Vol. 27, No. 4, JulyAugust 2008, pp.

585599
issn0732-2399 eissn1526-548X 08 2704 0585
informs

doi 10.1287/mksc.1070.0319
2008 INFORMS
Practice Prize Report
The Power of CLV: Managing Customer
Lifetime Value at IBM
V. Kumar
J. Mack Robinson College of Business, Georgia State University, Atlanta, Georgia 30303,
[email protected]
Rajkumar Venkatesan
Darden Graduate School of Business, University of Virginia, Charlottesville, Virginia 22904,
[email protected]
Tim Bohling
Americas Market Intelligence, IBM Corporation, New York, New York 10589, [email protected]
Denise Beckmann
Americas Market Intelligence, IBM Corporation, Atlanta, Georgia 30327, [email protected]
C
ustomer management activities at rms involve making consistent decisions over time, about: (a) which
customers to select for targeting, (b) determining the level of resources to be allocated to the selected
customers, and (c) selecting customers to be nurtured to increase future protability. Measurement of customer
protability and a deep understanding of the link between rm actions and customer protability are critical
for ensuring the success of the above decisions. We present the case study of how IBM used customer lifetime
value (CLV) as an indicator of customer protability and allocated marketing resources based on CLV. CLV
was used as a criterion for determining the level of marketing contacts through direct mail, telesales, e-mail,
and catalogs for each customer. In a pilot study implemented for about 35,000 customers, this approach led to
reallocation of resources for about 14% of the customers as compared to the allocation rules used previously
(which were based on past spending history). The CLV-based resource reallocation led to an increase in revenue
of about $20 million (a tenfold increase) without any changes in the level of marketing investment. Overall,
the successful implementation of the CLV-based approach resulted in increased productivity from marketing
investments. We also discuss the organizational and implementation challenges that surrounded the adoption
of CLV in this rm.
Key words: customer relationship management; customer lifetime value; eld experiment; return on marketing
contacts; missing value imputation
History: This paper was received September 21, 2006, and was with the authors 3 months for 2 revisions;
processed by John Roberts. Published online in Articles in Advance May 7, 2008.
1. IBM Background
IBM is one of the leading multinational high-
technology rms that markets hardware, software,
and services to B2B customers. IBM intends to proac-
tively manage individual customer relationships to
maximize overall rm protability. A variety of mar-
keting factors including product/service innovation,
product/service pricing, transaction channel, mass
market advertising, and individual customer contacts
are expected to impact customer protability. Among
these factors, the level of customer contacts is most
applicable for customization across customers, and is
the focus of differential resource allocation for man-
aging customer protability at IBM. Customers are
contacted through several channels, such as salesper-
son, direct mail, telesales, e-mail, and catalogs. At the
time of this study, the channels used for contacting
a customer depended largely on the past relationship
with a customer. For example, customers who have
employees ranging from 100 to 999 (the midmarket)
are contacted primarily through direct mail, telesales,
e-mails, and catalogs.
Each year, IBM sorts customers based on their score
on a customer-level metric and prioritizes marketing
contacts to customers based on this score. The choice
of the metric (used to score customers) is considered
as one of the primary factors that determines the
return on marketing and is constantly rened by IBM
with the objective of maximizing potential value from
customers. Through the 1990s, a customer spending
score (CSS) was used to score customers. CSS was
dened as the total revenue that can be expected from
585
Kumar, Venkatesan, Bohling, and Beckmann: Practice Prize Report: The Power of CLV: Managing Customer Lifetime Value at IBM
586 Marketing Science 27(4), pp. 585599, 2008 INFORMS
a customer in the next year. Customers from the top
one or two CSS deciles, depending on the customer
segments (small to medium businesses or large com-
panies), are selected for future targeting.
Over the years, additional components were added
to augment CSS. Ultimately, the CSS score was
abandoned because it focused primarily on customer
revenues (i.e., the top line) and largely ignored the
variable cost of serving a customerand therefore,
the bottom line. Trials of customer protability and
customer lifetime value (CLV) were proposed as an
alternative for scoring customers. In this study, we
present the case study of how CLV was measured,
evaluated, and implemented in IBM for prioritiz-
ing marketing resources among customers in the
midmarket.
The objectives of the customer selection process at
IBM are summarized by the following questions:
Which customers should be selected for tar-
geting?
Is there a way to determine the level of
resources to be allocated to those customers?
How can the selected customers be nurtured in
order to increase future protability?
In trials we tested a simple management belief:
Can an increase in contacts to the right customers
create high value from low-value customers when
all other drivers are similar? To accomplish these
objectives, IBM adopted the following CLV manage-
ment framework. The CLV management framework
is intended to guide the marketing activity directed
towards customers each year.
2. Customer Lifetime Value
Management Framework and
Impact
The measurement of CLV and the alignment of
marketing resources to customers with the highest
potential value are at the core of the proposed CLV
management framework illustrated in Table 1.
Although the components of this framework have
been proposed separately in past literature (Gupta
and Zeithaml 2006), this is the rst study to pro-
vide an integrated framework that is also suitable for
eld implementation. In this paper, we discuss imple-
mentation, through a eld study, of the CLV man-
agement framework for a sample of IBMs customers
(the midmarket), and the impact of this implementa-
tion on marketing productivity. Although eld studies
are relatively scarce in the literature, they are essen-
tial for establishing the external validity of marketing
strategies and vastly improve the adoption of the pro-
posed strategy by rms across industries. The eld
study that we report here also illustrates the interplay
Table 1 Customer Lifetime Value Management Framework
Process Purpose
Measure customer lifetime
value (CLV)
To obtain a measure of the potential
value of IBM customers.
Identify the drivers of CLV So that managers can inuence the CLV
Determine optimal level of
contacts for each customer
that would maximize their
respective CLV
To guide managers on the level of
investment required for each
customer
Develop propensity models to
predict what product(s) a
customer is likely to purchase
To develop a product message when
contacting a customer
Reallocate marketing contacts
from low CLV customers to
high CLV customers.
To maximize marketing productivity
between the short-term tactical decisions such as mar-
keting contacts and the long-term strategic focus on
CLV. In addition, this paper discusses some initiatives
at IBM that further expand upon the foundations of
the eld study reported here.
The results from the eld study are encouraging.
Marketing resources were reallocated based on CLV
calculations for 14% of the customers. The resource
reallocation involved contacting a certain set of cus-
tomers in 2005 who were not contacted until 2004,
but who are predicted to have high CLV. On aver-
age, the revenue from these customers increased ten-
fold. The total increase in revenues in the test sample
for the midmarket customers, as compared to previ-
ous years, is about $20 million dollars. This increase in
revenue is obtained without any changes in the level
of marketing investment.
The rest of the article is organized as follows. In 3,
we provide an overview of how the customer prof-
itability management framework was implemented
in IBM. Section 4 explains the context in which the
study was conducted, and 5 provides details on
the rst stage, model development and measurement.
The successful application of the customer protabil-
ity management framework to IBMs customer base
has been remarkable, and has been described, along
with some extensions, in 6. In 7, we discuss the
organizational challenges most companies would face
when implementing the framework. The paper con-
cludes with a summary in which further issues such
as transferability to other industries and the limitation
of the customer protability management framework
are discussed.
3. Assessing the Impact of the
Customer Lifetime Value
Management Framework
We now describe the two-stage process used to
determine whether the CLV management framework
Kumar, Venkatesan, Bohling, and Beckmann: Practice Prize Report: The Power of CLV: Managing Customer Lifetime Value at IBM
Marketing Science 27(4), pp. 585599, 2008 INFORMS 587
Figure 1 Implementing Customer Lifetime Value Management Framework
Purchase history Marketing information
Computation of customer selection metrics
Compare customer selection metrics in
choosing the valuable customer.
Who to Target
Identifying the optimal contact strategy for
customer selection using CLV
Developing ContactContact strategy
How many times to ContactContact
Field study
Contact group No contact group
Compare
performance
Propensity modeling
What products to Pitch
Experimentation
time period
Estimation
time period
Gross margin data
achieved its objectives. In the rst stage, several
models were developed to generate inputs for imple-
mentation. In the second stage, a eld study was con-
ducted based on recommendations from the models
developed in the rst stage. In the eld experiment,
we tested an initial theoretical assumption within IBM
that, an increase in contacts to the right customer creates
high value from low value customers when all other drivers
are similar.
3.1. First Stage: Model Development
IBM used CLV, coupled with other qualitative inputs
from the sales team, to base their marketing decisions
for a sample of midmarket customers. Figure 1 illus-
trates the phased approach that was implemented,
and the various steps that were executed in order are
explained below:
Phase I: Dened and developed a model to mea-
sure CLV for each customer.
Phase II: Conducted a prediction exercise to com-
pare the performance of customers rank ordered based
on the traditionally used metrics with that of CLV.
Phase III: To maximize CLV, we developed an opti-
mal contact strategy to allocate contact resources to
each customer. Once the guidelines for the optimal
number of contacts (in different channels of commu-
nicationtelephone, catalog, e-mail, and direct mail)
with the customers are established, we then observed
the performance in the marketplace based on our
recommendations.
Phase IV: Propensity models were built for each
product category to identify the product to feature in
addition to the CLV measure, which helps to select the
customers for targeting; and the optimization process,
which suggests the contact strategy.
3.2. Second Stage: Implementation
Phase V: Customers were split into two groups
(1) not contacted so far (Not Contacted Until 2004
group), and (2) previously contacted (Contacted by
2004 group). Marketing contacts were then reallocated
to align resources to the high CLV customers.
Phase VI: Customers with potential for higher CLV,
but not contacted so fari.e., the Not Contacted until
2004 groupwere contacted in 2005 as per the rec-
ommendations of the CLV-based approach. The per-
formance of this group of customers was compared
between the years 2004 and 2005 to illustrate the
impact of the model recommendations. Further,
the model recommendations were evaluated even for
the intersection of the Contacted by 2004 group and the
Contacted in 2005 group to see if we missed out on
existing source of revenues. In other words, if we rec-
ommended that someone be contacted in 2005, did
the customers provide higher revenue than the cus-
tomers who were not contacted?
4. Study ContextData
Data on midmarket companies were used to evalu-
ate the impact of the CLV-based framework. The mid-
market customers represent companies with number
Kumar, Venkatesan, Bohling, and Beckmann: Practice Prize Report: The Power of CLV: Managing Customer Lifetime Value at IBM
588 Marketing Science 27(4), pp. 585599, 2008 INFORMS
of employees in the range 100999 at the enterprise
level and with total enterprise revenues to IBM in
the last three years (20012003) that were greater than
$25,000. The total number of enterprises in the data
set was greater than 20,000. Because some enterprises
can have more than one establishment (e.g., different
locations making independent decisions), we conduct
our analysis at the establishment level. We aggregate
all the customer transactions over a month and con-
duct our analyses at the monthly level because at
least a month is required to sort out the complexities
involved in selling to other businesses, and most of
the transactions registered within a certain month in
the customer database correspond to the same cus-
tomer order. We have a total of 2.5 million observa-
tions (72 months each for the 35,131 establishments)
in the analysis group.
5. First-Stage Model Development
and Measurement
5.1. Phase I: Measurement of CLV
We adopt the always-a-share approach for measuring
CLV because it is more appropriate for the noncon-
tractual setting of IBM (Reinartz and Kumar 2000,
Rust et al. 2004, Venkatesan and Kumar 2004). The
always-a-share approach assumes that there is only
dormancy in a customer-rm relationship and that
customers never terminate their relationship with a
rm. This assumption allows for a customer to return
to purchasing from a rm after a temporary dor-
mancy, and when the customer returns to the relation-
ship they retain the memory about their prior rela-
tionship with the rm.
We dene the lifetime value for customer i as the
net present value of cash ows they provide over
a three-year period (36 months). Theoretically, CLV
models should estimate the value of a customer over
the customers lifetime. However, in many rms,
including IBM, three years is considered to be a good
estimate for the horizon over which the current busi-
ness environment (with regard to technology, com-
petition, etc.) would not change substantially. Hence,
most customer relationship management (CRM) deci-
sions are made based on CLV estimates over a rolling
three-year window. Further, in most cases the major-
ity of a customers lifetime value is captured within
the rst three years (Gupta and Lehmann 2005). For
example, if the retention rate is equal to 75%, and
the discount rate is equal to 20%, then three years
accounts for about 86% of the CLV.
1
We therefore
measure CLV as,
CLV
i
=
T +36
_
=T +1
p(8uj
i
=1)

CM
i
(1 +r)
T

MT
i
MC
(1 +r)
T
, (1)
1
We thank the area editor for providing us with this example.
where,
CLV
i
= Lifetime value for customer i,
p(8uj
i
) = Predicted probability that customer i will
purchase in time period ,

CM
i
= Predicted contribution margin provided
by customer i, in time period ,

MT
i
= Predicted level of marketing contacts (or
touches) directed towards customer i in
time period ,
MC= Average cost for a single marketing con-
tact, this is assumed to be $7 in the study,
= Index for time periods; months in this
case,
T = Marks the end of the calibration or obser-
vation time frame, and
r = Monthly discount factor; 0.0125 in this
case (amounts to a 15% annual rate).
Computation of CLV requires predictions on three
aspects: (a) the level of marketing contacts directed
towards customer i in time period (MT in Equa-
tion (1)), (b) the probability that a customer would
purchase in each time period (p(8uj) in Equa-
tion (1)), and (c) the contribution (in $s) provided by
the customer in each time period (CM in Equa-
tion (1)).
Model Likelihood. The three aspects involved in the
computation of CLV are inherently correlated. The
level of marketing contacts directed towards a cus-
tomer depends on customer characteristics, past cus-
tomer behavior, and the past level of marketing
resources allocated towards the customer. The prob-
ability that a customer would purchase is likely to
be dependent on the level of marketing resources
directed towards the customer, and nally, the cus-
tomer provides prots to the rm only if they
purchase. In our model framework we allow for
correlations among these aspects of rm and cus-
tomer behavior. We model marketing contacts, prob-
ability of purchase, and contribution margin jointly
through a seemingly unrelated regression-based
model structure. The likelihood that summarizes our
model structure is provided below,
|(MT, 8uj, CM)

i=1
T

=1
Pr(8uj

i
0, MT
i
)
18uj
i
Pr(CM

i
=CM
i
, 8uj

i
>0, MT
i
)
8uj
i
(2)
where
8uj

i
=customer is latent utility for purchasing in
time period .
Please refer to Appendix A for further details on
the likelihood. As indicated in Table 2 and explained
in Appendix A, the proposed framework extends the
Kumar, Venkatesan, Bohling, and Beckmann: Practice Prize Report: The Power of CLV: Managing Customer Lifetime Value at IBM
Marketing Science 27(4), pp. 585599, 2008 INFORMS 589
Table 2 Comparison of Model Contributions
Imputing missing values Joint estimation Modeling Field Forward-looking cost
Studies in contribution margin of CLV components rms decisions experiment allocation strategy
Reinartz et al. (2005) No No No No No
Venkatesan and Kumar (2004) No No No No No
Lewis (2004) No No Yes No No
Fader et al. (2005) No Yes No No No
Donkers et al. (2007) No Yes No No No
Current study Yes Yes Yes Yes Yes
previous individual level, and always-a-share-based
CLV models along four critical dimensions: (a) mod-
eling rm decisions, (b) providing a forward-looking
cost allocation strategy, (c) imputing missing contribu-
tion margin, and (d) accommodating for unobserved
dependence among levels of marketing contacts, pur-
chase incidence, and contribution margin.
We used the rst 54 months of data to estimate
model parameters, and the CLV score was computed
for each customer over the next 36 months (2002
through 2004), as described in Equation (1). Further
details on model estimation and model comparisons
are provided in the Technical Appendix, which is
available at http://mktsci.journal.informs.org.
5.1.1. Drivers of CLV. Although rms are inter-
ested in knowing the lifetime value of their customers,
they are also keen on identifying the factors that are in
their control that could increase the value of their cus-
tomers. Reinartz and Kumar (2003) identied the fac-
tors that explain the variation in the protable lifetime
duration among customers. The antecedents of prof-
itable lifetime duration are grouped as exchange char-
acteristics and customer heterogeneity. The exchange
characteristics dene and describe the nature of
the customer-rm exchange, whereas rmographic
variables capture customer heterogeneity. Different
exchange characteristics that we include as drivers of
purchase propensity and contribution margin include
past customer-spending level, cross-buying behavior,
purchase frequency, recency of purchase, past pur-
chase activity, and the marketing contacts by the rm.
We also evaluated interactions between the drivers to
accommodate for nonlinearity in the inuence of the
drivers on purchase incidence and contribution mar-
gin. To accommodate for the diminishing returns to
marketing efforts (Venkatesan and Kumar 2004), we
use the logarithm of the marketing contacts variables
as predictors in the purchase incidence and contribu-
tion margin equations.
2
2
We also evaluated both a linear and a quadratic term of marketing
contacts as independent variables in the model. We found that the
in-sample t and predictive accuracy were better when log of mar-
keting contacts was used, given the range of marketing contacts in
our data.
Drivers of marketing contacts were based both on
theory and discussions about how marketing contacts
are executed at IBM. The drivers of contact history
include past customer spending, past levels of mar-
keting contacts, customer relationship (or exchange)
characteristics such as cross buying and recency, and
past purchase activity. The past levels of marketing
contacts were by denition equal to zero for estab-
lishment without contact history. Similar to the pur-
chase propensity and contribution margin equations,
we included interactions between the drivers in the
marketing contacts equation. When assigning predic-
tors, we ensured that there is at least one unique
predictor for each dependent variable, i.e., marketing
contacts, purchase propensity, and contribution mar-
gin, to ensure model identication (Greene 1993).
Customer rmographics were included as drivers
of a customer-specic intercept (o in Equation (A7) in
Appendix A) in all three components of our model
framework: marketing contacts, purchase propensity,
and contribution margin. The various rmographics
included were the sales of an establishment (a mea-
sure of the size of the establishment), an indicator
for whether the establishment belonged to the B2B or
B2C industry category, and the installed level of PCs
in the establishment (PcQ), a measure of the level of
demand for IT products in the establishment.
The coefcient estimates of the drivers of CLV are
reported in Table 3. The reported values are the poste-
rior means and variances. A parameter is considered
not signicant if a zero exists within the 2.5th per-
centile and 97.5th percentile values of the posterior
distribution for that parameter.
Regarding the level of marketing contacts, the
results reect IBM practices related to contacting cus-
tomers. The level of marketing contacts for a cus-
tomer depends on the recent purchase behavior of the
customer, which is captured through the covariates
lagged indicators of purchase incidence, lagged con-
tribution margin, and the lagged number of purchases
made by a customer so far. Whereas a customers past
purchase behavior in general determines whether a
customer is contacted, the specic level of marketing
contacts directed towards a customer in a particular
month is inuenced to a large extent by the level con-
tacts for the customer in the two prior months. Finally,
Kumar, Venkatesan, Bohling, and Beckmann: Practice Prize Report: The Power of CLV: Managing Customer Lifetime Value at IBM
590 Marketing Science 27(4), pp. 585599, 2008 INFORMS
Table 3 Estimation Results
Coefcients

Independent variables Mean Variance

Level of marketing contacts
Lagged level of contacts 07366 00316
Two-period lagged level of contacts 03239 00333
Lagged average number of purchases 05836 02158
Two-period lagged indicator of purchase 47114 14023
Interaction of cross buying and recency 00149 00035
Lagged contribution margin 06016 01128
Purchase incidence
Lagged indicator of purchase 06573 00902
Two-period lagged indicator of purchase 02172 00891
Lagged average level of contribution margin 00056 00026
Log of lagged level of contacts 00041 00012
Interaction of cross buying and recency 00047 00074
Interaction of log of lagged level of contacts 00004 00002
and lagged indicator of purchase
Contribution margin
Lagged contribution margin 08612 00247
Lagged average contribution margin 07442 00325
Cross buying 02858 01075
Frequency of purchases 73692 1887
Log of lagged level of contacts 0079 00105
Interaction of cross buying and recency 0038 00565

Mean and variance are computed using the 5th through 95th percentiles
of the posterior sample.
the interaction of cross buying and recency of pur-
chase reects IBMs strategy of focusing marketing
activities on customers who have been actively pur-
chasing products across several categories.
The signicant positive inuence of the two lagged
indicators of purchase indicates the existence of sta-
tionarity in customer purchase incidence. Also, cus-
tomers who have spent more (as captured by the
average level of contribution margin), and customers
who have made a recent purchase (indicating that the
customer is active) and have purchased across a wider
range of product categories (as captured by cross buy-
ing) are more likely to purchase in the current month.
Finally, marketing contacts have a signicant positive
inuence on customer purchase incidence, and the
inuence of marketing contacts is enhanced for cus-
tomers who have made a recent purchase.
Similar to purchase incidence, past customer spend-
ing seems to have a positive reinforcing effect on
spending in the future. Over and beyond the lagged
effects of customer spending, customers who have a
greater breadth of relationship (i.e., cross buying) and
customers who have purchased frequently in the past
provide a higher contribution margin. Also, the con-
tribution margin potential is further enhanced for cus-
tomers who have both a greater relationship breadth
and have made a recent purchase.
The heterogeneity distribution for the customer-
specic intercepts (provided in Technical Appendix
W1) indicate that although IBM allocates more mar-
keting contacts for customers who have a higher sales,
the purchase incidence and contribution margin is
lower for these customers. This is possible because
customers who have higher sales (or larger compa-
nies) in general split their purchases across several
vendors (Bowman and Narayandas 2004). Similarly,
customers in the B2B industry are allocated more mar-
keting contacts, but these customers have a lower pur-
chase incidence and lower contribution margin than
customers in the B2C industries. The coefcients for
customer sales and industry category imply that IBM
may benet from reallocating the marketing contacts.
Finally, customers who have a large installed base of
PCs (an indicator of the demand for IT-related prod-
ucts) have a higher purchase incidence and contribu-
tion margin and are contacted more by IBM.
5.2. Phase II: Comparing Traditional Customer
Selection Metrics with CLV
Difference Between CLV and the Traditionally Used
Metrics. Although recency-frequency-monetary value
(RFM), past customer value (PCV), and CSS are com-
monly used for computing the customers future
value, they suffer from the following drawbacks.
3
RFM and PCV are not forward looking and do
not consider whether a customer is going to be
active in the future. These measures consider only
the observed purchase behavior and assume that
past behavior reects future behavior. RFM assumes
that the recency, frequency, and monetary value of a
customers purchase explain the future value of the
customer. It fails to account for other factors (e.g.,
marketing actions) that help to predict the customers
future purchase behavior and the customers worth
to the rm. Also, the weights given for R, F, and M
greatly inuence the computation of a customers
worth. PCV fails to account for factors (e.g., cross
buying) inuencing future purchase behavior of cus-
tomers. It also does not incorporate the expected cost
of maintaining the customer in the future. This lim-
its its use as a valuable input in designing customer-
level marketing strategies. As mentioned before, CSS
focuses only on customer revenue and ignores the
cost of serving a customer. On the other hand, the
CLV measure incorporates the probability of a cus-
tomer being active in the future, the future contribu-
tion margin, and the marketing costs to be spent to
retain the customer. All these factors used to measure
CLV are essential for designing customer-level mar-
keting strategies that maximize rm value.
3
Details on the measurement of the traditional metrics are provided
in the Technical Appendix at http://mktsci.journal.informs.org.
Kumar, Venkatesan, Bohling, and Beckmann: Practice Prize Report: The Power of CLV: Managing Customer Lifetime Value at IBM
Marketing Science 27(4), pp. 585599, 2008 INFORMS 591
Table 4 A Comparison of Metrics for Customer Selection
Using the rst 54 months of data to predict the next 18 months of purchase behavior
Percent of cohort Customer lifetime Past customer
(selected from top) value CSS RFM value
15 Average revenue 30427 21789 22622 23542
Gross value 9184 6659 6966 7185
Variable costs 107 114 110 104
Net value 9077 6544 6856 7081
Notes. The reported values are in dollars (expressed as a multiple of the actual numbers) per customer and are cell
medians. The net result was identifying the top customers who provided the best customer value.
Metrics Evaluation. The CLV score for each cus-
tomer was computed using information from the pre-
dictions of marketing contacts, purchase incidence,
and contribution margin (obtained from the proposed
modeling framework illustrated in Appendix A), as
well as the unit marketing costs for each channel.
Seventy-two months of historical data were available
for model development. Traditional metrics were also
computed using the rst 54 months of data. Cus-
tomers were then rank ordered based on the CLV
measure as well as on the traditionally used metrics.
The comparative performance of the customers (i.e.,
the observed prots provided by the customers in the
last 18 months) in the top 15% of each metrics list
clearly shows the power of CLV to identify the best
customers for future targeting (see Table 4). Previous
research in contractual settings has found that cur-
rent prot is a good indicator of future protability
(Donkers et al. 2007). In contrast, our results indicate
that in noncontractual settings, at least with regard to
selecting high-potential customers for future target-
ing, current prot performs worse than estimates of
future protability.
5.3. Phase III: Optimal Contact Strategy
Forward-Looking Cost Allocation Strategy. In
Phase III, we use a genetic algorithm to obtain
the optimal contact strategy for each customer
(Venkatesan et al. 2007). Once the posterior distribu-
tion of the parameter estimates of the model used to
measure CLV (Phase I) are obtained, we vary the fre-
quency of marketing contacts for each customer and
then calculate the sum of the expected CLV of all the
customers in the sample. For each customer, we cal-
culate a CLV corresponding to each sampled value of
the posterior distribution of the parameter estimates
for that customer, and the average of the CLVs across
the posterior distribution for the customer provides
the expected CLV. The optimization algorithm maxi-
mizes the sum of expected CLVs for all the customers.
To compute the optimal contact strategy, we assume
that IBM would make the same number of contacts
to a particular customer every month over the three-
year prediction horizon. IBM contacts its midmarket
customers through direct mail, telesales, and e-mail.
Within IBM, the implementation of marketing con-
tacts within each channel is managed by separate
groups. Obtaining the support and participation of
each marketing channel group would have signi-
cantly delayed the implementation of the eld exper-
iment. Also, the eld experiment was intended to
illustrate the usefulness of the CLV management
framework within the organization. We expected the
success of the eld experiment to improve the chances
of organization-wide adoption of this framework and
also to help us obtain the participation of all the mar-
keting channel groups in future campaigns. Therefore,
we identify the optimal level of marketing contacts
across all channels, and not the optimal level for each
channel.
Given the posterior sample of model coefcients,
the optimization algorithm is implemented by vary-
ing the level of lagged marketing contacts and two-
period lagged marketing contacts for each customer.
For the rst time period in the prediction horizon ( =
T +1, in Equation (1)), we rst compute the level of
marketing contacts (Equation (A1) in Appendix A).
Given the predicted level of marketing contacts, we
then predict purchase incidence (Equations (A2) and
(A3) in Appendix A), and predict contribution mar-
gin (Equation (A4) in Appendix A) given the level of
marketing contacts and purchase incidence. The pre-
dicted quantities for the rst time period are then sub-
stituted back as independent variables to predict the
values in the second time period and so forth. These
predictions are carried forward to 36 months to calcu-
late CLV from Equation (1). This process is repeated
for each sampled value in the posterior distribution
of the parameters to compute the expected CLV for
each customer for a given level of marketing contacts.
Different values of lagged marketing contacts and
two-period lagged marketing contacts that are sim-
ulated from the optimization algorithm will lead
to different predictions of marketing contacts, pur-
chase incidence, and contribution margin, and hence
expected CLV. The objective of the optimization algo-
rithm is to nd the optimal level of marketing con-
tacts for each customer that would maximize the
Kumar, Venkatesan, Bohling, and Beckmann: Practice Prize Report: The Power of CLV: Managing Customer Lifetime Value at IBM
592 Marketing Science 27(4), pp. 585599, 2008 INFORMS
sum of expected CLVs of all the customers. The opti-
mization algorithm therefore maximizes the sum of
expected CLVs by varying 10,000 parameters (i.e.,
varying the lagged and two-period lagged levels of
marketing contacts for the 5,000 customers in the esti-
mation sample). We set the parameters in the genetic
algorithm as follows: population size =200, probabil-
ity of crossover =0.8, probability of mutation =0.25,
and convergence criteria =difference in optimal solu-
tion over the last 10,000 iterations is less than 0.1%.
Further details on the genetic algorithm used in this
study are provided in online Appendix W3. Note that
even though we estimate a single marketing contact
response parameter for all the customers, the optimal
level of marketing contacts that would maximize a
customers expected CLV is different because the val-
ues of the other covariates in the model vary by each
customer.
The output from the optimal resource allocation
model produced input into the decision-making pro-
cess about the number of contacts in each chan-
nel for each customer in various customer segments
described below. The differences in suggested opti-
mal contact frequencies across various customer seg-
ments are shown below. At the time of the study, the
rm was using CSS as a key metric for targeting cus-
tomers and allocation of contact resources. The seg-
ments themselves are xed. For example, one segment
could be recently acquired customers, and the other
segment could be dormant customers, and so on. Fig-
ure 2 shows the contact frequencies for each of these
Figure 2 A Comparison of Contact Strategies
A B C D E F
A
v
e
r
a
g
e
Customer segments
STATUS QUO-Using customer spending score
0
20
40
60
80
100
120
140
A
v
e
r
a
g
e
0
20
40
60
80
100
120
140
A B C D E F
Customer segments
RECOMMENDED-Using CLV
segments (whose names are not revealed due to con-
dentiality reasons). However, as shown in Table 5,
when we used the CLV metric, the contact frequen-
cies changed (as a result of maximizing CLV) for
these segments (quite the contrary to the CSS metric).
CLV, coupled with other techniques, provided further
insights for IBM to redistribute marketing contacts by
customer segments.
For example, some customers in Segment D are not
being contacted at all, and some other customers in
that segment are contacted at a very low frequency.
The proposed framework recommends that customers
in Segment D be contacted more so that the resulting
CLV is higher. This exercise indicates that the resource
allocation strategy suggested by CLV is different from
that suggested by a CSS metric. However, we would
be able to assess whether the different resource alloca-
tion strategy suggested translates into higher prots
for the rm only through a eld experiment.
The optimal level of marketing contacts is used to
guide the determination and budgeting of customer-
level marketing resources. However, recognizing the
fact that the optimization framework does not con-
sider competitive reactions and changes in the prod-
uct/service space, the optimal level of contacts served
only as a planning tool. During implementation, a
customer was contacted until a purchase, or when
an additional contact would result in negative prof-
itability. Because the customer responses in the cur-
rent period are augmented to the customer database,
repetition of the CLV measurements and the deter-
mination of the corresponding optimal level of con-
tacts over time are expected to reduce the discrepancy
between the planned optimal level of resources and
the realized level of resources.
5.4. Phase IV: Prediction of Purchase Probabilities
for Each Product Category
The CLV ranking determines which customers to tar-
get; the propensity model determines which prod-
ucts to pitch to the targeted customers. In Phase IV
we estimated a customers propensity to purchase in
each of the product categories. A series of binomial
logit models are estimated (one for each product cat-
egory) to predict customer purchase propensity. The
use of sophisticated modeling approaches that accom-
modate for coincidence of product category purchases
such as the multivariate probit model (Seetharaman
et al. 2006) are also being evaluated currently for
predicting purchase propensity. However, the sim-
ple binary logistic model approach was used in this
study because it was the primary methodology used
in the organization at the time, and we wanted to
clearly evaluate the benets from implementing the
CLV approach rst.
Propensity models were built to predict the pro-
pensity of buying products within the following
Kumar, Venkatesan, Bohling, and Beckmann: Practice Prize Report: The Power of CLV: Managing Customer Lifetime Value at IBM
Marketing Science 27(4), pp. 585599, 2008 INFORMS 593
Table 5 Optimization Strategies
Customer spending score (CSS)
Customer lifetime value Low High
High Direct mail/Telesales/Catalog/Email: Direct mail/Telesales/Catalog/Email:
Current interval is 4.82 days Current interval is 6.3 days
Optimal interval is 1.9 days Optimal interval is 5.3 days
Gross value: Gross value:
Current value is $10,936 Current value is $53,488
Optimal value is $17,809 Optimal value is $90,522
Low Direct mail/Telesales/Catalog/Email: Direct mail/Telesales/Catalog/Email:
Current interval is 9.7 days Current interval is 8.4 days
Optimal interval is 12.6 days Optimal interval is 8.3 days
Gross value: Gross value:
Current value is $743 Current value is $1,091
Optimal value is $1,203 Optimal value is $2,835
categorieshardware, software, and services. Our
assumption is that the customers need for certain
product types as well as familiarity with the focal cat-
egories are the key drivers of product category choice.
The drivers of product category choice that have a sig-
nicant inuence include (1) the proportion of same-
category purchases, i.e., the dominance of a category
over others; (2) the depth of same-category purchases
measured as the number of products purchased within
the focal category, i.e., the knowledge of the focal cate-
gory; and (3) the breadth of same-category purchases
measured as the number of different product types
purchased within the focal category. Marketing con-
tacts are not included as a predictor in the logistic
regressions because the customer database currently
does not register the product message in a market-
ing contact. The percentage of correct classications
for the propensity models were 88%, 84%, and 89%
for the hardware, software, and services categories,
respectively. A product message was used in a mar-
keting contact to a customer if the predicted purchase
propensity for the category was greater than 0.5 for
that customer. The predictions from the propensity
models were used to provide further insight on the right
product/service message to convey when contacting the
selected customers.
6. Second StageImplementation
The second stage involves the eld implementation
of the models developed and validated for measur-
ing CLV, and the selection of the product message in
the rst stage. The general guidelines for the second
stage, which involves the implementation of the CLV
management framework, are provided below:
1. Assess the distribution of customer value.
2. Identify customers with high potential value that
have been ignored in the past.
3. Select customers for contacting in the next year
by dropping customers that were targeted in the past
but are predicted to have low CLV, and picking cus-
tomers who are identied in Step 2.
4. Provide the new customer list to the salespeople.
Make them aware of the optimal level of resources
and the product message identied for each customer
in Phase III (in the rst stage) as a guideline.
5. Record the revenue obtained from each customer
and the resources that were required to obtain the
revenues.
4
6. Going forward to the subsequent year, reesti-
mate the CLV and propensity models developed in
the rst stage with the new data obtained the prior
year.
We now provide details about the implementation
of the above implementation guidelines in our study.
6.1. Phase V: Resource Reallocation
In Phase V, the marketing resources are aligned with
customers who have the highest potential. The cus-
tomers are rst divided into two groupscustomers
who have not been contacted at all (the Not Con-
tacted Until 2004 group), and customers who were
contacted previously (the Contacted by 2004 group). In
each group, the customers were further divided into
deciles, and the mean CLV in each decile is provided
in Table 6.
Table 6 shows that customers in Decile 10 for the
Contacted by 2004 group have a negative mean CLV,
and the customers in Decile 1 for the Not Contacted
Until 2004 group have a higher CLV than Decile 2
(and onwards) of the Contacted by 2004 group.
4
Salespeople also use their judgment and intuition when deciding
on the resources required for each customer. Further, salespeople
from the multiple product groups coordinate their sales calls as
deemed appropriate.
Kumar, Venkatesan, Bohling, and Beckmann: Practice Prize Report: The Power of CLV: Managing Customer Lifetime Value at IBM
594 Marketing Science 27(4), pp. 585599, 2008 INFORMS
Table 6 Resource Reallocation Based on CLV

Not contacted Contacted

Decile until 2004 ($) by 2004 ($) Customer segment
1 350,471 2,124,483 Super high CLV
2 993 125,460 High CLV
3 669 43,681
4 638 23,624
5 623 17,499 Medium CLV
6 611 13,675
7 534 10,513
8 444 8,051
9 369 5,023 Low CLV
10 80 (35)

The values reported here have been adjusted by a constant factor of the
actual gures.
Based on the distribution of CLVs across the deciles,
it was recommended that marketing contacts be real-
located from customers in Decile 10 in the Contacted
by 2004 group to customers in the top three deciles
of the Not Contacted Until 2004 group. For the cus-
tomers in the top three deciles of the Not Contacted
Until 2004 group, we then obtained the predictions
for purchase probability for the three product cat-
egories from Phase IV. Customers who had a high
purchase propensity for at least one of the three
product categories were selected as candidates for
resource reallocation. The candidate customers were
then rank ordered based on their CLV (provided in
Table 6). Marketing resources were allocated based
on this rank order (i.e., higher CLV candidate cus-
tomers were rst allocated resources) until all the
resources that were available from Decile 10 of the
Contacted by 2004 group were exhausted. The level
of resources allocated to each candidate customer was
determined based on the optimal contact strategy
described in Phase III. This process for resource reallo-
cation resulted in some customers in Deciles 1 and all
the customers in Deciles 2 and 3 in the Not Contacted
Until 2004 group being allocated marketing resources
for 2005.
6.2. Phase VI: Field Study
Recommendations from Phase V coupled with inputs
from other groups, including the sales managers,
formed the basis for the marketing contact strategy
implemented in 2005.
5
The process of implement-
ing the strategy involved contacting the customers
who were not contacted until 2004 but who exhibit
the potential for high CLV in 2005i.e., Deciles 1
through 3 in the Not Contacted Until 2004 group.
5
2005 can also be considered as a true holdout time period.
Figure 3 Measuring the Impact of the Contact Strategy
$X (2004)
$10X (2005)
R
e
v
e
n
u
e
Y% (2004)
1.6Y% (2005)
%
Average revenue/customer
(for the same group of customers)
Percent of establishments
w/purchase
No Contact until 2004 Contacted in 2005
6.2.1. Impact. The results reveal that on average
the revenue from the customers who were not con-
tacted until 2004, but contacted in 2005 increased by
10 times (across all customers), as shown in Figure 3.
To ensure that the increase in revenue is due to con-
tacts and pitching the right products/services, the
revenues for the customer group who requested not
to be contacted at allDo Not Contact groupwere
compared for the years 2004 and 2005. The average
revenue for this group in both years showed simi-
lar revenues (i.e., similar to the revenue provided in
2004 by the not contacted until 2004 but contacted
in 2005 group). We therefore infer that the higher
performance in 2005 for the group of customers who
were not contacted in 2004 was due to contacting the
right customers with the right message. As explained
before, the selection of the customers and the corre-
sponding messages were based on the models devel-
oped in Phases 1 through 5, thereby illustrating the
economic impact of these models. The effectiveness
of the propensity models was reected in the supe-
rior performance of the sales revenue metric. In other
words, if inappropriate products were pitched, then
the rm would not have realized higher revenues.
Further, the average number of contacts (8.9) per cus-
tomer across the customers was close to the predicted
number of contacts (8.1) for this group.
The net result was higher revenues (although it
is still not the net prot, given that certain selling
[general and administration] expenses have to be
accounted for, the comparison holds good), and
higher value in 2005 versus 2004 for the no con-
tact until 2004 but contacted in 2005 group of cus-
tomers (Deciles 1 through 3 in the Not Contacted Until
2004 group). The total increase in revenues for the
no contact until 2004 but contacted in 2005 group
of customers is about $19.2 million dollars. We use
the increase in revenue among the Contacted by 2004
group of customers from 2004 ($751 M) to 2005 ($793
M), which is about 5%, as the year-to-year growth
Kumar, Venkatesan, Bohling, and Beckmann: Practice Prize Report: The Power of CLV: Managing Customer Lifetime Value at IBM
Marketing Science 27(4), pp. 585599, 2008 INFORMS 595
Table 7 Further Analysis of the Not Contacted Until 2004 Group
Customers
Dormant customers active in
reactivated in 2005 2004 and 2005
Number of customers 273 1827
Change in revenue from 2004 ($) 114 M 7.6 M
Average increase in revenue 4176 M 4160
per customer ($)
in revenue from IBM customers. After adjusting for
the annual growth in customer revenue, the incre-
mental revenue due to resource reallocation among
the No Contact until 2004 group of customers is about
$19.1 M. The direct marketing contact cost for contact-
ing these establishments is about $95,200 (given the
average unit cost of $7 per contact). The model devel-
opment costs for this study are estimated to be about
$25,000. The ROI for IBM in this study is therefore
around 160.
6
Similar to Elsner et al. (2004), we investigate the
sources of incremental value. The incremental value
in 2005 for the No Contact Until 2004 but Contacted in
2005 group of customers is derived from two sources.
Table 7 shows that $7.6 million (approximately 40%) is
obtained from the increase in purchase amount from
customers who were active (or purchased) in 2004
and $11.4 million (approximately 60%) is obtained
from new purchases from customers who did not
purchase in 2004, i.e., reactivation of dormant cus-
tomers. About 273 customers who were dormant in
2004 were reactivated in 2005; therefore, the average
increase in revenue from reactivating dormant cus-
tomers was about $41,758. One thousand eight hun-
dred twenty-seven customers were active in both 2004
and 2005. Therefore, the average increase in revenue
from existing customers was $4,160. The resource
reallocation therefore seems to result in both grow-
ing existing customers and reactivating dormant cus-
tomers, with slightly more emphasis on reactivating
dormant customers.
The total revenue from the customers who were
contacted in 2004 and 2005 (based on our model rec-
ommendations) was over 750 million dollars (after
accounting for the direct marketing expenses). There-
fore, our model identied both existing and new
sources of revenue. The reallocation of marketing
resources led to about a 3% increase in prots from
2004 to 2005 for the midmarket customers from con-
tacting only 1% of the IBM customers. The impact
of the CLV-based campaign on the protability of
midmarket customers has accelerated the adoption
of the CLV management framework for other cus-
tomer segments in IBM. One such initiative that is
6
The ROI measure excludes any selling (general and administra-
tion) expenses.
notable is the implementation of the CLV manage-
ment framework for marketing multiple service solu-
tions to very large customers (i.e., customers with
greater than 1,000 employees), the large enterprise
segment. Within this initiative, sales team contacts are
prioritized based on the potential value of the cus-
tomers, and predictions from propensity models for
the various service options act as inputs for the mes-
sage of the sales team.
7. Organizational and Implementation
Challenges
The reorientation required to implement a CLV man-
agement framework has to be carefully managed by
the rm. In this section, we outline some of the
challenges faced and the steps that are necessary
to manage this change successfully. Customers are
more central to rms than brands and brand equity,
although current management practices in many rms
do not reect this shift (Rust et al. 2004). Tradition-
ally, in many companies, marketing and sales activ-
ities for individual product groups are implemented
by a product-specic sales team, sometimes with little
idea of what activities have been planned by the other
product groups targeted at the same customer.
Although many multiproduct rms may strive to
coordinate their messaging strategy, they fall short in
the customers mind when the rms various mar-
keting and sales groups set forth group-specic com-
munication goals that often are not in sync with
each other or with the overall strategy. Customers
often feel they are dealing with multiple compa-
nies because they receive uncoordinated commu-
nications from both sales and marketing. This is
possible because some customers have been observed
to have a high afnity for the rms products. On the
other hand, some potential high-value customers get
ignored. Applying the CLV management framework
helps a rm to understand which products, specic
customers, and/or customer segments are likely to
buy at varying time frames. Thus, to reap the benets
of the CLV management framework, a rm should
try to develop strategies and tactics from a customers
perspective rather than from a product perspective.
This transformation presents multiple organizational
and implementation challenges. We summarize some
of these challenges next.
Organizational challenges can be broadly catego-
rized along business and people dimensions. The
business dimension requires dening and articulat-
ing the business case for change and the desired
outcomes of change. Quantication of the projected
return on investment is required not just within spe-
cic business units but also across multiple key stake-
holders and across business units. Cross-organization
Kumar, Venkatesan, Bohling, and Beckmann: Practice Prize Report: The Power of CLV: Managing Customer Lifetime Value at IBM
596 Marketing Science 27(4), pp. 585599, 2008 INFORMS
collaboration is essential to identify and assess key
stakeholders, degree of risk, and cost of change in
the planning and implementation phases. As to the
human element, managers and professionals who
have traditionally been responsible for all aspects of
the marketing of a single brand now have to be
empowered with the ability and responsibility across
brands, products, and units. Thus, the following steps
are recommended:
1. Generate awareness of the need to change.
2. Create a desire to participate and support the
change through innovative rewards and incentives.
3. Disseminate knowledge of how to change.
4. Empower ability to implement the change on a
day-to-day basis.
5. Provide managerial reinforcement to keep the
change in place.
Success of campaigns implemented based on the
CLV management framework depends largely on
understanding customer-level campaign responses
and documenting the relationship progression to pro-
vide effective feedback to the system. The three criti-
cal aspects for campaign implementation are outlined
next.
Applying the Time-Based Probability Measure. The
measure for the probability to purchase in each prod-
uct category and customer value for each quarter or
any other time frame of interest is provided in the
campaign execution output le for each customer. The
recommendation from the output le may suggest a
signicant change in the customer contact pattern.
Figure 2, described earlier, illustrates the actual ver-
sus recommended contact pattern (aggregated across
customers within each segment, over three product
categories across four quarters) for each of the six dif-
ferent customer segments. The recommendation runs
counter to the current practice. The analysis suggests
that customer segments C, D, and E, which were
being contacted least frequently, should be contacted
most frequently to maximize CLV across various mar-
keting initiatives. If, after several contacts, the cus-
tomer does not purchase a product in the expected
quarter, it is prudent to reevaluate the contact fre-
quency with the customer in the next quarter and
onwards for that product category.
Documentation of Progress. To effectively assess per-
formance, the purchases made by the customer need
to be recorded in real time. For the customers who
have already made the purchase, the product cate-
gory purchased needs to be documented, and fur-
ther contact efforts need to be tailored based on this
information.
Performance Tracking and Monitoring. To provide the
inputs into the CLV computation, measures such
as marketing communications cost, manufacturing
cost, service delivery cost, retention and acquisi-
tion costs, and revenues need to be monitored and
reported. Other metrics that guide performance are
often reported to the management. Examples of these
are cross-sell ratio, win-back ratio, share of wallet,
increase in prots, and overall return on investment.
8. Conclusions and Future Steps
The CLV management framework proposed can allow
a rm to rene and improve its customer contact
strategy. Using the CLV-based approach, rms can:
Increase the return on their marketing invest-
ment by allocating resources towards customers who
are most likely to provide value in the next year.
Identify products to sell as bundles.
Reallocate the excess resources (after targeting
the most likely customers to buy in a given time
period) to other prospects (acquiring new customers
or reactivating dormant customers).
Plans are in place to measure the CLV for the larger
set of customers (greater than 1,000 employees) and
focus on even higher marketing productivity to
expand this concept to the entire organization. Two
major initiatives that are being considered based on
the CLV management framework include (a) segmen-
tation and proling, and (b) understanding customer
migration.
Segmenting and Proling. The segmentation and
proling initiative is intended to guide the acquisi-
tion activities. Customers are segmented into distinct
groups based on their CLV, share/size of wallet esti-
mates and rmographics-based proles of these cus-
tomer segments are built. The expected CLV and the
segment afliation of the CLV deciles for the midmar-
ket customers discussed in this study are presented
in Table 6. The customer deciles are grouped to form
four segments based on the CLV; Super High CLV,
High CLV, Medium CLV, and Low CLV. For these
customer segments the rmographics used to build
the customer proles primarily include the industry
category, the number of employees, the sales of the
establishment, and the installed level of PCs in the
establishment. Only those rmogaphics that had dis-
cernible differences across CLV segments are included
for prole analyses. The customer acquisition contacts
are then directed towards prospects that match the
proles of the Super High CLV and the High CLV
segments.
Understand Customer Migration. The customer mi-
gration initiative harnesses CLV measurements over
time to improve retention activities. Customer migra-
tion is dened as the movement of customers from
either the high-value segments (Super High CLV and
High CLV segments in Table 7) in a year to the low-
value segment (Low CLV segment in Table 7) in the
Kumar, Venkatesan, Bohling, and Beckmann: Practice Prize Report: The Power of CLV: Managing Customer Lifetime Value at IBM
Marketing Science 27(4), pp. 585599, 2008 INFORMS 597
next year, or vice versa. Separate propensity mod-
els are built to identify (1) the drivers of customer
migration to the low-protability segments and (2)
the drivers of customer migration to high-protability
segments. The various drivers that are considered
include the level and change in marketing resource
allocation, the rate of acquisition of new products by
the customer, the intensity of nancing availed by
the customer, and the changes in the depth of pur-
chasing exhibited by a customer. The drivers of cus-
tomer migration are used to guide customer retention
strategies.
We expect that the framework presented here has
immediate application in industries where a rm has
direct contact with its customers, such as online retail-
ing, telecommunications, hospitality, and direct-to-
consumer nancial services rms. In industries where
rms market through resellers or agents, rms may
adopt a multistage protability analysis where the
protability of the intermediary (i.e., the resellers or
the agents) is determined by the protability of the
end consumers associated with the intermediaries.
Marketing resource allocations can then be deter-
mined based on the intermediarys protability.
Limitations and Future Research. The CLV manage-
ment framework presented here has limitations that
could serve as fruitful avenues for future research.
The model does not include the impact of compet-
itive marketing efforts on CLV. Although we did
accommodate for competitive effects indirectly by
the imputation of unobserved contribution margin
values, a richer understanding of customer respon-
siveness to marketing contacts can be obtained by
including the level of marketing contacts initiated
by competitors at the same time. Accurate data on
competitive initiatives is of course hard to obtain,
but the customer transaction information can be aug-
mented with information from periodic surveys of
customers on the intensity of competitor actions and
the customers attitudes to IBMs products (Keller and
Lehmann 2006). The CLV management framework
presented in this study is similar in some aspects
to previous research in sales call planning (Lodish
1971, Zoltners and Sinha 2005). For example, both
our framework and CALLPLAN (Lodish 1971) con-
sider cost and revenue, as well as customer respon-
siveness, to optimize across customer allocation of
marketing resources. Whereas CALLPLAN estimates
customer responsiveness based on salesperson esti-
mates, our model framework estimates them based on
historical data on customer revenue and salescall fre-
quency. Also, our framework includes a component
on predicting customers product choices, which is
absent in the CALLPLAN framework. Future research
should evaluate the relative superiority of the pro-
posed framework with earlier sales-planning frame-
works such as CALLPLAN. The design of our eld
experiment does not allow us to estimate the con-
tribution of accurately predicting customers product
purchase propensities (Phase IV) towards the growth
in protability for the Not Contacted in 2004 group.
Research indicates that appropriate cross-selling mod-
els can improve the protability of sales campaigns
signicantly (Kumar et al. 2008). Future research that
compares the contribution provided by better models
that accurately select protable customers and mod-
els that accurately identify the product message (such
as the cross-selling models) would provide a valu-
able contribution to the marketing productivity litera-
ture. Although we measured CLV and determined the
customers to target based on CLV at the individual
customer level, we did not provide any recommen-
dations about optimal level of marketing resources
for each customer. As also suggested by Rust and
Chung (2006), studies that document the benets of
customer-level optimal marketing resource allocation
based on CLV would provide a good contribution to
the literature.
Our framework does not include the impact of
macroeconomic trends and new product introduc-
tions by competitors on CLV because the current
study dealt with a single instance of CLV measure-
ment. These are factors that are important to consider
when the CLV measurement is carried over mul-
tiple years and would prove especially critical for
understanding customer migration. The interaction
of customer appreciation initiatives such as loyalty
programs, and the frequency of communication in
each marketing channel, with customer protability-
based marketing resource allocation, is a fruitful
area for future research. Finally, the development of
sophisticated models based on the multivariate probit
framework for predicting customers basket choices
in the context of the CLV management framework is
a good avenue for future research.
In summary, the impact on marketing productiv-
ity realized in the eld study has been instrumental
in the adoption of the customer protability manage-
ment as a core framework for IBMs marketing strate-
gies for customer acquisition, retention, growth, and
win back.
Acknowledgments
The authors thank the editor, John Roberts, Gary Lilien, and
the members of the INFORMS Practice Prize Committee for
their encouragement and support of this study.
Appendix A. Details on Model Framework for
Measuring CLV
As mentioned in the text, measurement of CLV requires
predictions of (a) marketing contacts (MT) directed towards
a customer, (b) the purchase propensity of the customer
[p(8uj)], and (c) the contribution margin provided by the
customer given purchase (CM).
Kumar, Venkatesan, Bohling, and Beckmann: Practice Prize Report: The Power of CLV: Managing Customer Lifetime Value at IBM
598 Marketing Science 27(4), pp. 585599, 2008 INFORMS
We rst model the log of the level of marketing contacts
allocated by the rm towards customer i as
log(MT
i
) =o
1i
+x
T
1i
p
1
+u
1i
, (A1)
where x
1i
, p
1
, o
1i
, u
1i
are a vector of predictor variables,
a vector of corresponding coefcients, an individual-level
intercept, and an error term.
We assume that customer i, purchases from the rm
only when the customers latent utility for purchasing from
the rm (8uj

i
), exceeds a certain threshold, set to zero in
this case. We do not observe the latent utility of the cus-
tomer, but observe a binary outcome variable regarding
whether the customer purchased or did not purchase in
time period . The latent utility is mapped to the binary
outcome variable (8uj
i
) as follows:
(8uj

i
) >0, if 8uj
i
=1
(8uj

i
) 0, if 8uj
i
=0.
(A2)
The latent utility, (8uj

i
), for customer i to purchase from
the rm in time period is then modeled as a function of
predictor variables in a linear model
8uj

i
=o
2i
+x
T
2i
p
2
+u
2i
(A3)
where, similar to Equation (A1), x
2i
, p
2
, o
2i
, u
2i
are a vector
of predictor variables, a vector of corresponding coefcients,
an individual-level intercept, and an error term. Finally, we
assume that a latent variable, CM

i
, represents the amount
spent by customer i in time period , irrespective of whether
it is with the rm, as a function of predictor variables with
a linear structure
CM

i
=o
3i
+x
T
3i
p
3
+u
3i
(A4)
where, similar to Equations (A1) and (A3), x
3i
, p
3
, o
3i
, u
3i
are a vector of predictor variables, a vector of corresponding
coefcients, an individual-level intercept, and an error term.
If the customer purchased from the rm in time period ,
then the rm observes the contribution margin provided by
the customer, i.e.,
CM
i
= CM

i
, if 8uj
i
=1
= unobserved, if 8uj
i
=0. (A5)
When data about customer transactions with all the rms
in a product category are available (e.g., in scanner panel
data), we can treat the contribution margin provided by a
customer when there is no-purchase as equal to zero. How-
ever, in CRM databases, there is rich information about
customer transactions with a single rm, but almost no
information about the customers transactions with other
rms. Therefore, when we dont observe a purchase from a
customer in time period , we cannot assume that the cus-
tomer did not make a purchase at all. A no-purchase in a
CRM database would only imply that the customer has not
made a purchase with the focal rm, but may have a pur-
chase with a competitor. We therefore treat the time peri-
ods when a customer does not provide any revenue to the
rm as missing data. We then impute these missing values
based as random realizations from a normally distributed
prior distribution for the missing contribution margins. We
assume the mean and variance of the prior distribution
to be the empirical mean and variance in the contribution
margin provided by a customer to IBM in a calibration
sample.
For observations with no purchase incidence (i.e., 8uj
i
equals zero), there is a possibility that the customer has
made a purchase with a competitor. However, the imputa-
tion of the missing contribution margins should also take
into account a customers interpurchase time. We therefore
compute a customers average interpurchase time from the
calibration sample used to compute the mean and vari-
ance for the prior distribution. The imputation of contri-
bution margins is performed only for those observations
whose time since last purchase (or recency) is greater than
the average interpurchase time calculated in the calibra-
tion sample. For example, if a customers average inter-
purchase time in the calibration sample is three, then
only those missing contribution margin observations whose
recency is greater than three are imputed from the prior
distribution; the contribution margin is allowed to be zero
for the nonpurchase occasions whose recency is less than
three. We choose this approach for selecting the nonpur-
chase occasions that qualify for imputation based on the
assumptions about customer purchase patterns in IBM. We
specify the following prior for missing CMs in months ,
when the recency of purchase exceeds customer is aver-
age interpurchase time obtained from the calibration sam-
ple, |CM

i
recency
i
>

|
i
] = TN
|, 0]
(j
i, CM
, u
2
i, CM
), where

|
i
= average interpurchase time for customer i in the
calibration sample, j
i, CM
= average observed contribu-
tion margin provided by customer i in the calibration
sample, and u
2
i, CM
= variance in the observed contribu-
tion margin provided by customer i in the calibration
sample.
The imputations are obtained as a random draw from a
truncated normal distribution (whose parameters are deter-
mined in the calibration sample as explained before) with
limits and 0. In other words, we impute negative con-
tribution margin values (including zero) for the no purchase
incidence occasions reecting the contribution margin lost
to competitors on these occasions. The customer-specic
mean and variance of the prior distribution accommodates
for customer-specic variations in the spending levels. The
prior distribution parameters assume that a customer has
the same spending level with IBM as well as with the com-
petitors. We adopt this conservative assumption because
of a lack of a reliable share of wallet information about
the customers. If reliable share of wallet information were
available, then the product of a customers average contri-
bution margin with IBM and the share of the customers
competitor purchases can be used as the mean of the
prior distribution. Given that no information is available
about customer transactions with competition, the calibra-
tion sample provides us with the best estimate (or expecta-
tion) of customer interactions with competitors. We expect
that by treating a no-purchase occasion in this manner
we would obtain better estimates of a customers value,
the opportunity cost of a customer and of the customers
responsiveness to marketing communications. The impu-
tation of missing contribution margin is performed only
Kumar, Venkatesan, Bohling, and Beckmann: Practice Prize Report: The Power of CLV: Managing Customer Lifetime Value at IBM
Marketing Science 27(4), pp. 585599, 2008 INFORMS 599
in the estimation sample. When calculating CLV, the con-
tribution margin is set to zero for months when the pre-
dictions indicate no purchase incidence. The imputation
of missing contribution margins in the estimation sample
are hence relevant for providing better estimates of a cus-
tomers response elasticity, which is critical for accurately
calculating CLV. Our approach here is the rst step towards
accommodating the missing information about customers
transactions with competition. Future research that evalu-
ates other mechanisms for addressing this issue would be
very valuable.
The above formulation is similar to the seemingly unre-
lated regression model structure, and the predictor vari-
ables in Equations (A1), (A3), and (A4) need not be the
same. We assume that the covariance structure of the errors
in Equations (A1), (A3), and (A4) may be modeled as,
_
_
_
_
u
1i
u
2i
u
3i
_

_
N
_
_
_
_
_
_
_
_
0
0
0
_

_
,
_
_
_
_
1 u
12
u
13
u
12
u
22
u
23
u
13
u
23
u
33
_

_
_
_
_
_
=N
3
(0, l
j
), (A6)
affording the possibility of correlations among the three
residuals. We x u
11
to be equal to one to ensure model iden-
tication. The covariance structure of the errors accounts
for any unobserved dependence between a rms decision
to contact a customer (MT), a customers decision to pur-
chase from the rm (Buy), and the amount the customer
spends with the rm (CM). Letting p = |p
1
, p
2
, p
3
], and
o =|o
1
, o
2
, o
3
], the simultaneous equations model gives rise
to the likelihood specied in Equation (2). The customer-
specic intercept terms are obtained froma multivariate nor-
mal distribution;
o
i
AVN (AZ
i
, l
o
) (A7)
where,
Z
i
=a p 1 vector of customer characteristics;
A=a 3p matrix of coefcients for the customer charac-
teristics;
l
o
=a 3 3 variance-covariance matrix;
p = number of customer characteristics that are used to
capture heterogeneity.
We assume diffuse and conjugate priors for the model
parameters. We assume multivariate normal priors for p,
and A. Let d
p
denote the dimension of the p vector; then, the
prior specication of p is given as p AVN(j
p
, l
p
), where
j
p
is a d
p
dimensional column vector of zeros, and l
p
=
100
dp
;
dp
is a d
p
d
p
dimensional identity matrix. Similarly,
let d
A
denote the dimension of A; then the prior specica-
tion of A is given as p AVN(j
A
, l
A
), where j
A
is a d
A
dimensional column vector of zeros, and l
A
=
d
A
;
d
A
is a
d
A
d
A
, dimensional identity matrix. Inverse Wishart priors
are assumed for the variance parameters. The prior speci-
cation for l
o
is given as l
o
= W(j
d
A
, j), where j =15
and
d
A
is a d
A
d
A
, dimensional identity matrix. The prior
specication for l
j
is given as l
j
= W(15
3
, NT ), where
3
is a 3 3 dimensional identity matrix.
Further details on the data augmentation, the prior spec-
ications, and the Markov chain Monte Carlo (MCMC)
estimation algorithm can be obtained from Cowles et al.
(1996).
References
Bowman, D., D. Narayandas. 2004. Linking customer management
effort to customer protability in business marketing. J. Mar-
keting Res. 41(4) 433447.
Cowles, M. K., B. P. Carlin, J. E. Connett. 1996. Bayesian tobit
modeling of longitudinal ordinal clinical trial compliance data
with nonignorable missingness. J. Amer. Statist. Assoc. 91(433)
8698.
Donkers, B., P. C. Verhoef, M. G. de Jong. 2007. Modeling CLV:
A test of competing models in the insurance industry. Quant.
Marketing Econom. Forthcoming.
Elsner, R., M. Krafft, A. Huchzermeier. 2004. Optimizing Rhenanias
direct marketing business through dynamic multilevel model-
ing (DMLM) in a multicatalog-brand environment. Marketing
Sci. 23(2) 192206.
Fader, P. S., B. G. S. Hardie, K. L. Lee. 2005. RFM and CLV: Using
iso-value curves for customer base analysis. J. Marketing Res.
42(4) 415430.
Greene, W. H. 1993. Econometric Analysis, 3rd ed. Prentice Hall, Inc.,
Upper Saddle River, NJ.
Gupta, S., D. Lehmann. 2005. Managing Customers as Investments:
The Strategic Value of Customers in the Long Run. Wharton School
Publishing, Upper Saddle River, NJ.
Gupta, S., V. Zeithaml. 2006. Customer metrics and their impact on
nancial performance. Marketing Sci. 25(6) 718739.
Keller, K. L., D. R. Lehmann. 2006. Brands and branding: Research
ndings and future priorities. Marketing Sci. 25(6) 740759.
Kumar, V., R. Venkatesan, W. Reinartz. 2008. Performance implica-
tions of adopting a customer-focused sales campaign. J. Mar-
keting. Forthcoming.
Lewis, M. 2004. The inuence of loyalty programs and short-term
promotions on customer retention. J. Marketing Res. 41(August)
281292.
Lodish, L. M. 1971. CALLPLAN: An interactive salesmans call
planning system. Management Sci. 18(4) 2540.
Reinartz, W. J., V. Kumar. 2000. On the protability of long life-
time customers: An empirical investigation and implications
for marketing. J. Marketing 64(4) 1732.
Reinartz, W. J., V. Kumar. 2003. The impact of customer relationship
characteristics on protable lifetime duration. J. Marketing 67(1)
7799.
Reinartz, W., J. Thomas, V. Kumar. 2005. Balancing acquisition and
retention resources to maximize customer protability. J. Mar-
keting 69(January) 6379.
Rust, R., T. S. Chung. 2006. Marketing models of service and rela-
tionships. Marketing Sci. 25(6) 560580.
Rust, R., K. N. Lemon, V. A. Zeithaml. 2004. Return on marketing:
Using customer equity to focus marketing strategy. J. Marketing
68(January) 109127.
Seetharaman, P. B., S. Chib, A. Ainslie, P. Boatwright, T. Chen,
S. Gupta, N. Mehta, V. Rao, A. Strijnev. 2006. Models of multi-
category choice behavior. Marketing Lett. 16(34) 239254.
Venkatesan, R., V. Kumar. 2004. A customer lifetime value based
framework for customer selection and resource allocation strat-
egy. J. Marketing 68(October) 106125.
Venkatesan, R., V. Kumar, T. Bohling. 2007. Optimal CRM using
bayesian decision theory: An application to customer selection.
J. Marketing Res. 44(November) 579594.
Zoltners, A., P. Sinha. 2005. Sales territory design: Thirty years of
modeling and implementation. Marketing Sci. 24(3) 313331.