Upload
matt-fabian
View
219
Download
0
Embed Size (px)
Citation preview
8/2/2019 Sequential Xsell
1/28
1
Cross-Selling Sequentially Ordered Products: An Application to
Consumer Banking Services
Shibo Li, Baohong Sun and Ronald T. Wilcox
Shibo Li is an Assistant Professor of Marketing at Rutgers University. Baohong Sun is an Assistant Professor ofMarketing at Kenan-Flagler Business School of University of North Carolina at Chapel Hill. Ronald T. Wilcox is anAssociate Professor of Business Administration at the Darden Graduate School of Business Administration,University of Virginia. The authors would like to thank Peter Boatwright, Ron Goettler, Frenkel ter Hofstede, AjayKalra, Puneet Manchanda, Charlotte Mason, Alan Montgomery, and Kannan Srinivasan for valuable comments.
8/2/2019 Sequential Xsell
2/28
2
Cross-Selling Sequentially Ordered Products: An Application toConsumer Banking Services
Abstract
In service and consumer technology product industries, we often observe people sequentiallypurchasing multiple products and services from the same provider. These sequential purchases can takeplace over an extended period of time and can often be naturally ordered in terms of complexity andfunctionality, with the purchase of some products or services generally preceding that of others. Thiscommonly observed situation offers significant opportunities for companies carrying multiple productsand services to cross-sell other products and services to their existing customer base.
In this paper, we adopt a multivariate probit model, implemented in a Hierarchical Bayesframework, to investigate how customer demands for multiple products evolve over time and itsimplications on sequential acquisition patterns for naturally ordered products. This paper is the first paper
that formally models customers sequential demands and investigates the behavioral reasons drivingconsumers acquisition patterns for multiple products and services. We also study how the sequentialdemands for naturally ordered products develop differently across customers. This model helps managerspredict which product or service a consumer is likely to buy and in what sequence.
We investigate customer purchase patterns for products marketed by a large Midwestern bank. Inthis process we highlight interesting substantive findings and provide guidance to managers charged withallocating marketing dollars towards customers with the greatest incremental profit potential.
Keywords: Sequentially ordered products; Demand maturity; Cross-selling; Bank and financial products;
CRM; Multivariate probit model; Hierarchical Bayes framework
8/2/2019 Sequential Xsell
3/28
3
1. Introduction
We often observe consumers purchasing multiple products and services from the same provider
over time.1 These products often can be naturally ordered in terms of complexity and
functionality leading to the empirical regularity that the purchase of certain products generally
precedes the purchase of others. For example, we often observe that a person generally
establishes a checking account with a given bank before establishing a brokerage account. A
woman may repeatedly patronize a salon or spa for a haircut only before moving on to
purchasing facial treatments. A consumer may also sequentially purchase local and long distance
telephone service, cable television service, and Internet access from the same company. A person
who purchases a personal digital assistant may acquire Internet access, additional memory, and
software from the same provider in the future. The common thread running through each of these
examples is that consumers are more likely to purchase some products or subset of products
before others. The authors have directly observed this phenomenon in the market for consumer
banking services, but strongly suspect that it is prevalent in many other environments as well.
We term the over-time development of consumers complementary demand for multiple
products and services as natural ordering among these products and services.
Markets especially prone to this empirical regularity include those in which consumer
wants evolve after some preliminary consumption, those in which consumers face some
uncertainty about the quality of the product or service offering or markets in which some
consumer learning is required to receive the full benefit of the product. In such markets,
sequentially purchasing multiple products or services from the same provider can enhance the
relationship with the provider, raise switching costs associated with moving to a new provider,
lower uncertainty with respect to additional product purchases, and, in some cases, ensure proper
technical compatibility with products the consumer already owns (Kamakura, Ramaswami and
Srivastava 1991; Kamakura, Wedel, Rosa and Mazzon 2002).
The existence of sequentially developed demand for naturally ordered products offers
great opportunities for companies carrying multiple products and services to cross-sell other
products and services to their existing customer base. These companies are eager to explore
additional profitable opportunities with their current customer base by cross-selling because it is
1 Throughout this paper we will refer to products and services interchangeably. The model we develop is equallyapplicable to both.
8/2/2019 Sequential Xsell
4/28
4
often cheaper to cross-sell to their existing customers than to attract new ones. Further, customer
retention is enhanced with the cross-selling as customer switching cost increases with more
complex relationships (Kamakura, Ramaswami and Srivastava 1991). Indeed, many companies
have come to realize that their current customers are by far the best prospects for new or existing
products and services.
There have been a few descriptive studies over the past four decades that probe
consumers sequential purchases (e.g. Hauser and Urban 1986; Mayo and Qualls, 1987;
Boulding, Karla, Staelin andZeithaml 1993; Bitner and Zeithaml 2000). This research focuses
on describing the existence of sequential acquisition patterns. To our knowledge, the earliest
paper that formally models cross-selling opportunities is Kamakura, Ramaswami and Srivastava
(1991). This research applies latent trait analysis to position financial services and investors
along a common continuum. Using this approach, they obtain the estimates of the purchase
ordering of financial services as well as the latent financial maturity for each household based on
(only) the current ownership of financial services. Recently, Kamakura and Kossar (2001)
develop a split hazard rate model and focus on predicting each customers (physicians) time of
adopting a new product (drug) based on the timing of their past adoptions of multiple products.
Knott, Hayes and Neslin (2002) present four next-product-to-purchase models (discriminant
analysis, multinomial logit, logistic regression, and neural net) that can be used to predict what is
likely to be purchased next and when. Kamakura, Wedel, de Rosa and Mazzon (2002) develop a
mixed-data factor analyzer that is an extension of factor analysis to incorporate various types of
data (choice, counts, or ratings) and is tailored to identify the best prospects for each product
based on customer transaction data. They are not primarily interested in a behavioral
interpretation of the latent dimensions, but rather in a convenient low-dimensional graphical
display of the structure in the data. Edwards and Allenby (2002) propose a general approach to
handling identifying restrictions for the multivariate binomial probit model and develop an
algorithm that is efficient especially for a large number of response options. One of their three
applications is to survey data on ownership of financial products.2 Applicable to cross-sectional
data, the above papers are aimed at inferring cross-selling opportunities from comparison of one-
time measurements of ownership of other products across customers. The resulting next product
2 The proposed multivariate binomial model can be extended to incorporate time series and reflect a sequential set ofoutcomes.
8/2/2019 Sequential Xsell
5/28
5
to sell is usually consistent with the ranking of their market shares. This approach ignores the
over-time development of individual-level demands in favor of a cross-sectional approach. We
believe that to better understand the acquisition sequence of multiple products it is important to
formally model how customer demands for multiple products evolve over time. This approach
will allow us to predict acquisition sequence that is consistent with the development of consumer
demand maturity.
In this paper, we explicitly model the development of customer demand for multiple
products over time and derive a product acquisition sequence based on consumers individual
level of demand maturity. Using a multivariate probit framework applied to panel data, the
model is implemented in a Hierarchical Bayesian framework. In contrast to previous work in the
area, we seek the behavioral reasons that underpin and drive these purchase patterns. Our
empirical application will demonstrate the value of this approach.
2. An Individual-level Model of Sequential Service Acquisition
Consumers have demands for multiple products at any point in time. Economic theory has shown
that in the presence of finite resources, there is a priority structure among the demand
objectives and that this priority structure can be transformed into a priority structure for products
(Kamakura et. al. 1991). Our modeling approach positions multiple products and consumers
demand for those products along a common continuum in a way that reflects the development of
consumer demand maturity. The closer any individual consumers demand maturity is to the
position of a given product, the more likely this product is to satisfy the current need of the
consumer and the more likely he/she is to purchase the product. This follows the theoretical
foundation laid out by Kamakura et. al. (1991) which positions financial services and investors
financial maturity along the same continuum according to their ownership pattern.3 To formulate
this idea in a random utility framework suitable for our panel-data approach, we adopt the ideal
point model developed by Lehmann (1971) in which choice probability is postulated to be
inversely related to the distance of an object (ratings of the actual brands being analyzed) from
the ideal point (a consumers rating of an ideal brand).4 Formally, we assume that household i =
3 In Kamakura, Ramaswami and Srivastava (1991), a latent trait model is used to position products and consumersalong a common continuum in a way that reflects the ownership of the different services by each consumer. Theyallow the probability that a consumer owns a particular financial service to be a function of the consumers positionin this continuum relative to that particular product. See this paper for more detailed discussion.4 See Lehmann (1971) for detailed discussion.
8/2/2019 Sequential Xsell
6/28
6
1, ,Imakes binary purchase decisions (buy or not buy) on each of the productsj=1, , Jin the
product set J, at each time period t. We specify the latent utility of a given household choosing
productj = 1, 2, , Jat occasion tas:
ijtijtijitjiijt XDMOU ++= '|| 1 (1)
and the periodically realizedJpurchase decision variablesyijtare determined by
yijt= 1 ifUijt>0
0 otherwise. (2)
forall i=1, , I, j=1, , Jand t=1, , Ti. Oj forj=1, J is the positions of product j ranked
along the same continuum as demand maturity DMit-1. These are parameters to be estimated.
DMit-1 denotes the demand maturity of consumeri at the end of time t-1. The term || 1 itj DMO
represents the distance between demand maturity and product j. Thus, the probability of
purchasingj depends on the distance between Oj and DMit-1. i measures how the development
of consumer demand maturity, relative to productj, affects the demand for productj.
The variable DMit-1 is latent. However, demand maturity at time t-1 can be reasonably
proxied by some observed ownership or purchase information of each of the possibleJproducts.
These variables measure the demand currently satisfied for consumer i up to time t-1. For
example, the total number of purchases made in this product class, the total amount of money
spent or invested in these accounts, and the time elapsed since first owning each of the products.
The greater the number of purchases, the greater the amount of money placed in these accounts,
and the longer these accounts have been held the more likely it is that the individuals demand
maturity has reached or surpassed that point in the maturity continuum. This assumption
recognizes the marketplace reality that if a person has not had activity in a given set of accounts
for an extended time period he/she may have built up needs that have not been satisfied simply
because of myopia. On the other hand, if a customer has recently spent the fixed time costs
associated with changing their financial situation it is much more likely that their current account
holdings reflect their current state of demand maturity.
We assume the satisfied demand of product j is a linear function of these variables,
1
'ijtZ , where 1ijtZ is a vector ofL variables. Note that purchase activities of different products
contribute differently to the development of demand maturity. For example, a customer who
8/2/2019 Sequential Xsell
7/28
7
invests $5000 to a brokerage account is more likely to move further along the financial maturity
continuum than a customer who put $5000 to a checking account. In order to take into account
the fact that purchases of different products contribute differently to the development of demand
maturity, we weigh the contribution of the purchase of each product by its corresponding order
(or importance) Oj when constructing demand maturity.5 Formally,
)]([ 1'
1
11 =
= ijtJ
j
ijtjit ZDODM (3)
where Dijt-1is dummy variable indicating whether productj is owned by customeri at time t-1
The vectorXijt (K1) is a vector of product and consumer-specific covariates that affect
customers sequential decisions to purchase multiple products from the same company. It
includes variables designed to measure a customers relationship with the company. ij is a Kx1vector of coefficients forXijt. We allow the unobserved part of the utilities to be correlated.
],0[~ MVNit (4)
The correlations capture the co-incidence among different products as termed in Manchanda
et. al. (1999). It takes into account the contemporaneous purchases of different products at the
same time.
It is important to note here that the application of this model does not require the
managers a priori ordering of the products or services. The estimated coefficient pattern of Oj
will recover the natural ordering inherent in the data. Our model is flexible enough to handle a
situation where no managers initial beliefs about the ordering are available. In that case natural
ordering is inferred directly from the estimated parameters. Of course, if managers have strong
initial beliefs, which certainly may be true in some industries, these beliefs can be incorporated
into the prior distributions of the parameters and as such increase the efficiency of this modeling
approach. In our application, while we do have some initial beliefs, we will use noninformative
priors.
5 Note Equation (3) can also be written as )]([ 1'
1
121 =
+= ijtJ
j
ijtjitit ZDODMDM , where
1 ijtZ represents the change of variables represented by Z during the past period and
= =
1
1
1
t
ijijt ZZ
. This
equation shows that demand maturity evolves over time and is updated by the change inZof all the owned productsduring the past period.
8/2/2019 Sequential Xsell
8/28
8
To mitigate the well-known problems associated with not controlling for consumer-level
heterogeneity (Heckman 1981; Kamakura and Russell 1989; Krishna, Currim, and Shoemaker
1991; Gnl and Srinivasan 1996; Allenby and Rossi 1999) we specify i =(i1, i2, , iJ)and
ki=(ki1, ki2, kiJ) as linear functions of household specific variables.
ki
'
,
'
+=
+=
ikki
iii
W
eW(5)
fork = 1, , K. Wi is a m1 matrix containing m-1 household specific variables. and k are
mJmatrices with each element measuring the effect ofWivariable on i and ki respectively.
We assume eiN[0, 2] andkiMVN[0, ].
To solve the identification problem, we normalize one of Oj to be 1.6 We divide the
utility by its corresponding standard deviation and thus transform
to a correlation matrix. Weuse a hierarchical Bayesian approach to estimate the model. Readers interested in the details of
the estimation procedure are referred to Gelfand and Smith (1990), Allenby and Rossi (1999),
and Manchanda, Ansari and Gupta (1999).7
3. Customers Sequential Acquisition of Banking Services
3.1 Data Description
[Insert Table 1a About Here]
We use this model and household-level data collected from a large Midwestern bank to
explore consumers acquisition of financial services. Specifically, we have holding and
transaction information for 8 products of 1201 randomly selected households whose use this
bank as primary bank from July 1997 to June 1998.8 This data includes monthly observations on
which accounts/products the customer had purchased or held as an investment. We also had
access to demographic information for each of these customers. Finally, the bank provided us
6 For identification purpose, one ofOs needs to be normalized to a non-zero value. We haveJequations to identifyJ
Os. The normalized value affects the scales of without affecting the estimated sequential order.7 We have conducted an intensive simulation to study the identification and statistical properties of this model. Wehave also compared the scale of the estimated product ranking coefficients with the estimated demand maturity
based on both simulated data and our data. The two variables are measured at similar scales and the distancesbetween product positions and demand maturity explain purchases very well. Interested readers can contact theauthors for details on the simulation, prior distributions and the estimation of the model.8 The bank gave us access to holding and transaction information for 20 financial products it offered for 1201households. However, there are 11 products without any purchase or very little purchase information in this shortobservation window. There is also 1 product with only 16 purchase observations. In order to avoid the problem ofdata scarcity, we only focus on products with more than 29 purchases in the observation period.
8/2/2019 Sequential Xsell
9/28
9
with the results of a customer satisfaction survey completed by each of the customers in the
sample just before July 1997. Because our data was at the household level, we also observed
repeat purchases. Because we cannot determine whether these repeat purchases represent true
repeats by the same individual or new purchases by someone else in the household, our analysis
is at household level. A brief description of the variables used in this paper is shown in Table 1a.
[Insert Table 1b About Here]
In Table 1b, we report a conditional table which shows how many purchases of product j
being made during the observation window given previous purchases of all the other products.
We also report the total number of current ownerships and new purchases of each of the 8
products. For example, we observe that owning saving, the customer usually go for CD or
another saving account.
3.2 The Specifics of Consumer Banking Service Acquisition Model
We write the utility function of any household i for banking servicej at time tas:
ijtitijiijiij
itjiijt
SWITOVERSATCOMPET
DMOU
+++
+=
321
1 ||(6)
forj=1,..8. We define COMPETi as a dummy variable equal to one when household i has opened
an account in category j with another bank in the past six months. It controls for competition,
which directly affects the possibility of cross-buying more services from the same company. Wedefine OVERSATi as the overall satisfaction score of household i, as measured by the banks
customer satisfaction survey. This represents the banks efforts in improving service quality and
the relationship with its existing customers. Prior research has concluded that is an effective
marketing strategy for service industry (Boulding et. al. 1993).
We also included household-level switching costs (SWITit) in our analysis. It is defined as
one if the account owners profession is white collar, and the household has at least one non-
adult child, and the household owns more than the average number of accounts with this bank.
The logic of this inclusion follows closely the reasoning of Blattberg, Buesing, Peacock, and Sen
(1978), and more primitively the work of Becker (1965). In short, we believe that households
that have a high cost of time, as evidenced by increasing levels of education or the presence of
children, will tend to spend less time shopping around for banking services. Further, individuals
8/2/2019 Sequential Xsell
10/28
10
that have a relatively complex relationship with the bank (have a large number of accounts there)
will also suffer greater inconvenience if they choose to switch banking service providers. While
switching grocery stores is relatively easy, switching service providers can often entail a
significantly higher degree of discomfort. The reasons for this greater inconvenience are
reasonably self-evident, but include such unpleasantness as filling out all of the paperwork
necessary for opening up a series of accounts at a new bank. Therefore, our measure of switching
cost takes into account both intrinsic time costs as well as extrinsic costs associated with the
customers current level of relationship with the bank.
We define the financial maturity as a function of cumulative ownership, monthly balance,
and holding time of all available accounts weighted by corresponding importance of each
product. That is,
)]([ 1312111
11 =
++= ijtijtijtJ
j
ijtjit HoldingBALACCTNBRDODM
Since the data is collected at household level, multiple purchases of the same account can be
opened under the names of different household members. The term ACCTNBRijt-1 captures the
cumulative number of purchases of productj for household i in the past month. Previous research
has shown the importance of using ownership as a predictor variable (see Knott, Hayes and
Neslin 2002). In addition, we also include the monthly balance (BALijt-1) and time elapsed since
first opening this type of account (Holdingijt-1) as additional explanatory variables to better takeinto account the fulfillment of demand for each product. In order to obtain more easily
interpretable estimates ofOj, we rescale all three variablesto have zero mean and unit variance.
These three variables indicate the purchase experience or satisfied needs for product j and can be
interpreted as the realization of consumer demand forj at the beginning of time t. Cumulative or
satisfied demands for all of the products is then weighted by product order to construct the latent
variable, the financial maturity of household i at time t.
We write the utility coefficients as linear functions of basic demographic information.
This allows for household characteristics to influence the weight certain factors will play in the
utility function. Formally, we specify coefficient heterogeneity as:
8/2/2019 Sequential Xsell
11/28
11
+++++=
+++++=
kiikikikikkki
iiiiii
INCOMEAGEGENDEREDUCAT
eINCOMEAGEGENDEREDUCAT
43210
43210 (8)
fork=1, 2 and 3. The coefficients 1 , 2 and 3 capture how different customers have different
information costs associated with adopting product j. 4 reveals the resource constraint. For
example, we believe that better educated and male household heads believe themselves to be
more knowledgeable and better informed about the financial products. They are also more
confident in managing their investments (See Barber and Odean 2001). Thus, we conjecture that
the adoption speed or movement along the financial maturity continuum may be faster for
households that are better educated and with a male head. Higher income individuals may move
along the continuum more quickly because they are more motivated to pay attention to and find
solutions for their financial needs. Similarly, different households may also respond differently
to competition, satisfaction and switching cost.
In the interest of comparing our model to other plausible specifications, we estimate four
competing models and our own. The first is a nave first order Markov model using the sample
conditional purchase probabilities (see Table 1b). Assuming independence among products, it
predicts purchases in the next period based on current ownership. The second model also
assumes the product choices are independent. This model specification is comprised of product
specific binary probit models, one for each category. This particular model specification ignores
potential unobserved correlation across categories, the allocation of assets, switching costs and
heterogeneity. We include this model because this specification is commonly adopted by the
industry to predict probabilities of cross-selling. This model is also similar to Kamakura,
Ramaswami, and Srivastava (1991) and Knott, Hayes, and Neslin (2002) in the sense that they
assume consumers make independent decisions about the purchase of different products. The
third model is a multivariate probit model that is similar to Manchanda, Ansari, and Gupta
(1999). It captures the contemporaneous (non-over-time) complementary relationship among
products on the same purchase occasion. The fourth model adds to the third model thedifferences in intrinsic preferences. Although not directly comparable, this model can be likened
to Edwards and Allenby (2002) in the sense that the predicted purchase order is derived from the
order of intrinsic preferences. The final model is our proposed model. It adds demand maturity to
the fourth model and thereby allows us to take into account customers previously satisfied
8/2/2019 Sequential Xsell
12/28
12
financial needs. The Independent model was estimated using the Probit procedure in SAS. The
remaining four models were estimated using an MCMC (Gibbs Sampler) approach. We use the
Bayes factor criterion for model comparison.
[Insert Table 2 about here]
We report the Deviance Information Criterion (DIC) which penalizes a complex model for
additional parameters (Spiegelhalter, Best, Carlin, and Linde 2002). A lower DIC indicates
better fit. As shown in Table 2, we find that all models that allow products to be correlated rather
than independently acquired (Model 3, 4, and 5) fit better than those that do not (Model 1 and 2).
Model 5 fits better than Model 3 and 4 indicating that taking into account contemporaneous
complementary demand and/or using brand constants is not adequate in capturing sequential
order in our application. The sequential order of financial products is different from market
share. Our proposed model (Model 5) fits the data better than any of the competing models and
demonstrates that taking into account the development of customer financial demand is
important in determining service adoption.
3.3The Substantive Results of our Consumer Banking Application
[Insert Table 3a About Here]
Now we want to turn our attention to our application itself and the substantive results that
arise from it. We first look at the ordering among products by examining the ranking of Oj
reported in Table 3a. The ranking of the parameters indicate a product sequence of checking-
>saving->debit-card->credit-card->installment loan->CD->money market (MM)->brokerage.
This is consistent with our intuition that customers usually invest more aggressively into
financial instruments promising stable returns (e.g. CD, MM) after they obtain basic financial
services (e.g. checking, saving, debit, credit, loan) and into high risk and high return brokerage
accounts only after they have invested in other very low risk products. The implied order of the
eight financial products is consistent with Kamakura, Ramaswami and Srivastavas (1991) idea
of the development of financial maturity from convenience products to stable income products
and then to risky income products. The ordering that was recovered by our model is exactly what
8/2/2019 Sequential Xsell
13/28
13
we a priori expected. It is also consistent with our conversations with bank managers as well as
the hierarchy suggested by Kamakura, Ramaswami, and Srivastava (1991).
The slope parameter indicates how the distance between the financial maturity and the
corresponding product affects the probability of purchase as the customers latent financial
maturity changes. The sign of the coefficient is negative indicating that the closer the distance to
the product, the more likely the product is to be purchased. This is consistent with the
interpretation of distance model by Lehmann (1971). 1, 2 and 3 are all positive and
significant indicating that the cumulative number of account purchases, cumulative balance and
holding time increase financial maturity.
Now referring to the other parameter estimates presented in Table 3a, notice that all
estimated coefficients are significant in the classical sense. Owning any account with another
financial institution takes away business from the bank. This effect is more powerful for higher-
end products than for lower-end products. This is probably due to the fact that the marketplace
for brokerage and other advanced financial services has grown increasingly competitive in recent
years. Whereas it is possible to purchase products from full services brokerage houses like
Schwab and Fidelity that mimic some aspects of a checking account, it is even easier to purchase
products that substitute, or dominate, the more advanced financial service offerings of a
consumer bank.
Overall satisfaction with the bank increases the banks ability to cross-sell its products toexisting customers, and this service quality has a higher impact on the future demand for
advanced financial services than for lower end products. This result is intuitive in that advanced
financial services, brokerage for example, require much more interaction with the bank than say
a CD account which requires almost no human interaction.
Switching costs, which indicate the customers relationship built over the year as well as
the time value of the households, plays an important role in providing the company with
opportunities to cross-sell their other products. These Becker-type effects are generally important
in markets where changing product or service providers is particularly costly in terms of the time
needed to complete the change. Switching costs play a powerful role in influencing households
in all eight product categories. If time costs are high, customers are unlikely to switch banks even
in the face of some amount of dissatisfaction. Following the logic above, we would expect this
effect to be more potent for convenience and advanced services than for cash reserves and indeed
8/2/2019 Sequential Xsell
14/28
14
it is. This result does raise some interesting questions about which customers are most important
for the bank to satisfy. While conventional wisdom would certainly dictate placing a great deal
of emphasis on satisfying the banks wealthier customers our results indicate that some
customers may be trapped by the bank owing to the substantial implicit costs a given customer
might face in switching from one bank to another.
[Insert Table 3b About Here]
Table 3b provides the estimated correlations among the unobserved components of the
utilities. This is the Co-incidence correlation termed in Manchanda, Ansari, and Gupta (1999).
A positive correlation between two products indicate that there are complementary demands for
these two products at each purchase occasion. For example, a household may open checking and
ask for a debit card together simply because they are bundled by the bank. Similarly, if a bank
can sell a customer a saving account, it is more likely to also sell the customer a brokerage
service. Note that these correlations capture the current inter-product relationship. This is
different from the sequential demand for naturally ordered products driven by financial maturity
and knowledge.
[Insert Table 4 About Here]
We now turn our attention to the consumer heterogeneity expressions and examine how
demographic variables like education, gender of the head of the household, age and income
affect the acquisition pattern of sequentially ordered products. Table 4 reports the effects of these
variables on utility weight heterogeneity. Most of the estimates are significant. These estimates
describe how knowledge, risk bearing, life stage and financial resources characterize the speed of
adoption for financial products. The negative coefficients of education and gender in the
heterogeneity equation fori indicate that, for the same amount of financial maturity (the same
distance between financial maturity and the position of checking), greater education or male
household heads are more likely to move faster along the financial continuum than other
customers. These results show that households with greater education or male-headed
households are more aggressive in pursuing products that are more advanced in the sequential
ranking. These results are consistent with the research findings of Barber and Odean (2001) who
find that men are more confident about their ability of managing risky assets. Similarly, we
found that older customers and higher income customers are also more likely to move faster
along the continuum. This is because they have more resources to purchase multiple products.
8/2/2019 Sequential Xsell
15/28
15
The coefficients of education in the heterogeneity equation for the coefficient of competition
directly suggests that those households are less likely to be cross-sold once they open accounts with the
banks competitors. It appears that highly educated households are, or believe themselves to be,
more knowledgeable and confident about financial products and hence are more likely to take
advantage of these outside opportunities. Interestingly, more highly educated customers seem to
care more about customer service and switching costs as determinants of future purchases. They
are in general more sophisticated customers who are also busier than average customers. Owning
accounts from other institutes does not deter households with male household head from opening
other accounts with this bank except debit card and saving. Male household heads seem to care
less than their female counterparts about both satisfaction and switching cost in determining
future purchases. This result may be at least partially explained by the impact of education and
gender on a customers organizational ability, an ability which would certainly impact theirsensitivity to service quality and propensity to switch, or patronize multiple, financial
institutions. In a completely different application, Narasimhan (1986) showed that the education
of the female head-of-household was critical in determining sufficient organizational skills to
effectively use coupons in grocery settings. While the applications are quite different, we suspect
that the logic that governs behavior may well be the same.
Owning accounts in other financial institutions is more likely to deter older customers
from cross-buying other products from the same bank than average customers due to the fact that
older customers have more time to spend to finding the best financial service provider. These
same customers are much more sensitive to satisfaction and less sensitive to switching cost. This
is a potentially important finding as older customers also have the potential to be some of the
banks best customers. Taken together, these results strongly suggest that making attempts to
satisfy older customers should be important to the bank.
Following the same reasoning and also arising from our estimated model, higher income
customers have less time to spend searching for the best product, service, or the lowest price.
They are more likely to be cross-sold other convenience products once they become customers of
the bank. They are less sensitive to satisfaction and more sensitive to switching costs on these
products. These are customers who are somewhat trapped and seek convenience to save time.
However, for non-convenience products like CDs, money market accounts and brokerage
accounts, these customers are the opposite, they are more sensitive to competition and
8/2/2019 Sequential Xsell
16/28
16
satisfaction. We believe that this is because higher income customers are more sensitive to the
return on investment and to the availability of specific investment choices. They are also more
likely to have better knowledge of the options provided by competing banks.
3.4 Predicting Cross-Selling Opportunities
Because part of the focus of this application is to predict the probability of acquiring new
products in order to help bank managers more efficiently allocate their targeting efforts, we now
compare the predictive ability of our proposed model with the other benchmark models. To
accomplish this we divided our original sample of 1201 households into an estimation sample
and a holdout sample. The estimation sample has three quarters of the households, and the
holdout sample has the remaining quarter.
[Insert Figure 1a and Figure 1b about here]
Figure 1a and Figure 1b depict the mean absolute error between the purchase probability
and the actual purchase realization for each of the five models using the estimation sample and
holdout sample, respectively. The worst predictive accuracy arises from the Independent Model
(Model 2). Including co-incidence to allow for co-purchase of multiple products and
heterogeneity (Model 3) increases the predictive accuracy. Thus, the approach suggested by
Manchanda, Ansari, and Gupta (1999), although developed in studying complementary demands
(shopping basket) for frequently purchased products, does indeed capture unobserved factors that
cause multiple purchases at the same purchase occasion in this setting. Comparing Model 4 with
Model 5, we find that a multivariate probit model with sequential ordering effects provides a
much more accurate description of households future purchase decisions. The predictive
accuracy of our proposed model is fairly remarkable. For example, the mean absolute error rate
of convenience services is less than 0.5 percent. While we do not claim that all applications will
achieve such a high degree of accuracy, we believe that the evidence does weigh in favor of
including order effects in attempts to formally model choice in this type of environment.
[Insert Figure 2 Here]
Given the CRM nature of our application, a gains chart is a common approach in
evaluating predictive performance when there are low response rates. Figure 2 includes gains
charts for each of the tested models. The gains chart approach to model evaluation and
comparison begins by selecting the 10 percent of the customers from our holdout sample who,
8/2/2019 Sequential Xsell
17/28
17
according to prediction from the model, are most likely to make a purchase. This is done for each
model. Next we compute the number of accurately predicted responses, relative to the total
number of responses in the entire sample; this percentage is the gain due to using the model and
represents the value of sorting the data via the predictions from the models. The No Model line
depicts a situation in which customers are grouped randomly. So, if we have a randomly drawn
group that constitutes 10% of the sample it will on average contain 10% of the overall number of
purchases etc. Analogous values are computed for each percentile of the holdout sample (top
20%, top 30% etc.). The greater the difference between the gains curve and the baseline model,
the better the model. The results of this analysis systematically support the conclusion that the
predictive accuracy of our proposed model is superior to that of the baseline case and each of the
competing models examined. The results are robust at each percentile.
[Insert Table 5 About Here]
Although our main objective is to derive the natural product order to guide the sequence
of cross-selling, our method also allows us to estimate the predicted purchase probability for
product-next-to-buy for each household. For demonstration purposes, we report the cross-selling
probabilities for the first two periods of five randomly selected households (Table 5). We can see
that the cross-selling probability predictions yielded from our proposed model are quite
consistent with the actual purchase data. These provide managers with information on whom and
when to target in order to cross-sell which products at time t. Note that the purchase predictions
on and beyond t+2 depend on prediction of ownership information on and beyond t+1. This
implies that our model will be more accurate in predicting next-to-purchase rather than purchase
incidence more distant in the purchase sequence.
In summary, the most important things we have learned from our application are:
The sequential purchase of multiple products are driven by a households evolving
financial maturity.
Households with higher education or male household heads move more quickly along the
financial maturity continuum than those with lower education or female household heads.
Households that are older and with higher income also move along this continuum more
quickly.
The switching costs created by owning multiple products creates opportunities to cross-
sell other products to the same customer.
8/2/2019 Sequential Xsell
18/28
18
Customer satisfaction or service quality is a significant influencer of a customers future
purchase decisions, especially for more advanced financial products (e.g. brokerage).
This effect is particularly strong for older customers.
A multivariate probit model with heterogeneity, switching cost and sequential ordering
effects provides the most accurate description and prediction of households future
purchase decisions.
4. Discussion, Limitations, and Directions for Future Research
This research explicitly models the development of customer demands for multiple
products over time and derives a product purchase sequence under the assumption that different
products are designed to meet the requirements of customers at different stages of their demand
maturity. Being the first cross-selling model applicable to panel data, this paper provides a
behavioral explanation for the over-time development of consumer demand and for purchasing
multiple products from the same provider. Our research also shows that sequential demands for
naturally ordered products develop differently across different customers and points to some of
the antecedents of these differences. It provides guidance to managers charged with allocating
marketing dollars towards customers with the greatest incremental profit potential.
Our model directly suggests that it is important for a company to collect information to
determine their customers stages of demand maturity before spending money to customers whomay not be ready to adopt a new product or service. In addition, it appears that in this particular
market it is more effective to target customers with higher education, males and those with
higher incomes as these customers move faster along the demand maturity continuum and are
faster in adopting new products. Though not the most important factor, increasing the depth of
the customer relationship significantly enhances the cross-selling opportunities, especially for
customers who need some financial advice (e.g. older customers).
Our banking application uncovered several interesting behavioral findings that may be
instructive of consumers service adoption more generally. We say may here because the
study was limited in several ways. First, we do not have detailed information on competition. In
the future, it would be interesting to consider the ownership information with competitors when
calculating demand maturity. This would allow us to derive a more accurate ordering of the
products. Second, while we were able to obtain relatively detailed information on customer-level
8/2/2019 Sequential Xsell
19/28
19
account activity, we did not have access to data detailing the marketing activity that each
individual in our sample was exposed to over this time period. If such information were
available, we could more formally explore the impact of marketing variables (e.g. price,
advertising) on cross-selling tactics and opportunities. This would be a valuable contribution to
our knowledge of these markets. Third, our data only covered a time span of one year. In an
environment such as banking in which a customer relationship may last for many years, but new
service acquisitions are relatively infrequent, one year may not be sufficient to capture the
richness of the phenomenon. We expect that data that covers a longer time span would yield
even greater differences between our modeling approach and those that do not account for
temporal ordering. Finally, a product or service provider is not only interested in whether or not
she is likely to sell an additional service to an existing customer but also in the amount of profit
such a sale is likely to generate. We did not have access to data that would allow such analysis.
Future research should more explicitly model the flow of funds, type of usage and cost to
examine the impact of cross-selling on profit.
The market for services is a huge and growing segment of the U.S. economy. Even the
more narrowly defined market for financial services is a truly enormous segment of the
economy. We have seen the benefits of predictive modeling in the arena of packaged goods. The
state of our knowledge in this market has grown dramatically over the past two decades.
Conversely, disproportionately little attention has been paid to the market for services. There is
no doubt that this neglect is at least partially explained by the relative difficulty of obtaining
quality and timely data from service providers. Data from service providers tends to be more
convoluted than data generated by scanner technology. Yet, in spite of these challenges, the
benefits of exploring this marketplace are truly remarkable. We believe our research is one step
in that direction.
8/2/2019 Sequential Xsell
20/28
20
Table 1a. Operationalization of Variables
Variables Definitions
Mean
or Freq
COMPET 1 if the household opens some accounts in another bank during the last six months,0 otherwise
13.3%
OVERSAT Households overall satisfaction (1 to 7) 4.33SWIT 1 if the account owners profession is white collar and the household has at least on
non-adult child and the household owns more than average number of accounts withbank, 0 otherwise
6.6%
EDUCAT The household heads education measured on a 1-5 scale(1 = some high school , , 5 = post graduate
1.21
GENDER 1 if the account owner is a male, 0 otherwise 82.5%
AGE Account owners age 67.01INCOME Households income (1-7) where 1 -
8/2/2019 Sequential Xsell
21/28
21
Table 2. Model ComparisonDIC Log-Marginal
Density
Model 1 20249.9 -10118.1Model 2 22466.0 -11226.4Model 3 20189.9 -10098.0Model 4 20096.7 -10046.4Model 5 18327.3 -9156.3
8/2/2019 Sequential Xsell
22/28
22
Table 3a. Estimation Results of Proposed Model
Product / Covariate COMPET OVERSAT SWIT D. MATURITYRELATIVE TO
EACH
PRODUCT
ORDER (Oj)
Installment Loan -11.60(0.062)
2.55(0.002)
2.55(0.098)
-0.018(0.003)
0.71(0.030)
Debit Card -9.59(0.026)
2.01(0.003)
2.38(0.100)
-0.018(0.003)
0.57(0.028)
Checking -12.04(0.049)
2.77(0.003)
4.40(0.119)
-0.018(0.003)
-0.09(0.032)
Credit Card -8.92(0.065)
2.49(0.007)
1.39(0.095)
-0.018(0.003)
0.69(0.025)
Saving -11.21(0.079)
2.77(0.006)
2.71(0.176)
-0.018(0.003)
-0.03(0.002)
CD -8.40(0.068)
2.33(0.008)
1.80(0.079)
-0.018(0.003)
0.79(0.108)
Money Market -9.54(0.051)
3.10(0.004)
1.20(0.164)
-0.018(0.003)
0.84(0.030)
Brokerage -14.63(0.064)
3.44(0.005)
5.05(0.061)
-0.018(0.003)
1
Estimates ACCTNBR BAL DURATION
D.Maturity
0.135(0.025)
0.028(0.002)
0.057(0.011)
Table 3b. Unobserved Correlations Across Categories
Installment
Loan
Debit
Card
Checking Credit
Card
Saving CD Money
Market
Brokerage
InstallmentLoan
1 0.005(0.013)
0.045(0.012)
-0.219(0.022)
-0.024(0.011)
0.264(0.034)
-0.075(0.020)
-0.125(0.009)
Debit Card 1 0.199(0.025)
0.135(0.019)
0.194(0.018)
0.108(0.031)
0.098(0.013)
0.119(0.011)
Checking 1 0.170(0.018)
0.246(0.013)
0.097(0.031)
0.099(0.013)
0.092(0.012)
CreditCard
1 0.259(0.012)
-0.063(0.012)
0.176(0.027)
0.217(0.014)
Saving 1 0.100(0.014)
0.093(0.020)
0.088(0.014)
CD 1 0.114(0.021)
0.047(0.007)
MoneyMarket
1 0.343(0.015)
Brokerage 1
8/2/2019 Sequential Xsell
23/28
23
Table 4. Estimation Results for Heterogeneity EquationsHeterogeneity Equation -
i : coefficient of DM
Intercept Education Gender Age Income
-0.05(0.014)
-0.03(0.007)
-0.09(0.018)
-0.06(0.001)
-0.01(0.003)
Heterogeneity Equation - i1 : coefficient of COMPET
Intercept Education Gender Age Income
InstallmentLoan
-9.94(0.053)
-3.04(0.107)
0.20(0.080)
-0.04(0.0010)
1.35(0.016)
Debit Card -10.85(0.047)
-2.46(0.037)
-1.52(0.089)
-0.01(0.0008)
1.46(0.019)
Checking -6.10(0.079)
-2.61(0.040)
2.76(0.127)
-0.10(0.0013)
1.01(0.015)
Credit Card -17.98(0.033)
-0.40(0.080)
0.45(0.116)
-0.02(0.0015)
2.97(0.028)
Saving -1.21(0.061)
-0.73(0.039)
-3.37(0.124)
-0.15(0.0012)
0.10(0.020)
CD 5.28(0.065)
-4.00(0.078)
6.27(0.118)
-0.10(0.0011)
-1.13(0.019)
MoneyMarket
-9.83(0.051)
-1.80(0.055)
0.07(0.150)
0.04(0.0013)
-0.11(0.026)
Brokerage -16.74(0.040)
1.71(0.041)
2.40(0.086)
0.02(0.0007)
-0.52(0.016)
Heterogeneity Equation -i2 : coefficient of OVERSAT
Intercept Education Gender Age Income
InstallmentLoan
2.48(0.039)
0.71(0.032)
-0.08(0.114)
0.002(0.0011)
-0.21(0.022)
Debit Card 1.11(0.049)
0.46(0.033)
0.09(0.129)
0.007(0.0012)
-0.03(0.022)
Checking 4.23(0.055)
0.17(0.034)
-0.60(0.176)
0.003(0.0009)
-0.39(0.027)
Credit Card 3.22(0.042)
0.03(0.032)
-0.88(0.069)
0.002(0.0006)
-0.23(0.012)
Saving 2.60(0.104) 0.25(0.038) -0.55(0.130) 0.011(0.0013) -0.23(0.031)CD 2.01
(0.061)
0.16
(0.040)
0.39
(0.036)
-0.004
(0.0004)
0.09
(0.011)MoneyMarket
3.51(0.207)
0.16(0.045)
-0.94(0.104)
-0.007(0.0014)
0.02(0.044)
Brokerage 2.99(0.106)
0.22(0.043)
-1.15(0.088)
-0.014(0.0009)
0.25(0.046)
Heterogeneity Equation -i3 : coefficient of SWIT
Intercept Education Gender Age Income
InstallmentLoan
3.31(0.119)
0.74(0.097)
-7.19(0.113)
-0.022(0.001)
0.38(0.022)
Debit Card 7.53(0.134)
1.30(0.061)
-11.77(0.294)
-0.099(0.001)
3.82(0.030)
Checking 2.43(0.134)
1.32(0.111)
-14.28(0.196)
-0.062(0.002)
0.44(0.030)
Credit Card 7.63(0.055)
0.50(0.063)
-13.92(0.085)
-0.032(0.002)
0.29(0.030)
Saving 6.42(0.148)
0.68(0.083)
-5.85(0.272)
-0.108(0.001)
1.68(0.047)
CD 4.10(0.045)
2.94(0.035)
-15.53(0.100)
-0.028(0.001)
0.62(0.012)
MoneyMarket
2.96(0.103)
1.08(0.067)
-11.67(0.141)
-0.053(0.001)
2.61(0.038)
Brokerage 0.79(0.076)
1.95(0.075)
-12.62(0.209)
0.051(0.001)
2.01(0.013)
8/2/2019 Sequential Xsell
24/28
24
Figure 1a: Mean Absolute Error Across M odels
- Estimation Sample
0
0.01
0.02
0.03
0.04
0.05
0.06
Model 1 Model 2 Model 3 Model 4 Model 5
Figure 1b: Mean Absolute Error Across Models- Holdout Sample
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
Model 1 Model 2 Model 3 Model 4 Model 5
Figure 2: Gains Chart for Model Comparison Across All Product Categories
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
0% 10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
% of Customers
Cu
mGains No Model
Model 1
Model 2
Model 3
Model 4
Model 5
8/2/2019 Sequential Xsell
25/28
25
Table 5. Selected 10 Households for Cross-Selling Efforts (Proposed Model)
Actual Data Predicted ProbabilitiesHousehold
t
IL DBC CHECK CRC IL DBC CHECK CRC
1 0*(0)
0(0)
0(0)
0(1)
0.0121 0.0102 0.0001 0.00081
2 0(0)
0(0)
0(0)
0(1)
0.0105 0.0093 0.0001 0.0008
1 0(1)
0(0)
0(0)
0(0)
0.0099 0.0044 0.0079 0.00042
2 0(1)
0(0)
0(0)
0(0)
0.0097 0.0034 0.0067 0.0005
1 0(0)
1(1)
0(2)
0(3)
0.0001 0.1931 0.0147 0.17823
2 0(0)
0(2)
0(2)
1(3)
0.0004 0.1322 0.0334 0.2386
1 0(0)
0(3)
0(3)
0(2)
0.0005 0.0168 0.0061 0.00044
2 0
(0)
1
(4)
0
(3)
0
(2)
0.0001 0.1485 0.0052 0.0001
1 0(0)
0(2)
0(1)
0(2)
0.0002 0.0492 0.0001 0.00695
2 0(0)
0(2)
0(1)
0(2)
0.0016 0.0147 0.0014 0.0017
Actual Data Predicted ProbabilitiesHousehold
t
SAV CD MM BRK SAV CD MM BRK
1 0*(0)
0(0)
0(0)
0(0)
0.0007 0.0173 0.0001 0.00001
2 0(0)
0(0)
0(0)
0(0)
0.0006 0.0168 0.0001 0.0000
1 0(0)
0(0)
0(0)
0(0)
0.0014 0.0079 0.0028 0.00072
2 0(0)
0(0)
0(0)
0(0)
0.0015 0.0065 0.0026 0.0006
1 0(1)
0(0)
0(0)
0(0)
0.0004 0.0223 0.0001 0.00163
2 0(1)
0(0)
0(0)
0(0)
0.0005 0.0301 0.0001 0.0015
1 0(0)
0(0)
0(0)
0(0)
0.0000 0.0008 0.0000 0.00424
2 0(0)
0(0)
0(0)
0(0)
0.0001 0.0005 0.0000 0.0053
1 0
(0)
1
(1)
0
(0)
0
(0)
0.0001 0.5644 0.0000 0.00075
2 0(0)
1(2)
0(0)
0(0)
0.0002 0.5743 0.0000 0.0008
* The number 0 or 1 indicates whether the household purchases (1) some service in the corresponding category ornot (0) during the sample period. The numbers in the parentheses are the numbers of accounts that the householdopened in the corresponding category with the bank before the current period.
8/2/2019 Sequential Xsell
26/28
26
References
Ainslie A., and P. Rossi (1998), Similarities in Choice Behavior Across Product Categories, MarketingScience, 17(2): 91-106.
Allenby, Greg M. and Peter E. Rossi (1999), Marketing Models of Consumer Heterogeneity, Journal of
Econometrics, 89: 57-78.
Barber B. and T. Odean (2001), Boys will be Boys: Gender, Overconfidence, and Common Stock Investment,"Quarterly Journal of Ecomomics, February 2001, Vol. 116, No. 1, 261-292
Becker, G. (1965), A Theory of the Allocation of Time,Economic Journal, 75: 493-517.
Bitner, Mary Jo and Valarie Zeithaml (2000).Services Marketing. McGraw-Hill, 2000.
Blattberg, R., T. Buesing, P. Peacock, and S. Sen (1978), Identifying the Deal Prone Segment, Journal ofMarketing Research, 15, 3: 369-377.
Blattberg, R., P. Peacock, and S. Sen (1976), Purchasing Strategies Across Product Categories, Journal ofConsumer Research, 3: 143-154.
Boulding, William, Ajay Karla, Richard Staelin and Valarie Zeithaml (1993), "A Dynamic Process Model ofService Quality: From Expectations of Service,"Journal of Marketing Research, Vol. 30, 1; pg. 7-2.
Chen, Y., J. Hess, R. Wilcox, and Z. Zhang (1999), Accounting Profits versus Marketing Profits: A RelevantMetric for Category Management,Marketing Science, 18(2), 129 -143
Chiang J., and R. Wilcox (1998), A Cross-Category Analysis of Shelf-Space Allocation, Product Variety, andRetail Margins,Marketing Letters, 8(2): 183-191.
Clarke, Y. and G. N. Soutar (1982), Consumer Acquisition Patterns for Consumer Durables: Some AustralianEvidence,Journal of Consumer Research, 8: 456-460.
Cunningham, R. (1956), Brand LoyaltyWhat, Where, and How Much, Harvard Business Review, 34,JanuaryFebruary: 116-128.
Dickson, P. R., R. F. Lusch and W, L. Wilkie (1983), Consumer Acquisition Priorities for Home Appliances:A Replication and Reevaluation,Journal of Consumer Research, 9(1): 432-435.
Drozdenko, R.G. and P.D. Drake, Optimal Database Marketing: Strategy, Development, and Data Mining,(Sage Publications 2002).
Edwards, Y. D. and G. Allenby (2000), Multivariate Analysis of Multiple Response Data, under review atJournal of Marketing Research.
Erdem and Sun (2000), An Empirical Investigation of Spillover Effects of Marketing Mix Strategy inUmbrella Branding, forthcoming,Journal of Marketing Research.
Fader, P., and L. Lodish (1990), A Cross-Category Analysis of Category Structure and Promotional Activityfor Grocery Products,Journal of Marketing, October: 52-65.
8/2/2019 Sequential Xsell
27/28
27
Gelfand, A E., and A.F. M. Smith (1990), Sampling-Based Approaches to Calculating Marginal Densities,Journal of the American Statistical Association, 85, 410: 398-409.
Gonul, Fusun, and Kannan Srinivasan (1996), Estimating the Impact of Consumer Expectation of Coupons onPurchase Behavior: A Dynamic Structural Model,Marketing Science, 15, 3: 262-279.
Hauser, J.R. and G. L. Urban (1986), The Value Priority Hypotheses for Household Budget Plans, Journal ofConsumer Research, 12: 446-462.
Hebden, J.J., and J. F. Pickering, Patterns of Acquisition of Household Durables, Oxford Bulletin ofEconomics and Statistics, (1974): 67-94.
Heckman, James J., Dummy Endogenous Variables in a Simultaneous Equation System,Econometrica, 46, 4(1978): 931-959.
Heckman, James J., The Incidental Parameters Problem and The Problem of Initial Conditions in Estimating aDiscrete Time-Discrete Data Stochastic Process, in Structural Analysis of Discrete Data with Econometric
Applications, edited by C. F. Manski and D. McFadden, (Cambridge: MIT Press 1981) 179-195.
Information Week, Outlook for 2001, January 8, 2001.
Jeffries H., Theory of Probability (3rd ed.), (Oxford, U.K., Oxford University Press 1961).
Kamkura, Wagner A., and Gary Russel, A Probabilistic Choice Model For Market Segmentation andElasticity,Journal of Marketing Research, 26, 4, 1989: 379-391.
Kamakura W., S. Ramaswami, and R. Srivastava (1991), Applying Latent Trait Analysis in the Evaluation ofProspects for Cross-Selling of Financial Services,International Journal of Research in Marketing, 8: 329-349.
Kamakura, Wagner A. and Bruce S. Kossar (2001), Identifying Innovators for the Cross-selling of NewProducts, working paper, Duke University.
Kamakura, Wagner A. M. Wedel, F. de Rosa and J. A. Mazzon (2002), Cross-selling Through DatabaseMarketing: A Mixed Data Factor Analyzer For Data Augmentation and Prediction, forthcoming,InternationalJournal of Marketing Research.
Kass, R., and A. Raferty (1995), Bayes Factors,Journal of the American Statistical Association, 90, 430: 773-795.
Kasulis, J.J., R. F. Lusch and E. F. Stafford (1979), Consumer Acquisition Patterns for Durable Goods,Journal of Consumer Research, 6(2): 47-57.
Kim, B., K. Srinivasan, and R. Wilcox (1999), Identifying Price Sensitive Consumers: The Relative Merits ofDemographic vs. Purchase Pattern Information,Journal of Retailing, 75(2): 173-193.
Knott Aaron, Andrew Hayes and Scott A. Neslin (2002), Next-Product-to-Buy Models for Cross-SellingApplications, working paper, Dartmouth College.
Krishna, Aradhna, Imran S. Currim and Robert W. Showmaker (1991), Consumer Perceptions of PromotionalActivity,Journal of Marketing, 55, 2: 4-17.
8/2/2019 Sequential Xsell
28/28
Lum S., and B. Moyer, Gross Domestic Product by Industry for 1997-99, Survey of Current Business,December 2000: 24-35.
Manchanda, P., A. Ansari, and S. Gupta (1999), The Shopping Basket: A Model for Multicategory PurchaseIncidence Decisions,Marketing Science, 18: 95-114.
Massy, W., R. Frank, and T. Lodahl,Purchasing Behavior and Personal Attributes (Philadelphia, University ofPennsylvania Press, 1968).
Mayo, M. C. and W. J. Qualls (1987), Household Durables Goods Acquisition Patterns: A LongitudinalStudy, P. Anderson and M. Wallendorf (eds.),Advances in Consumer Research. (Ann Arabor. MI: Associationfor Consumer Research), 14.
McFall, John (1969), Priority Patterns and Household Behavior,Journal of Marketing: 50-55.
Narasimhan, C., S. Neslin, and S. Sen (1996), Promotional Elasticities and Category Characteristics,Journalof Marketing, 60: 17-30.
Network World, Spending on the Rise, January 29, 2001.
Paroush, Jacob (1965), The Order of Acquisition of Household Durables,Econometrica: 225-235.
Pyatt, F. Graham, Priority Patterns and The Demand for Household Durable Goods, (Cambridge, MA:Cambridge University Press, 1964).
Raju, J. (1992), The Effect of Price Promotions on Variability in Product Category Sales, Marketing Science,11(3): 207-220.
Rossi, Peter E., Robert E. McCulloch and Greg M. Allenby (1996), The Value of Purchase History Data in
Marketing,Marketing Science, 15(4): 321-340.
Spiegelhalter, D. J., N.G. Best, B.P. Carlin, and A. Van der Linde, (2002), Bayesian Measures ofModel Complexity and Fit (with discussion),Journal of the Royal Statistical Society(B), 64(3): 583-639.
Wall Street Journal, Citigroup is Ready to Realize the Benefits of Cross-Selling, New York, March 15, 1999.
Wind Y., and R. Frank (1969), Interproduct Household Loyalty to Brands,Journal of MarketingResearch, 6:434-435.