($24 million/$716. the return on marketing communication to the organization is approximately $44 million ($44 million/$1 million). with respect to marketing resources for a firm. we find that a resource allocation strategy that maximizes CLV results in an increase in profits by approximately 83%. In our study. the multimodal nature of the optimization surface (given the support for a mixture distribution for purchase frequency) motivated us to use a parallel-search technique. the customer contact levels across different channels appear in the revenue and cost sides of Equation 2 and thus avoid the scope of corner solutions. Narayandas. Implications. and Kumar 2004). and the purpose of the optimization is to find the level of contacts across various channels with each individual customer that would maximize CLV. and CLD. we evaluated the benefits of designing marketing communications that maximize CLV.188). To design a CLV-based resource allocation strategy. We find that. The total cost of communication (over three years). we resorted to a search algorithm to find the optimal resource allocation levels. level of channel contacts) for each customer that would maximize CLV. the goal of which is to measure CLV. In summary. we assumed that the level of contacts is equal across the prediction period. We now discuss the implications of our study and how managers can use this knowledge to design efficient marketing programs. The return on marketing communication to the organization. Zeithaml. we maintained the resource allocation levels for the most recent year and calculated CLV over a three-year period. Third. PCV. and marketing costs. the models provide less insight into decisions about how to manage individual customers in a way that accounts for the heterogeneity. parallel-search algorithms that have been used in econometrics (Dorsey and Mayer 1995.e. Liang and Wong 2001) to obtain optimal solutions when the complexity of the optimization function tends to be intractable and multidimensional. Specifically. and Further Research The objective of our study was to analyze the usefulness of CLV as a metric for customer selection and resource allocation strategy. to simplify our optimization routine. October 2004 . depending on the cost of each mode of communication and the responsiveness of the customer (in terms of both purchase frequency and contribution margin) to each channel of communication. Our resource allocation algorithm uses Equation 2 as the objective function.

both of which ultimately influence CLV. customers who return products within a certain period may do so because they have inherent trust in the supplier and because they expect future benefits. firms can further improve the efficiency of communication strategy by appropriately sequencing their customer contacts across different channels. We find an inverted U-shaped relationship between returns and purchase frequency. To summarize. satisfaction. in our study. Thus. we tested parts of the CUSAMS framework and found empirical support for the parts we tested. we provide a framework for maximizing CLV with marketing communications. which ultimately influences CLV. marketing communications influence the expected contribution margin from a customer. In addition. reward programs) affect a customer’s price perceptions. Dobscha.. Further research can investigate whether the customers selected for high levels of marketing communications are the same when the resource allocation strategy focuses on maximizing CLV or on maximizing incremental CLV. However. Further studies need to investigate whether the results are generalizable to other industries and settings. We find that in addition to influencing purchase frequency. This supports Fournier. Our analyses indicate that CLV is preferred to incorporate the dynamics of customer purchase behavior into the customer selection process. However. further research needs to develop models that combine forecasts of aggregate competitive responses to marketing actions and customer brand switching with individual-level models of direct marketing. and it provides accountability for strategies geared toward managing customer assets. In addition. In addition. it is possible that customers that previously had high CLV continue to have high CLV in the future. and CLD. The results of this study are from a customer database in the high-technology industry. Our analyses suggest that there is a potential for substantial improvement in CLV through appropriate design of marketing contacts across various channels. When firms design resource allocation rules. with an appropriately designed marketing strategy. and Mick’s (1997) expectations that too much communication between suppliers and customers can be disruptive. direct mailings. However. a customer’s returning too many times may indicate erosion of trust with the firm or a lower level of future activity. Rust. Limitations and Further Research The study has limitations that further studies can address. The proposed resource allocation strategy can be a basis for evaluating the potential benefits of CRM implementations in organizations. especially through rich modes of communication. For example. Resource allocation strategy. Managers can substantially improve their return on marketing investments by using a dynamic. Our analyses of customer selection investigate how firms can enhance strategies. We also find that breadth of purchases (cross-buying) and depth of buying (upgrading) affect purchase frequency. In addition. We find empirical support for marketing instruments’ effects on purchase frequency (rich and standardized modes of communication and relationship benefits) and contribution margin (total marketing efforts). Also in the B2B scenario. the online channel can provide an ideal setting for B2B firms to enhance their customer relationships. customer-level measure of CLV for scor- ing rather than using the other metrics suggested in the literature and by prioritizing contact programs.g. the sum of optimal CLVs for each individual customer need not lead to the optimal customer equity in Customer Lifetime Value Framework / 121 . and Lemon (2004) propose that firm strategies and tactical marketing actions affect the marketing productivity chain. and commitment. In this study. irrespective of the level of marketing communications. industry category and size seem to be important factors that influence the magnitude of contribution margin. Customer selection. We also consider only the average levels of optimal communication strategy in a channel. our results indicate that managers need to be cautious when designing marketing communication strategies across different channels and need to be wary of contacting customers too many times. and CLD. Zeithaml. We also find that customers that establish contact through the online channel of communication exhibit high frequency of purchases and have high involvement.asset management framework (Bolton. A CLV metric better identifies customers that provide higher future profits than do PCR. and that some customers that have had low CLV transform to high-CLV customers. which then ultimately influence CLV. •CLV performs better than other commonly used customerbased metrics for customer selection such as PCR. note that we do not compare our proposed resource allocation strategy (that focuses on maximizing CLV) with a strategy that focuses on allocating resources to customers for which the increment in CLV from appropriate design of marketing communication is highest. A possible benefit from customers who return products could be the opportunity to understand the reasons for dissatisfaction. Lemon. these affect the length. Therefore. Enhancement of marketing productivity. PCV. Specifically. and breadth of a relationship. Finally. The results from our study highlight the importance of firms’ considering individual customers’ responsiveness to marketing communication as well as the costs involved across various channels of communication when making resource allocation decisions. and Verhoef 2004). depth. PCV. such as improvements in the quality of the product. the major conclusions that we derive from our study are the following: •Marketing communication across various channels affects CLV nonlinearly. In turn. and our analysis of optimal resource allocation investigates how firms can improve tactical marketing actions. and •Managers can improve profits by designing marketing communications that maximize CLV. they can realize the increase in profits by incorporating the differences in individual customer responsiveness to various channels of communication and the potential value provided by the customer. the CUSAMS framework proposes that marketing instruments (e. we find support for a nonlinear relationship between supplier communications and purchase frequency.

We treated several antecedents used in our framework (e. Appendix A Results from Data Analyses Purchase Frequency Model Model estimation. We find that Model 3 has the highest value for both AIC and BIC (AIC = –4484 for Model 3. the value that a customer provides to a firm is a function of both the expected time frame until the next purchase and the contribution margin at that particular period. number of returns. –5910 for Model 2. a higher value indicates a better model. and –6639 for Model 1). Our empirical analyses support all the proposed effects of the antecedents on purchase frequency. Values of √ R that are closer to 1 reveal that all the chains have converged to true posterior distribution. Given the variation in expected frequencies in each subgroup. we used the initial 40. We also used multiple prior values for the estimates and did not find a significant impact on the posterior samples. The autocorrelation functions revealed that every fifth sample is unrelated to every other fifth sample. which we then included in a net present value function (Equation 1) that includes both costs and margins. We used the lagged variables of these to account for potential endogeneity. νk.2 purchases in a year.54 and . We simulated the posterior distribution using five parallel chains with overdispersed starting values.. the loglikelihood for the generalized gamma model with the probit link function (with two subgroups) and time-varying covariates (Model 3) is –2237. Overall. October 2004 . a latent-class finite mixture model with two segments provides a likelihood of –2938. We assessed the convergence of the algorithm from the line plots of the posterior sample and by evaluating Gelman and Rubin’s (1992) √ R statistic. We performed our optimization with a budget constraint. Increasing the number of subgroups in Models 2 or 3 did not result in significantly higher log-likelihoods.000 samples of the MCMC algorithm (for details of the algorithm. respectively. In this case. In our analyses. upgrading.000 samples). we found that the value of √ R ranged from 1.46. In addition. and the mean expected frequency for Subgroup 2 is 1. cross-buying. respectively. –5825 for Model 2. Table 4 shows that a simple generalized gamma model without the probit link function or timevarying covariates (Model 1) provides a log-likelihood of –3298. we assessed the indirect effect of the covariates on both costs and margins. the posterior statistics are based on 2000 samples (using every fifth sample from 10. We then used the mean posterior values to compute the expected time until next 122 / Journal of Marketing. we term Subgroup 1 the “active state” and Subgroup 2 the “inactive state. number of Web-based contacts) as exogenous variables in our analysis. All the distribution parameters (αk.” Influence of antecedents and covariates. Our approach to investigating within-chain autocorrelations.82. these values have been found to indicate convergence of the MCMC chains). The component masses for Subgroup 1 (ϕ1) and Subgroup 2 (ϕ2) are .9 (in previ- ous research [Cowles and Carlin 1996].91 (we calculated the log-likelihood using Newton and Raferty’s [1994] log marginal density measure). Finally. see Allenby. it can be expected that margins change over time. We used all but the last observation for each customer (calibration sample) to estimate the parameters of each model. We also computed the AIC and BIC for model selection.2 and γ2 = . The mean expected interpurchase time for Subgroup 1 is 4. BIC = –4683 for Model 3. a field experiment that tests the recommendations of such a framework on a test group against a control group that is managed according to existing norms would provide a stronger justification for CRM-based efforts.and supplier-specific antecedents used in the customer response model also can directly affect costs and thus margins. computing Gelman and Rubin’s statistic. Model selection. For both measures. We implemented the estimation algorithm using the GAUSS software package. the results show that the generalized gamma model with two subgroups provides a good fit to the data. We used “slice-sampling” (Neal 2000) to obtain random samples. Although our study is a step in the right direction to assess the accountability of marketing actions. Distribution parameters. and –6606 for Model 1.000 iterations as burn-in and used the last 10. Thus.01 purchases in a year. and visually inspecting the sampling plots has been recommended widely for convergence diagnostics (Cowles and Carlin 1996). We used relative absolute error (RAE) to evaluate the forecasting accuracy of the generalized gamma model compared with that of a naive moving average model (the moving average model is implemented as an updated average of every consecutive interpurchase time for a customer). The values of γ in the generalized gamma model are γ1 = 1.9 for Subgroups 1 and 2. this is because we used diffuse prior values and because of the large sample size of our data set. and we performed the convergence diagnostics using CODA. and Jen 1999.000 iterations to obtain posterior statistics. Further studies can develop and test hypotheses that directly relate CRM efforts to costs and margins. and θk) have more than 99% of their samples different from zero. The customer. the results are based on 50. In addition. For Cohorts 1 and 2. In each chain. A generalized gamma model with the probit link function (two subgroups) but no time-varying covariates (Model 2) provides a log-likelihood of –2908. However.the case of a budget constraint. Leone. Finally. Appendix). Further research can investigate more sophisticated techniques that explicitly treat these variables as endogenous.66. because we estimated the customer response model in a single step.2 to . We assessed the autocorrelations of the posterior samples to perform thinning. our optimization algorithm is flexible enough to allow for inclusion of the budget constraint without any substantial adjustments. bidirectional communication. Out-of-sample forecasting accuracy. In addition. and there were no changes in the substantive results of the study. a notable issue that arises from our analyses is whether the recommendations from an optimization framework pan out when implemented in the real market.g.

In the holdout sample. for example. Overall. Model 3 provides the best improvement in forecasting accuracy compared with Model 2 (RAE = . We evaluated the performance of the contribution margin model in the holdout sample on the basis of our estimates from the calibration sample. Finally. The mean unit cost of standardized modes of communication is approximately $3. The purchase frequency model predicts the expected time in months until next purchase for each customer. t is an index for time period (t = 0 for the period of customer initiation.458) Notes: All reported values are in dollars and are rounded to the nearest integer. Out-of-sample forecasting accuracy. Based on the RAE measure. Table A1 provides the descriptive statistics of the observed contribution margin and the predicted contribution margin in the holdout sample (period t + 1). We projected the profit in each year to current terms using a discount factor.538 (22. which is based on the lending rate that is appropriate for the time of the study. The contribution margin model provides an adjusted R2 of .68. the contribution margin in period t + 1 is an independent variable and is obtained from the prediction in period t + 1. and the mean unit cost of rich modes of communication is $60.881 (1. The RAE is given by the ratio of the model MAD to the MAD based on the moving average measure. We then compared the forecast time until the next purchase with the holdout sample (formed from the last observation for each customer) to assess the mean absolute deviation (MAD). the generalized gamma model with time-varying covariates (Model 3) has an RAE of .purchase from Equation 3.598) Minimum –171 (–181) –178 (–159) Maximum 1.938.981) 1. and the mean observed contribution margin is approximately $64. we assessed the number of purchases in the holdout sample that the model also correctly predicted as a purchase. the mean predicted contribution margin in period t + 1 is approximately $67. the model currently identifies 89% of them.84) and Model 1(RAE = . As with typical time-series models.396. T is TABLE A1 Comparison of Descriptive Statistics Between Observed Contribution Margin and Predicted Contribution Margin in the Holdout Sample Mean Observed Predicted 50. the model currently identifies 90% of them. we advanced the independent variables by one period to forecast the dependent variable in period t + 1.283) Standard Deviation 23. We estimated the contribution margin model on annual data from 1997 (t – 4) to 2000 (t) for Cohort 1 and from 1998 to 2000 for Cohort 2.t is the contribution margin for customer i in period t. We observe that among customers who bought within 12 months. The PCV calculation is as follows: ( B1) PCVi = t=0 ∑ (CM T it − MCit ) × (1 + r )T − t . We define PCV as the cumulative profits obtained from a customer until the current period.836) 24. compared with 10. Appendix B Description of Customer Selection Metrics CLV We entered predictions from the purchase frequency (Equation 3) and contribution margin (Equation 4) models into Equation 2 to obtain the net present value of future profits (period t + 1) from each customer. when the contribution margin model is used to predict period t + 3.010. We computed the unit cost of communication for each customer as the ratio of the total contacts for a given channel in a given year to the total cost of contact for a given mode in a given year. The MAD from Model 3 is 5.24 months for the moving average measure. We assigned a 30% margin after accounting for cost of goods sold (the managers who provided the database informed us that 30% was a nominal margin for most of their products).t denotes marketing costs for customer i in period t.729 (74. compared with that of a naive moving average technique.199 (49. We obtained the magnitude of revenue in period t + 1 by adding the predicted value from the contribution margin model to the base revenue in period t. we used an annual discount rate of 15% for each customer. MCi.420. Contribution Margin Model Estimation and influence of antecedents and covariates. and t = 0 for 1998 for Cohort 2 customers). In other words. Values in parentheses represent Cohort 2. t = 0 for 1997 for Cohort 1 customers.729.257 (1.897. The coefficients of the contribution margin model are provided in Table 5. and we computed the variable costs using costs of communication. and among customers who did not buy within 12 months. The contribution margin model predicts the growth in revenue from period t to t + 1. Specifically. We also assessed the predictive capability of the purchase frequency model using hit rate. PCR and PCV We define PCR as the revenue provided by the customer in the most recent observed purchase. the results of the analyses support all the hypothesized relationships. The cumulative profits are calculated annually from a customer’s initiation until the current period.850 (24. Customer Lifetime Value Framework / 123 .51. The revenue in 2001 (t + 1) is a holdout sample for Cohorts 1 and 2.21 months. where CMi.95).229) 57.

the unit marketing cost through standardized modes (average of direct mail and telephone sales) is $3. The P(Alive) measure uses the previous purchase pattern to predict the probability that a customer is still alive at each period in the prediction window. and Nada Nasr Bechwati (2001). A. Robert C. “Managing Customer-Initiated Contacts with Manufacturers: The Impact on Share of Category Requirements and Word-of-Mouth Behavior. and r is the discount rate. ———. John Deighton. Our last prediction for the long-term analysis is in period 2003. Leone. the cost of communication through rich modes is approximately 20 times the cost of communication through standardized modes. Our database provides information on the approximate unit cost of communication through rich modes and standardized modes to each customer. Higher values of P(Alive) indicate longer lifetime duration. 17–30. Carlin (1996). Bowman. 77–85. Robert P. 38 (August). Ruth and Katherine N. and Kumar 2004) in deriving optimal strategies for complex marketing problems. 42 (8). and Gordon Swartz (2002). 65 (January). Balakrishnan. and Peter C.” Journal of American Statistical Association. “Manage Marketing by the Customer Equity Test. Department of Marketing. 271–92. Blattberg. Thus.. we set the time frame for our optimization framework as three years. Marks. Jacob (1996).” Journal of Marketing. “The Allocation of Promotion Budget to Maximize Customer Equity. On average. Roland T. Krishnan.000 iterations should be less than . Elten Briggs. Werner J. which we do not have. “Genetic Algorithms of Product Design.” Journal of Service Research. We set the parameters in the genetic algorithm as follows: population size = 200. Venkatesan. and Creed Terry (2002). “Markov Chain Monte Carlo Convergence Diagnostics: A Comparative Review. V. October 2004 . 7When forecasting contribution margin for three years ahead. our optimization algorithm maximized the objective function by varying 432 parameters in Cohort 1 (216 customers and 2 parameters for each customer [levels of rich and standardized modes]). “Evaluating the Predictive Accuracy of Long-Term Value Models for Customer Relationship Management Strategies. 12 (Winter). 49–61. Joseph P. 29 (1). Cannon. Badi H.25. Greg M. Cowles. The sum of the optimal CLVs from each individual customer provides the optimal customer equity of the analysis sample. “Buyer–Seller Relationships and Customer Firm Costs. we used the average of marketing-mix variables from 1997 to 2001 for imputing the values in 2002. Such a cost index is commonly encountered.” OMEGA. probability of crossover = . the optimized marketing-mix variables are substituted for the 2002 values. ——— and Nada I.7 Specifically.” Journal of Marketing Research. Zeithaml. and Cooper 1997. P. 39–54. and the unit marketing cost through rich modes (salesperson contacts) is $60.the current period. “A Dynamic Model of Customers’ Usage of Services: Usage as an Antecedent and Consequence of Satisfaction. “The Theoretical Underpinnings of Customer Asset Management: A Framework and Propositions for Future Research. ———. 124 / Journal of Marketing.” working paper. Rust. and Bradley P. and Varghese A. Kumar. which we set at 15%. Bolton.” Journal of Marketing Research.” Journal of Interactive Marketing.” Management Science. 94 (June). we varied the contact levels for each customer and then calculated the sum of CLV of all customers in the sample. We ran the genetic algorithm at least 50 times and used the mean of the resource levels corresponding to the maximum CLV from each run as the resource reallocation rule for each customer. 365–74. Bolton.V.8. Reinartz. 5 (1). Douglas Bowman. and Jacquelyn S. Ruth N. However. for REFERENCES Allenby.” Journal of Service Research. CLD In our analysis. Douglas and Das Narayandas (2001). Econometric Analysis of Panel Data. we evaluated the probability that a customer is alive or dead in the planning window using the P(Alive) measure that Schmittlein and Peterson (1987) and Reinartz and Kumar (2002) recommend. which we need marketing variables in 2002. and convergence criteria = difference in solution in the last 10. In this case.” Journal of American Statistical Association. and Lemon 2004). David. Thomas (2001). 136–44. 5 (1). The genetic algorithm proceeds by searching for the optimal level of contact for each customer that maximizes CLV. Following research in customer equity (Rust. Nasr (1998). and Christian Homburg (2001). Paul D. Our objective was to calculate the maximum value for this sum of the CLVs. 1105–1118. Northwestern University. Berger. Midley. “Marketing Actions and the Value of Customer Assets: A Framework for Customer Asset Management. Verhoef (2004). and Lichung Jen (1999). Mary K. Mantrala. note that we have data on marketing activities from 1997 to 2001. probability of mutation = . Parasuraman. 74 (July–August). 36 (May). Naik. Boston: Harvard Business School Press. “A Dynamic Model of Purchase Timing with Application to Direct Marketing. 32 (3). 883–904. Customer Equity: Building and Managing Relationships as Assets. Baltagi. Gary Getz.01%. 281–97.” Journal of the Academy of Marketing Science. (1998). Appendix C Genetic Algorithms to Develop Resource Allocation Strategies Researchers in marketing have only recently begun to recognize the potential benefits of using genetic algorithms (Balakrishnan and Jacob 1996. 91 (June). ——— and Edward Malthouse (2003). ———. “Seven Barriers to Customer Equity Management. 171–86. and Sawyer 1998. and John Deighton (1996). Bell.” Harvard Business Review. 29–43. Lemon (1999). New York: John Wiley & Sons. For forecasting the contribution margin in 2003. We thank a reviewer for bringing this to our attention. for the optimization phase. ———. “Customer Lifetime Value: Marketing Models and Applications.

3d ed. 11–27. Edward L. 30–55. M. Winer (1999). “A Simple Genetic Algorithm. “Relationship in Consumer Markets: Antecedents and Consequences. What. 75 (January–February).” Journal of Marketing.” Harvard Business Review. John E. 68 (January). 5 (1). Customer Lifetime Value Framework / 125 . Fournier. Lehmann. 2004).” Harvard Business Review. New York: McGraw-Hill. P. 214–35. Robert E. Rindfleisch. “Transaction Cost Analysis: Past. Trichy V. 61 (October). Jap. Neal. “Customer Profitability in a Supply Chain. Villas-Boas. Boston: Addison-Wesley Longman.. Morgan. Berlin: Springer-Verlag. 3–8. Kumar (2004). 2005. Sunil (1988). Morrison (1985). Das Narayandas. Schurr.. “Developing Buyer–Seller Relationships. 58 (April). 653–66. 13 (1). Roland T. Naik. 379–90. Upper Saddle River. (1989).” Journal of Marketing. Niraj. and A. “CRM and the Internet: Leading Edge Strategies for a Multi-Channel World. Marks. Hunt (1994). 26 (November).” Technical Report No. 342–55. Rajendra K. David (2001). Dwyer. 17 (3). Department of Statistics. Sawyer (1998). and Shelby D. 53–67. Mario Merino.” Marketing Science.” Harvard Business Review. “Linking Customer Assets to Financial Performance. 69–76. and Katherine N. “Is the Customer Alive?” The American Statistician. “Tackling the Retailer Decision Maze: Which Brands to Discount. “The Mismanagement of Customer Loyalty. 36–51. 86–94.. 5 (August).” Journal of Marketing. 255–71. 77–99. Jakki and John R. “Approximate Bayesian Inference with the Weighted Likelihood Bootstrap.” Journal of Marketing. Jacquelyn S.” Journal of Marketing. Heide (1997). 271–300.. 2d ed. 56). Krishnan. “Inference from Iterative Simulation Using Multiple Sequences.” Management Science.htm]. Mahendra Gupta. 96 (454). 1324–38. Sawhney. 67 (January). Verhoef (2002). The New Direct Marketing: How to Build a Profit-Driven Database Marketing Strategy. Raferty (1994). Grewal. eds. “Evolutionary Estimation of Macro-Level Diffusion Models. 80 (7). Corner. and D. and L. 41–67. Mayer (1995). and How Much to Buy. 26–38. Gupta. “The Impact of Customer Relationship Characteristics on Profitable Lifetime Duration. “Real-Parameter Evolutionary Monte Carlo with Application to Bayesian Mixture Models. Present. and Atul Parvatiyar (1995). 65 (July). When. ——— and ——— (2002). [available at http://www. “On the Profitability of Long-Life Customers in a Noncontractual Setting: An Empirical Investigation and Implications for Marketing. and Peter C. Miclo (2001). and V. Robert F. (2001). Rubin (1992).” Marketing Science. and Chakravarthi Narasimhan (2001). and David Glen Mick (1997). 7 (3).” Journal of Marketing. Libai. Liang. Werner J. Sandy D. Venkatesan. Bob (2001). 81 (May). B. Roberts.” Journal of Personal Selling & Sales Management. Mohr. A. “Endogeneity in Brand Choice Models. Sheth. and Lee G. “Slice-Sampling. and V. Gerard and Fred Zufryden (1995). and Why?” Marketing Science. Susan. 1–16. Thompson. 95–108. Murali K. and Raj Mehta (2001). “An Investigation into the Antecedents of Organizational Participation in Business-to-Business Electronic Markets. “The Impact of Sales Promotion on When. Mohanbir (2001). 20–38. 23–53. Zeithaml. 1–19. and Russell S. Nash. Rogers. David and Donald G. “Determinants of Long-Term Orientation in Buyer–Seller Relationships.” Journal of Marketing. Miguel J. Peterson (1994). Ganesan. How Much. and Alan G. 81 (July–August). Donald R. Shepard.” Journal of Marketing. Mantrala. Kamakura. 54 (October). Wagner and Gary Russell (1989). Shankar (1994).” Statistical Science. Hogan. 45 (10). NJ: Prentice Hall. “Hedging Customers. Direct Marketing Management. “Breeding Competitive Strategies. Rajdeep. ——— and Robert A. 101–108. Kumar (2000). Robert E.” in Theoretical Aspects of Evolutionary Computation. “Return on Marketing: Using Customer Equity to Focus Marketing Strategy.” Journal of Service Research. Aric and Jan B. 13 (1).Del Moral. forthcoming. Paul H. Synchronize.” Journal of Business and Economic Statistics. 457–511. Valarie A.” Journal of Marketing Research. 291–95. 39 (4). and A. Susan Dobscha.E.” Journal of Marketing. 21 (2). 51 (April). Dorsey. David E.” Management Science. Rajkumar. “The Commitment–Trust Theory of Relationship Marketing. “Asymptotic Results for Genetic Algorithms with Applications to Nonlinear Estimation. Thomas.” Journal of the Academy of Marketing Science. Jagdish N. Faming and Wing Hung Wong (2001). James M. Naudts. Ravi and Rashi Glazer (2003). “Preventing the Premature Death of Relationship Marketing. David F. Dhar. 257–75. Barak.A. Cooper (1997). 10–14. “The Strategic Role of Salesforce in Developing Customer Satisfaction Across the Relationship Lifecycle. (1993). Schmittlein. Rust. Midley. and Future Applications.” Marketing Science. “Toward an Individual Customer Profitability Model: A Segment Based Approach. “A Probabilistic Choice Model for Market Segmentation and Elasticity Structure. Tellis.” Journal of Marketing Research.com/adsections/ care/crm/crm_index. 43 (3).” Journal of Marketing. Rakesh. University of Toronto. L. and Walter J. “Don’t Homogenize. ——— and ——— (2003). 14 (3).” Journal of the American Statistical Association.” in Genetic Algorithms in Search Optimization and Machine Learning. 17–34.. Prasad A. Mary Lou and Paul Berger (1999). Nevin (1990). and Sejo Oh (1987). 439–94.” Journal of the Royal Statistical Society. 23 (4).” Harvard Business Review. 17–35. “Customer Base Analysis: An Industrial Purchase Process Application. 2–8. 3–48. Reinartz. “Genetic Algorithms for Estimation Problems with Multiple Optima. Kallel.” (accessed May 19. “Planning Media Schedules in the Presence of Dynamic Advertising Quality. 65 (July). Radford M.” Journal of Service Research.businessweek. 25 (November). New York: McGraw-Hill. Newton. Database Marketing: The Ultimate Marketing Tool. Goldberg. (2000). Srivastava. “Communication Strategies in Marketing Channels: A Theoretical Perspective. Gelman. 58 (July). Robert M. (Series B.B. 64 (October). and Clive Humby (2002). Lemon (2004).