Direct marketing decision support through predictive customer response modeling

David L. Olson
a
, Bongsug(Kevin) Chae
b,

a
Department of Management, University of Nebraska, Lincoln, NE 68588-0491, United States
b
Department of Management, Kansas State University, Manhattan, KS 66506, United States
a b s t r a c t a r t i c l e i n f o
Article history:
Received 8 June 2011
Received in revised form 12 May 2012
Accepted 19 June 2012
Available online 3 July 2012
Keywords:
Customer response predictive model
Knowledge-based marketing
RFM
Neural networks
Decision tree models
Logistic regression
Decision support techniques and models for marketing decisions are critical to retail success. Among different
marketing domains, customer segmentation or proling is recognized as an important area in research and
industry practice. Various data mining techniques can be useful for efcient customer segmentation and
targeted marketing. One such technique is the RFM method. Recency, frequency, and monetary methods pro-
vide a simple means to categorize retail customers. We identify two sets of data involving catalog sales and
donor contributions. Variants of RFM-based predictive models are constructed and compared to classical data
mining techniques of logistic regression, decision trees, and neural networks. The spectrum of tradeoffs is an-
alyzed. RFM methods are simpler, but less accurate. The effect of balancing cells, of the value function, and
classical data mining algorithms (decision tree, logistic regression, neural networks) are also applied to the
data. Both balancing expected cell densities and compressing RFM variables into a value function were
found to provide models similar in accuracy to the basic RFM model, with slight improvement obtained by
increasing the cutoff rate for classication. Classical data mining algorithms were found to yield better pre-
diction, as expected, in terms of both prediction accuracy and cumulative gains. Relative tradeoffs among
these data mining algorithms in the context of customer segmentation are presented. Finally we discuss prac-
tical implications based on the empirical results.
2012 Elsevier B.V. All rights reserved.
1. Introduction
The role of decision support techniques and models for marketing
decisions has been important since the inception of decision support
systems (DSSs) [25]. Diverse techniques and models (e.g., optimization,
knowledge-based systems, simulation) have emerged over the last ve
decades. Many marketing domains, including pricing, new product de-
velopment, and advertising, have beneted from these techniques and
models [16]. Among these marketing domains, customer segmentation
or proling is recognized as an important area [18,19,26,43]. There are
at least two reasons for this. First, the marketing paradigm is becoming
customer-centric [41] and targeted marketing and service are suitable.
Second, unsolicited marketing is costly and ineffective (e.g., low re-
sponse rate) [15,30]. Along with these reasons, there are increasing ef-
forts on collecting and analyzing customer data for better marketing
decisions [9,26,30]. The advancement of online shopping technologies
and database systems has accelerated this trend.
Data mining has been a valuable tool in this regard. Various data
mining techniques, including statistical analysis and machine learn-
ing algorithms, can be useful for efcient customer segmentation
and targeted marketing [4,26,38]. One such technique is RFM, stand-
ing for recency, frequency, and monetary. RFM analysis has been used
for marketing decisions for a long time and is recognized as a useful
data mining technique for customer segmentationandresponse models
[3,30]. A survey [43] also shows that RFM is among the most popular
segmentation and predictive modeling techniques used by marketers.
RFMrelies onthree customer behavioral variables (howlong since the
last purchase by customer, howoften the customer purchases, howmuch
the customer has bought) tondvaluable customers or donors anddevel-
op future direct marketing campaigns. Having a reliable and accurate cus-
tomer response model is critical for marketing success since an increase
or decrease in accuracy of 1% could have a signicant impact on their
prots [1]. While there could be many other customer-related factors
[e.g.,42], previous studies have shown that RFM alone can offer a pow-
erful way of predicting the future customer purchase [1,3,17].
Our research builds customer response models using RFM variables
and compares theminterms of customer gains and predictionaccuracy.
The paper aims to increase understanding of how to nd knowledge
hiddenin customer andtransactional databases using data mining tech-
niques. This area is called knowledge-based marketing [26]. The next
section briey reviews various data mining techniques for building cus-
tomer response or predictive models. Section 3 describes methodology.
All the response models will be built upon the three RFM variables,
while different data mining techniques are used. Then, we present a re-
search design, including two direct marketing data sets with over
100,000 observations, a process of predictive modeling building, and
methods to measure the performance of models. Section 4 includes
analysis and results. There could be different methods to increase the
Decision Support Systems 54 (2012) 443451
Corresponding author.
E-mail addresses: [email protected] (D.L. Olson), [email protected]
(B.(K.) Chae).
0167-9236/$ see front matter 2012 Elsevier B.V. All rights reserved.
doi:10.1016/j.dss.2012.06.005
Contents lists available at SciVerse ScienceDirect
Decision Support Systems
j our nal homepage: www. el sevi er . com/ l ocat e/ dss
prediction performance of an RFM-based predictive model and sophis-
ticated data mining techniques (decision tree, logistic regression, and
neural networks) appear to outperform more traditional RFM. These
ndings are further discussed in Section 5, comparing results with pre-
vious studies of customer response models and in the broad contexts of
knowledge-based marketing. We also discuss practical implications from
the ndings and offer conclusions.
The contribution of this study is to demonstrate how RFM model
variants can work, and supports general conclusions consistently
reported by others that RFM models are inferior to traditional data
mining models. This study shows that RFM variables are very useful
inputs for designing various customer response models with different
strengths and weaknesses and the ones relying on classical data min-
ing (or predictive modeling) techniques can signicantly improve the
prediction capability in direct marketing decisions. These predictive
models using RFM variables are simple and easy to use in practice
than those with a complex set of variables. Besides descriptive model-
ing techniques popular in practice [43], thus, marketers should adopt
those advanced predictive models in their direct marketing decisions.
2. Customer response models using data mining techniques
2.1. Marketing DSS and customer response models
The use of DSS in marketing goes back to the 1960s and 1970s
[22,44] and has been applied in various areas, including marketing
strategy, pricing, new product development, and product analysis
and management [16]. There has been an increase of DSS use in
customer-side marketing activities, such as customer segmentation
(or proling), direct marketing, database marketing, and targeted ad-
vertising. This reects advances in database management and com-
plex model building [11,16,35]. More convenient methods are
available for the acquisition and storage of large amounts of customer
and transactional data. In addition, knowledge-based systems or intel-
ligent systems using data mining techniques (e.g., neural networks)
[37] have emerged in the marketing domain.
This trend is broadly termed knowledge-based marketing.
Knowledge-based marketing is both data-driven and model-driven:
that is the use of sophisticated data mining tools and methods to
nd knowledge discovery from customer and transactional databases
[26]. Overall, this leads to more efcient and effective communication
with potential buyers and an increase in prots. An important ap-
proach to knowledge-based marketing is to understand customers
and their behavioral patterns. This requires such transactional charac-
teristics as recency of purchases, frequency of purchases, size of pur-
chases, identifying customer groups, and predicting purchases [35].
The RFM model and other data mining-based customer response
models have proven useful to marketers.
2.2. Data mining techniques for customer response models
2.2.1. RFM
R represents the period since the last purchase. F is the number of
purchases made by a customer during a certain period. M is the total
purchase amount by a customer over that period. It is common practice
for each R, F, and Mto have ve groups or levels and thus there are 125
(=5
*
5
*
5) customer segmentation groups. Each customer is segment-
ed into one cell or group. This model allows markets to differentiate
their customers in terms of three factors and to target the customer
groups that are likely to purchase products or services. This technique
is known as the benchmark model in the area of database marketing [3].
Since its introduction in a major marketing journal [5], RFM has
received a great deal of interest from both academic and industry
communities [3,17]. Many studies [1,13,17] have recognized these
three variables as important to predict the future responses by cus-
tomers to potential direct marketing efforts. Certain limitations in
the original RFM model have been recognized in the literature [31,45].
Some previous studies have extended the original RFM model either
by considering additional variables (e.g., socio-demographics) [1] or
by combining with other response techniques [6,7]. Because of the
high correlation between F and M, Yang [45] offered a version of RFM
model collapsing the data to a single variable Value=M/R. To over-
come the problem of data skewed in RFM cells, Olson et al. [31] pro-
posed an approach to balance observations in each of the 125 RFMcells.
Other variables that may be important include customer income,
customer lifestyle, customer age, product variation, and so on [14].
That would make traditional data mining tools such as logistic regres-
sion more attractive. However, RFM is the basis for a continuing
stream of techniques to improve customer segmentation marketing
[12]. RFMhas been found to work relatively well if expected response
rate is high [24]. Other approaches to improve RFM results have in-
cluded Bayesian networks [1,8] and association rules [46].
2.2.2. Classical data mining tools
Common data mining practice in classication is to gather a great
number of variables and apply different standard algorithms. Given
the set of predened classes and a number of attributes, these classi-
cation methods can provide a model to predict the class of other un-
classied data. Mathematical techniques that are often used to
construct classication methods are binary decision trees, neural net-
works, and logistic regression. By using binary decision trees, a tree
induction model with YesNo format can be built to split data into
different classes according to its attributes. Such a model is very
easy to apply to new cases, although the algorithms often produce
an excessive number of rules. Neural networks often t nonlinear re-
lationships very well, but are difcult to apply to new data. Logistic
regression models are easy to apply to new data, although the prob-
lem of a cutoff between classes can be an issue [32].
Relative performance of data mining algorithms has long been un-
derstood to depend upon the specic data. Since data mining soft-
ware is widespread, common practice in classication is to try the
three basic algorithms (decision trees, neural networks, logistic re-
gression), and use the one that works best for the given data set.
Studies have compared these algorithms with RFM. Levin and Zahavi
[20] compared RFM with decision trees (specically CHAID), pointing
out that decision trees are more automatic (RFM requires extensive
data manipulation), but involve modeling issues such as controlling
tree size and determining the best split for branches and leaves.
Kim and Street [19] proposed a neural network model and applied
feature selection mechanisms to reduce input variables, enabling
focus upon the most important variables. Baesens et al. [1] also ap-
plied neural networks to customer response models (adding custom-
er prole indicators to RFM), obtaining better prediction accuracy.
That is a consistent nding data mining algorithms will be expected
to better predict customer response than RFM. However, RFM re-
mains interesting because it relies upon the three fundamentally
basic inputs that are readily available.
3. Methodology
3.1. Problem description and data set
This research design includes two studies (Study 1 and Study 2
hereafter) using two datasets obtained from the Direct Marketing
Educational Foundation. Study 1 uses a dataset including 101,532 in-
dividual purchases from 1982 to 1992 in catalog sales. Study 2 is
based on the data of 1,099,009 individual donors' contributions to a
non-prot organization collected between 1991 and 2006. The pur-
chase orders (or donations) included ordering (or donation) date
and ordering amount. The last four months (AugDec) of the data
were used as the target period: AugDec 1992 for Study 1 and
444 D.L. Olson, B.(K.) Chae / Decision Support Systems 54 (2012) 443451
AugDec 2006 for Study 2. The average response rates in Studies 1
and 2 are 0.096 and 0.062 respectively.
Data preparation and manipulation are an important stage of
knowledge discovery and learning in knowledge-base marketing
[35]. Fig. 1 describes our approach. The raw data contained customer
behavior represented by account, order (or donation) date, order (do-
nation) dollars, and many other variables. We followed the general
coding scheme to compute R, F, and M [17]. Various data preparation
techniques (e.g., ltering, transforming) were used during this pro-
cess. The order date of last purchase (or the date of last donation)
was used to compute R (R1, R2, R3, R4, R5). The data set contained
order (or donation) history and order dollars (or donation amounts)
per each customer (or donor), which were used for F (F1, F2, F3, F4,
F5) and M(M1, M2, M3, M4, M5). We also included one response var-
iable (Yes or No) to the direct marketing promotion or campaign.
3.2. Predictive models
3.2.1. RFM
RFM analysis typically divides the data into 125 cells, designated
by the 5 groups. The most attractive group would be 555, or Group
5 for each of the 3 variables [17].
3.2.2. RFM with balanced cells
Dividing customers or donors into 125 cells tends to result in the
skewness that the data is not evenly distributed among those cells. This
skewness has been recognized as one of the problems with RFM
[13,27,31]. Our approach to this issue was through more equal density
(size-coding) to obtain data entries for all RFM cells. We accomplished
this by adjusting cell limits to obtain more equal counts for cells in the
training set.
3.2.3. RFM with Yang's value function
Previous studies [19] have pointed out a strong correlation between
F and M as a limitation of RFM. The value function [45] compresses the
RFM data into one variable V=M/R.
3.2.4. Logistic regression (LR)
The purpose of logistic regression is to classify cases into the most
likely category. Logistic regression provides a set of parameters for
the intercept (or intercepts in the case of ordinal data with more than
two categories) and independent variables, which can be applied to a
logistic function to estimate the probability of belonging to a specied
output class [32]. Logistic regression is among the most popular data
mining techniques in marketing DSS and response modeling [24].
3.2.5. Decision tree (DT)
Decision trees in the context of data mining refer to the tree struc-
ture of rules. They have been applied by many in the analysis of direct
marketing data [39,40]. The data mining decision tree process in-
volves collecting those variables that the analyst thinks might bear
on the decision at issue, and analyzing these variables for their ability
to predict outcome. Decision trees are useful to gain further insight
into customer behavior, as well as lead to ways to protably act on
results. One of a number of algorithms automatically determines
which variables are most important, based on their ability to sort
the data into the correct output category. The method has relative
advantage over neural network and genetic algorithms in that a
reusable set of rules are provided, thus explaining model
conclusions.
3.2.6. Neural networks (NN)
Neural networks are the third classical data mining tool found in
most commercial data mining software products, and have been ap-
plied to direct marketing applications [4,8,19,36]. NN are known for
their ability to train quickly on sparse data sets. NN separates data
into a specied number of output categories. NN are three layer net-
works wherein the training patterns are presented to the input layer
and the output layer has one neuron for each possible category.
3.3. Performance evaluation measures
There are different methods to assess customer response model
performances. We use prediction accuracy and cumulative gains to
discuss the performance of different predictive customer response
models. Gains show the percentage of responders in each decile. Mar-
keters can gure out how many responders (or what proportion of
responders) can be expected in a specic decile. For example, we
can say that given a same mailing size (e.g., 40% of the total cus-
tomers) a model capturing 70% of the responders is better than a
model capturing only 60% of the responders [47]. Through cumulative
gain values we can evaluate the performances of different data min-
ing techniques [21]. Another way is using prediction accuracy rate
of each technique. The data set employed in this research has the
information about who responded to the direct marketing or cam-
paign. Using R, F, and M as three predictive variables, each data
mining technique will develop a binary customer response model
based on the training data set and apply the model to the test
data set. This will generate prediction accuracy rate the percentage
of customers classied correctly [21]. The model building process is
shown in Fig. 1.
4. Analysis and results
The analysis process consisted of model building using each data
mining technique and model assessment. For Study 1, customer re-
sponse models were developed using RFM, RFM with balanced cells,
RFM with Yang's value function, logistic regression (LR), decision tree
(DT), and neural networks (NN). Model assessment is presented with
gains and predictive accuracy.
4.1. Study 1
An initial correlation analysis was conducted, showing that there
was some correlation among these variables, as shown in Table 1.
Fig. 1. Research design building predictive models using RFM variables.
445 D.L. Olson, B.(K.) Chae / Decision Support Systems 54 (2012) 443451
All three variables were signicant at the 0.01 level. The relation-
ship between R and customer response is negative, as expected. In
contrast, F and M are positively associated with customer response.
R and F are stronger predictors for customer response.
RFM was initially applied, dividing the scales for each of the three
components into ve groups based upon the scales for R, F, and M.
This was accomplished by entering bin limits in SPSS. Table 2 shows
boundaries. Group 5 was assigned the most attractive group, which
for R was the minimum, and for F and M the maximum.
Note the skewness of the data for F, which is often encountered.
Here the smaller values dominate that metric. Table 3 displays the
counts obtained for these 125 cells.
The proportion of responses (future order placed) for the data is
given in Table 4.
In the training set, 10 of 125 possible cells were empty, even with
over 100,000 data points. The cutoff for protability would depend
upon cost of promotion compared to average revenue and rate of
prot. For example, if cost of promotion were $50, average revenue
per order $2000, and average prot rate $0.25 per dollar of revenue,
the protability cutoff would be 0.1. In Table 4, those cells with return
ratios greater than 0.1 are shown in bold. Those cells with ratios at 0.1
or higher with support (number of observations) below 50 are indi-
cated in italics. They are of interest because their high ratio may be spu-
rious. The implication is fairly self-evident seek to apply promotion to
those cases in bold without italics. The idea of dominance can also be
applied. The combinations of predicted success for different training
cell proportions are given in Table 5.
The RFM model from the Excel spreadsheet model yields predictive
model performance shown in the Appendix A for the line Basic on 0.1
(because the cutoff used was a proportion of 0.1) along with results
from the other models. This model was correct (13,961+1337=
15,298) times out of 20,000, for a correct classication rate of 0.765.
The error was highly skewed, dominated by the model predicting
4113 observations to be 0 that turned out to respond. An alternative
model would be degenerate simply predict all observations to be 0.
This would have yielded better performance, with 18,074 correct re-
sponses out of 20,000, for a correct classication rate of 0.904. This
value could be considered a par predictive performance. This data is in-
cluded in the Appendix A, where we will report results of all further
models in terms of correct classication.
Increasing the test cutoff rate leads to improved models. We used
increasing cutoffs of 0.2, 0.3, 0.4, and 0.5, yielding the results indicat-
ed in the Appendix A. Only the model with a cutoff rate of 0.5 resulted
in a better classication rate than the degenerate model. In practice,
the best cutoff rate would be determined by nancial impact analysis,
reecting the costs of both types of errors. Here we simply use classi-
cation accuracy overall, as we have no dollar values to use.
The correlation across F and M (0.631 in Table 1) can be seen in
Table 3, looking at the R=5 categories. In the M=1 column of
Table 3, F entries are 0 for every F5 category, usually increasing
through M=2 through M=5 columns. When F=5, the heaviest
density tends to be in the column where M=5. This skewness is
often recognized as one of the problems with RFM [13,27,31]. Our ap-
proach to this issue was through more equal density (size-coding) to
obtain data entries for all RFM cells. We accomplished this by setting
cell limits by count within the training set for each variable. We
cannot obtain the desired counts for each of the 125 combined cells
because we are dealing with three scales. But we can come closer,
as in Table 6. Difculties arose primarily due to F having integer
values. Table 6 limits were generated sequentially, starting by divid-
ing R into 5 roughly equal groups. Within each group, F was then
sorted into groups based on integer values, and then within those
25 groups, M divided into roughly equally sized groups.
The unevenness of cell densities is due to uneven numbers in the
few integers available for the F category. The proportion of positive
responses in the training set is given in Table 7.
If M=5, this model predicts above average response. There is a
dominance relationship imposed, so that cells 542 and better, 532
and better, 522 and better, 512 and better, 452 and better, 442 and
better, and 433 and better are predicting above average response.
Cells 422, 414, and 353 have above average training response, but
cells with superior R or F ratings have below average response, so
these three cells were dropped from the above average response
model. The prediction accuracy ((13,897+734)/ 20,000) for this
model was 0.732 (see the balance on 0.1 row in the Appendix A). In
this case, balancing cells did not provide added accuracy over the
basic RFM model with unbalanced cells. Using the cutoff rate of 0.5,
the model is equivalent to predict the combination of R=5, F=4 or 5,
and M=4 or 5 as responding and all others not. This model had a cor-
rect classication rate of 0.894, which was inferior to the degenerate
case. For this set of data, balancing cells accomplished better statistical
properties per cell, but was not a better predictor.
Since F is highly correlated with M (0.631 in Table 1), the analysis
is simplied to one dimension. Dividing the training set into groups of
5%, sorted on V, generates Table 8.
Table 1
Variable correlations.
R F M Ordered
R 1
F 0.192

1
M 0.136

0.631

1
Ordered 0.235

0.241

0.150

Correlation is signicant at the 0.01 level (2-tailed).

Table 2
RFM boundaries.
Factor Min Max Group 1 Group 2 Group 3 Group 4 Group 5
R 12 3810 1944+ 12911943 6881290 306687 12305
Count 16,297 16,323 16,290 16,351 16,271
F 1 39 1 2 3 45 6+
Count 43,715 18,274 8206 6693 4644
M 0 4640 020 2138 3965 66122 123+
Count 16,623 16,984 15,361 16,497 16,067
Table 3
Count by RFM cell training set.
RF R F M1 M2 M3 M4 M5
55 R 12305 F 6+ 0 0 16 151 1761
54 F 45 2 18 118 577 1157
53 F 3 9 94 363 756 671
52 F 2 142 616 1012 1135 559
51 F 1 2425 1978 1386 938 387
45 R306687 F 6+ 0 1 11 101 1018
44 F 45 0 16 87 510 927
43 F 3 6 88 316 699 636
42 F 2 150 707 1046 1140 616
41 F 1 2755 2339 1699 1067 416
35 R6881290 F 6+ 0 1 5 70 799
34 F 45 1 16 122 420 832
33 F 3 9 88 319 706 589
32 F 2 163 697 1002 1128 645
31 F 1 2951 2567 1645 1078 437
25 R12911943 F 6+ 0 0 9 56 459
24 F 45 0 22 72 372 688
23 F 3 9 95 290 678 501
22 F 2 211 749 1096 1128 561
21 F 1 3377 2704 1660 1108 478
15 R 1944+ F 6+ 0 0 3 22 170
14 F 45 1 11 74 243 409
13 F 3 9 122 261 511 380
12 F 2 268 878 1108 995 522
11 F 1 4145 3177 1641 908 449
Totals 16,623 16,984 15,361 16,497 16,067
446 D.L. Olson, B.(K.) Chae / Decision Support Systems 54 (2012) 443451
Lift is the marginal difference in a segment's proportion of re-
sponse to a promotion and the average rate of response. Target cus-
tomers are identied as the small subset of people with marginally
higher probability of purchasing. Lift itself does not consider prot-
ability. In practice, this needs to be considered. For our purposes, we
demonstrate without dollar values (which are not available), noting
that the relative cost of marketing and expected protability per seg-
ment will determine the optimal number of segments to market.
Fig. 2 shows lift by value ratio.
In Fig. 2, the most responsive segment has an expected return of
slightly over 20%. The lift line is the cumulative average response as
segments are added (in order of response rate).
Using the value ratio as a predictive classier, the training data
was used to identify cells with better responses. Model t is shown
in the Appendix A in the row value function. This model has a correct
classication rate of 0.721. This is inferior to a degenerate model that
would simply classify all cases as no response, indicating that the
value function was non-productive in this case. While it is easier to
manipulate than the RFM model, in this case the t was inferior to
the basic RFM model.
Data mining is rich in classication models [2,23]. Three classical
data mining classication models were applied to the data: logistic
regression, decision trees, and neural networks. We next applied
these three basic data mining algorithms using SPSS.
A logistic regression model was run on RFM variables. The model
results were as shown in Table 9. The beta values of R and F are
found to be signicant.
Note that F was not included at all. This is explainable by the high
correlation between M and F, and the dominance of R in obtaining a
better t. This model did very well on the test data, with a correct
classication rate of 0.984.
The neural network model used a popular architecture called mul-
tilayer perceptron (MLP) [10]. This model built a hidden layer. The
rate of neural network was 0.911, as shown in the Appendix A.
Another performance measure used in this study is gains, which is a
useful tool for evaluating the value of predictive models in direct mar-
keting [21]. We use gains to compare the performance of RFM score
model and classical data mining-based predictive models. Table 11
Table 4
Response ratios by cell.
RF R F M1 M2 M3 M4 M5
55 R 12306 F 6+ 0.687 0.563 0.558
54 F 45 0 0.500 0.415 0.426 0.384
53 F 3 0.111 0.426 0.342 0.381 0.368
52 F 2 0.296 0.289 0.281 0.283 0.256
51 F 1 0.173 0.196 0.201 0.158 0.152
45 R307687 F 6+ 0 0.273 0.238 0.193
44 F 45 0.125 0.092 0.112 0.123
43 F 3 0 0.091 0.082 0.089 0.101
42 F 2 0.060 0.075 0.069 0.081 0.078
41 F 1 0.047 0.049 0.052 0.053 0.041
35 R6881286 F 6+ 1.000 0 0.100 0.125
34 F 45 0 0.063 0.107 0.107 0.103
33 F 3 0.111 0.023 0.066 0.059 0.075
32 F 2 0.049 0.047 0.061 0.063 0.060
31 F 1 0.030 0.031 0.029 0.026 0.021
25 R12871943 F 6+ 0.111 0.054 0.078
24 F 45 0.091 0.028 0.065 0.060
23 F 3 0 0.053 0.048 0.049 0.064
22 F 2 0.043 0.020 0.039 0.041 0.039
21 F 1 0.018 0.021 0.018 0.020 0.019
15 R 1944+ F 6+ 0.000 0.045 0.041
14 F 45 0 0.091 0.024 0.025 0.039
13 F 3 0.111 0.041 0.050 0.033 0.053
12 F 2 0.019 0.046 0.036 0.031 0.044
11 F 1 0.021 0.015 0.016 0.020 0.016
Table 5
Basic RFM models by cutoff.
Cutoff R F M
0.1 R=5 Any Any
R=4 F=5 M=3, 4, or 5
R=3 F=4 M=4 or 5
F=3 M=5
F=4 or 5 M=3, 4, or 5
0.2 R=5 F=2, 3, 4, or 5 Any
R=4 F=5 M=3, 4, or 5
0.3 R=5 F=3, 4, or 5 M=2, 3, 4, or 5
0.4 R=5 F=4 or 5 M=2, 3, 4 or 5
0.5 R=5 F=5 M=3, 4, or 5
Table 6
Balanced group cell densitiestraining set.
RF M1 M2 M3 M4 M5
55 186 185 149 223 187
54 185 186 185 185 186
53 187 185 188 186 187
52 184 184 185 184 185
51 186 187 186 187 186
45 268 265 270 289 246
44 269 269 268 274 264
43 272 267 280 251 296
42 263 263 265 245 283
41 268 261 261 259 277
35 331 330 349 316 330
34 324 325 322 325 324
33 332 331 329 332 335
32 330 330 330 331 330
31 323 324 323 326 324
25 733 730 735 737 733
24 735 736 735 737 734
23 747 746 751 749 748
22 705 704 707 704 707
21 731 733 730 735 732
15 1742 1746 1739 1740 1744
14 1718 1715 1713 1713 1716
13 1561 1809 1689 1675 1684
12 1768 1775 1771 1779 1762
11 1830 1831 1832 1824 1839
Table 7
Training set proportion of responses by cell.
RF M1 M2 M3 M4 M5
55 0.129 0.178 0.101 0.673 0.818
54 0.059 0.118 0.189 0.541 0.629
53 0.064 0.130 0.287 0.392 0.647
52 0.076 0.103 0.200 0.424 0.605
51 0.054 0.102 0.274 0.406 0.527
45 0.037 0.109 0.141 0.211 0.378
44 0.041 0.108 0.116 0.281 0.417
43 0.033 0.052 0.125 0.072 0.483
42 0.049 0.118 0.098 0.073 0.544
41 0.045 0.038 0.092 0.116 0.531
35 0.045 0.067 0.138 0.060 0.458
34 0.052 0.043 0.059 0.080 0.448
33 0.042 0.048 0.058 0.093 0.433
32 0.027 0.045 0.058 0.097 0.379
31 0.050 0.040 0.062 0.080 0.414
25 0.037 0.051 0.056 0.084 0.254
24 0.024 0.046 0.052 0.076 0.309
23 0.051 0.047 0.055 0.080 0.273
22 0.027 0.040 0.055 0.068 0.246
21 0.027 0.038 0.048 0.076 0.242
15 0.017 0.021 0.025 0.051 0.146
14 0.016 0.017 0.033 0.054 0.167
13 0.010 0.019 0.034 0.052 0.156
12 0.018 0.021 0.036 0.043 0.137
11 0.016 0.022 0.014 0.044 0.154
447 D.L. Olson, B.(K.) Chae / Decision Support Systems 54 (2012) 443451
displays cumulative gains for different deciles for RFMscore model, de-
cision tree, logistic regression, and neural networks. This gain-value in-
formation is well aligned with the prediction accuracy. The predictive
response models based on decision, logistic regression, and neural net-
work signicantly outperformed RFMscore model. For example, if only
10% of the total customers are selected for direct marketing promotion,
the RFM-based predictive model can include only 32.1% of actual re-
spondents in that sampling customer group, those of logistic and neural
network are 38.7% and 42.9% respectively. With selecting a group of
only 20% of customers, the decision tree-based predictive model can in-
clude almost 95% of actual buyers.
For Study 1, we used J48, one of the most popular decision tree algo-
rithms. The J48 decision tree algorithm using 10 fold cross-validation
[10] was applied to the dataset. The resultant decision tree was as
shown in Table 10.
4.2. Study 2
An initial correlation analysis was conducted, showing that there
was signicant correlation among these variables, as shown in Table 12.
F and M appear to have a strong correlation [45]. R and F appear to
be strong predictors for customer response [1]. Table 13 shows RFM
limits for this dataset and cell counts.
We built an RFM model by following the same procedures de-
scribed in Study 1. An RFM model using a cutoff rate of 0.1 was
built on half of the dataset, and tested on the other half. This yielded
a model with a correct classication rate of 0.662, as reported in the
Appendix A. This was far worse than any of the other models tested.
Difculties arose in balancing cells due to F being only a few inte-
ger values (1, 2, 3, 4, 5+) and highly skewed, letting a majority of the
data assigned into F group1.
Fig. 3 displays the lift chart for the V models. The lift chart shows that
the 5% of cases with the most likely response is much more likely to re-
spond than the least responsive 50%. The proportion of responses in the
test set for the 5% highest training set V scores had a response ratio of
0.311, comparedtoless than0.010for the worst 50%. We applieddifferent
Vlevels (0.05 and up; 0.10 and up; 0.15 and up; 0.20 and up; 0.25 and up;
and 0.30 and up). These six models had very consistent results as shown
in the Appendix A, just slightly inferior to the degenerate model. When
datasets are highly skewed as this is, with roughly only 5% responding,
the degenerate model becomes very hard to beat.
All three predictive data mining models (DT, LR, NN) were built as
in Study 1. The result is that those three models are performed equal-
ly in terms of accuracy (0.938), as shown in Appendix A. We also per-
formed the gain analysis reported in Table 14. The predictive models
using decision tree, logistic regression, and neural networks out-
performedthe RFMScore model. The performance gap is more signicant
when a small sample size (e.g., 20%) is chosen for donor solicitation.
5. Discussion and conclusion
Marketing professionals have found RFM to be quite useful
[17,18,35,43], primarily because the data is usually at hand and the
technique is relatively easy to use. However, previous research
Table 8
V values by cell.
Cell Min V UL Hits N Success
1 0.0000 4077 91 4076 0.0223
2 0.0063 8154 69 4077 0.0169
3 0.0097 12,231 116 4077 0.0285
4 0.0133 16,308 109 4077 0.0267
5 0.0171 20,385 120 4077 0.0294
6 0.0214 24,462 119 4077 0.0292
7 0.0263 28,539 151 4077 0.0370
8 0.0320 32,616 174 4077 0.0427
9 0.0388 36,693 168 4077 0.0412
10 0.0472 40,770 205 4077 0.0503
11 0.0568 44,847 258 4077 0.0633
12 0.0684 48,924 256 4077 0.0628
13 0.0829 53,001 325 4077 0.0797
14 0.1022 57,078 360 4077 0.0883
15 0.1269 61,155 408 4077 0.1001
16 0.1621 65,232 542 4077 0.1329
17 0.2145 69,309 663 4077 0.1626
18 0.2955 73,386 827 4077 0.2028
19 0.4434 77,463 1134 4077 0.2781
20 0.7885 81,540 1686 4070 0.4143
Total/avg 7781 81,532 0.0954
Fig. 2. Lift by value ratio cell.
Table 9
Regression betas for logistic regression.
Variable Beta Signicance
Constant 1.5462 0.05
R 0.0015 b0.05
F 0.2077 b0.05
M 0.0002
Table 10
J48 Decision Tree.
R M Yes Total P (yes) P (no) Conclusion Error
036 1 1 1.000 Yes
37152 41 619 0.066 0.934 No 41
153 605 606 0.998 0.002 Yes 1
154257 53 1072 0.049 0.951 No 53
258260 449 500 0.898 0.102 Yes 51
261516 0 2227 0.000 1.000 No
517519 119 144 0.826 0.174 Yes 25
520624 0 1219 0.000 1.000 No
625 206 227 0.907 0.093 Yes 21
626883 0 2047 0.000 1.000 No
884 51 68 0.750 0.250 Yes 17
885989 0 1116 0.000 1.000 No
990 135 160 0.844 0.156 Yes 25
9911248 0 1773 0.000 1.000 No
1249 31 37 0.838 0.162 Yes 6
12501354 0 985 0.000 1.000 No
1355 85 108 0.787 0.213 Yes 23
13561612 0 1290 0.000 1.000 No
16131614 17 28 0.607 0.393 Yes 11
16151720 0 786 0.000 1.000 No
1721 36 36 1.000 0.000 Yes
17222084 14 1679 0.008 0.992 No 14
20852086 18 18 1.000 0.000 Yes
20872343 0 831 0.000 1.000 No
23442345 7 7 1.000 0.000 Yes
23462448 0 404 0.000 1.000 No
24492451 M>44 21 24 0.875 0.125 Yes 3
Mb=44 8 12 0.667 0.333 No 8
24522707 0 665 0.000 1.000 No
27082710 3 5 0.600 0.400 Yes 2
2711+ 26 1306 0.020 0.980 No 26
Total 1926 20,000 0.096 0.904 327
448 D.L. Olson, B.(K.) Chae / Decision Support Systems 54 (2012) 443451
suggests that it is easy to obtain a stronger predictive customer re-
sponse model with other data mining algorithms [e.g., 1, 19, 20, 24].
RFM has consistently been reported to be less accurate than other
forms of data mining models, but that is to be expected, as the origi-
nal RFM model segmenting customers/donors into 125 cells and is
prescriptive rather than predictive.
That expected result was conrmed in this research. RFMhelped nice-
ly structure millions of records in each dataset into 125 groups of cus-
tomers using only three variables. The model offers a well-organized
description of people based on their past behaviors, which helps mar-
keters effectively identify valuable customers or donors and develop a
marketing strategy. However, this descriptive approach is less accurate
in predicting future behavior than more complex data mining models.
There have been proposed improvements to RFM. In the models
seeking to improve RFM, our study showed that increasing the cutoff
limit will lead to improvement in prediction accuracy. However, RFM
models at any cutoff limit have trouble competing with degenerate
models. Degenerate models have high predictive accuracy for highly
skewed datasets, but provide no benet as they simply conclude it
is not worth promoting to any customer prole.
Balancing cell sizes by adjusting the limits for the three RFM vari-
ables is sound statistically, but did not lead to improved accuracy in
our tests. In both Study 1 and Study 2, the basic RFM model signi-
cantly underperformed other predictive models, except the V func-
tion model in Study 1. These results indicate that balancing cells
might help improve t, but involves signicant data manipulation
for very little predictive improvement in the data set we examined.
Using the V ratio is an improvement to RFM that is useful in theo-
ry, but in our tests the results are mixed. In Study 1, the technique did
not provide better predictive accuracy. In Study 2, it did yield an im-
proved classication rate but underperformed the degeneracy model.
Thus, this technique deserves a further inquiry. Overall, the results
above indicate that some suggested alternatives to the traditional
RFM have limitations in prediction.
The primary conclusion of our study, as was expected, is that clas-
sical data mining algorithms outperformed RFM models in terms of
both prediction accuracy and cumulative gains. This is primarily be-
cause decision tree, logistic regression, and neural networks are
often considered the benchmark predictive modeling techniques
[4,10,29,42]. The demand of predictive modeling or analytics is in
high demand in many industries [9], including direct marketing [4].
This implies that marketers can make more effective marketing deci-
sions by embracing advanced predictive modeling techniques, be-
sides popular descriptive models. It often is the case that decision
tree, logistic regression, and neural networks vary in their ability to
t specic sets of data [34]. Furthermore, there are many parameters
that can be used with neural network models and decision trees. All
three of these model types have the advantage of being able to
consider external variables in addition to R, F, and M. Here, we ap-
plied them to these three variables without adding other explanatory
variables. All three model types did better than the degenerate case,
or any of the other variants we applied.
The best overall predictive t was obtained using the decision tree
model. This model also outperformed other predictive models in cu-
mulative gains in both studies. Decision tree tends have advantages
over low dimensionality datasets [34] like those used in this research.
This characteristic of decision tree may explain this result. Thus, we
do not contend that decision tree will always be best. However,
there is a major relative value for decision trees that they provide
an easily understandable model. For example, Table 10 presents the
decision tree rule sets obtained in Study 1, which amounts to enu-
merating ranges of R that had high densities of response. There was
only one range where M was used (R=2449 to R=2451). And
looking at Table 10, the t would have been improved if the decision
tree had not differentiated and called all of these cases Yes. (There
would have been 4 fewer errors out of 20,000, yielding essentially
the same t with the same correct response of 0.984.) There is the
downside for decision trees that they often overt the data (as they
did in Table 10), and can yield an excessive number of rules for
users to apply. Table 15 presents a comparison of methods based on
inferences from our two studies.
While our study uses predication accuracy along with cumulative
gains for model comparison, in practice the type of error can be consid-
ered in terms of relative costs, thus enabling inuence on prot. For ex-
ample, our study shows that increasing the cutoff level between
predicting response or not can improve correct classication. However,
a more precise means to assess this would be to apply the traditional
cost function reecting the cost of the two types of error. This is to be
a consideration inevaluating other predictive models as well. Thus, spe-
cic models should be used in light of these relative costs.
The good performance of those data mining methods (particularly
decision tree), in terms of prediction accuracy and cumulative gains,
indicates that three variables (R, F, and M) alone can be useful for
building a reliable customer response model. This echoes the impor-
tance of RFM variables in understanding customer purchase behavior
and developing response models for marketing decisions [17,18,33,35].
Previous research [e.g., 1] also shows that inclusion of non-RFMattributes
(e.g., income) is likely to slightly improve the model performance.
Table 11
Gains.
10% 20% 30% 40% 50%
RFM score 32.12 49.83 62.24 72.26 81.05
LR 38.79 61.67 70.67 79.01 84.85
DT 89.62 95.67 96.94 98.21 99.48
NN 42.95 60.28 70.21 79.12 84.80
Table 12
Variable correlations.
R F M Response
R 1
F 0.237

1
M 0.125

0.340

1
Response 0.266

0.236

0.090

Correlation is signicant at the 0.01 level (2-tailed).

Fig. 3. Lift chart for study 2.
Table 13
RFM boundaries.
Factor Min Max Group 1 Group 2 Group 3 Group 4 Group 5
R 1 4950 2811+ 19322811 9351932 257935 1257
Count 220,229 219,411 220,212 219,503 219,654
F 1 1027 1 2 3 4 5+
Count 599,637 190,995 95,721 57,499 155,157
M 0 100,000 09 1024 2539 4089 90+
Count 248,639 343,811 77,465 209,837 219,257
449 D.L. Olson, B.(K.) Chae / Decision Support Systems 54 (2012) 443451
However, a sophisticated model with too many variables is not very ef-
fective for marketing practitioners [19] and reducing variables is impor-
tant for practical use of predictive models [28]. Marketers should be
aware of this tradeoff between a simple model (with fewer variables)
and a sophisticated model (witha large number of variables) and devel-
op a well-balanced model using their market and product knowledge.
To repeat the contributions of this paper given in the Introduction,
we have demonstrated how RFM models and variants can be
implemented. RFM models have the relative advantage that they are
simple in concept, and thus understandable to users. However, they
can easily be improved in terms of predictive accuracy (or protabil-
ity, given situational data) by using classical data mining models. Of
these traditional data mining models, decision trees are especially at-
tractive in that they have easily understood output. These advanced
predictive models are muchbenecial inthe practice of direct marketing
since they can use only three behavioral input variables and generate
the results signicantly better than the traditional RFM model and
other variants.
Appendix A. Comparative model results Study 1
Comparative model results Study 2
References
[1] B. Baesens, S. Viaene, D. den Poel, J. Vanthienen, Bayesian neural network learning
for repeat purchase modelling in direct marketing, European Journal of Opera-
tional Research 138 (2002) 191211.
[2] N. Belacel, H. Raval, A. Punnen, Learning multicriteria fuzzy classication method
PROAFTNfromdata, Computers and Operations Research 34 (7) (2007) 18851898.
[3] R. Blattberg, B. Kim, S. Neslin, Database Marketing: Analyzing and Managing Cus-
tomers, Chapter 2 RFM Analysis, Springer, New York, 2008.
[4] I. Bose, X. Chen, Quantitative models for direct marketing: a review from systems
perspective, European Journal of Operational Research 195 (2009) 116.
[5] J. Bult, T. Wansbeek, Optimal selection for direct mail, Marketing Science 14 (4)
(1995) 378394.
[6] C. Cheng, Y. Chen, Classifying the segmentation of customer value vis RFM model
and RS theory, Expert Systems with Applications 36 (3) (2009) 41764184.
[7] W. Chiang, To mine association rules of customer values via a data mining procedure
with improved model: an empirical case study, Exper Systems with Applications 38
(2011) 17161722.
[8] G. Cui, M. Wong, H. Lui, Machine learning for direct marketing response models:
Bayesian networks with evolutionary programming, Management Science 52 (4)
(2006) 597612.
[9] T. Davenport, J.G. Harris, R. Morison, Analytics at Work: Smarter Decisions, Better
Results, Harvard Business Press, Boston, MA, 2010.
[10] D. Delen, A comparative analysis of machine learning techniques for student re-
tention management, Decision Support Systems 49 (2010) 498506.
Table 14
Gains.
10% 20% 30% 40% 50%
RFM score 40.38 62.39 84.66 95.63 97.90
LR 43.24 66.22 86.10 95.75 99.75
DT 44.68 70.75 87.41 96.63 97.96
NN 43.64 67.58 86.12 95.75 99.77
Table 15
Comparison of methods.
Model Relative
advantages
Relative
disadvantages
Inferences
Degenerate Tends to have
high
accuracy when
outcome highly
skewed
Mindless
Simply says no
Provides no
marginal value
If cost of missing good
responses is low, don't do
anything
Basic RFM Widely used
Data readily
available
Software
obtainable
Predictive
accuracy
consistently weak
Can do better using
conventional data mining
(RFM implicitly a special
case)
RFM with
balanced
data
Better statistical
practice
May not actually
improve accuracy
Not worth the trouble
Value
function
Easy to apply
(uses 2
of the 3 RFM
variables, so
data readily
available)
Focuses on
uncorrelated
variables
Not necessarily
more accurate
Value function is superior
to RFM
Logistic
regression
Can get better t
Can include
many variables
Model
statistically
interpretable
Logistic output
harder to interpret
than OLS for
managers
Decision trees easier to
interpret
Neural
network
Can get better t
Can include
many variables
Output not
conducive to
interpretation
Can't apply model
outside of software
used to build model
Decision trees easier to
interpret
Decision
trees
Can get better t
Can include
many variables
Output easily
understandable
by managers
Model may involve
an excessive
number of rules
Best option, if can control
the number of rules
obtained (through
minimum required
response parameter)
Model Actual
no response,
model response
Actual
response,
model no
response
Correct
response
Overall
correct
classication
Degenerate 0 1926 18,074 0.904
Basic RFM on 0.1 4113 589 15,298 0.765
Basic RFM on 0.2 1673 999 17,328 0.866
Basic RFM on 0.3 739 1321 17,940 0.897
Basic RFM on 0.4 482 1460 18,058 0.903
Basic RFM on 0.5 211 1643 18,146 0.907
Balance using 0.5 1749 379 17,872 0.894
Value function 623 4951 14,426 0.721
Logistic regression 1772 91 18,137 0.907
Neural network 119 1661 18,220 0.911
Decision tree 185 142 19,673 0.984
Model Actual
no response,
model response
Actual
response,
model no response
Correct
response
Overall correct
classication
Degenerate 0 34,598 515,123 0.9371
Basic RFM 4174 181,357 364,190 0.6625
Value
function >5
6212 30,418 513,091 0.9334
Value
function>10
3344 31,830 514,547 0.9360
Value
function>15
2296 32,475 514,950 0.9367
Value
function>20
1712 32,867 515,142 0.9371
Value
function>25
1400 33,136 515,185 0.9372
Value
function>30
1153 33,330 515,238 0.9373
Logistic
regression
821 32,985 515,915 0.9385
Neural network 876 32,888 515,957 0.9386
Decision tree 393 33,373 515,955 0.9386
450 D.L. Olson, B.(K.) Chae / Decision Support Systems 54 (2012) 443451
[11] E. Eisenstein, L. Lodish, Marketing decision support and intelligence systems: pre-
cisely worthwhile or vaguely worthless? In: in: B. Weitz, R. Wensley (Eds.), Hand-
book of Marketing, Sage Publication, London, 2002.
[12] R. Elsner, M. Krafft, A. Huchzemeier, Optimizing Rhenania's main-order business
through dynamic multilevel modeling, Interfaces 33 (1) (2003) 5066.
[13] P. Fader, B. Hardie, K. Lee, RFM and CLV: using iso-value curves for customer base
analysis, Journal of Marketing Research 42 (4) (2005) 415430.
[14] M. Fitzpatrick, Statistical analysis for direct marketers in plain English, Direct
Market 64 (4) (2001) 5456.
[15] R. Gopal, Ad mediation: new horizons in effective email advertising, Communica-
tions of the ACM 19 (1) (2001) 1730.
[16] M. Hart, Systems for supporting marketing decisions, In: in: F. Burstein, C.
Holsapple (Eds.), Handbook on Decision Support Systems, 2, Springer, 2008,
pp. 395418.
[17] A. Hughes, Strategic Database Marketing, Third ed. McGraw-Hill, New York, 2006.
[18] J. Jonker, N. Piersma, R. Potharst, A decision support system for direct mailing deci-
sions, Decision Support Systems 42 (2006) 915925.
[19] Y. Kim, W. Street, An intelligent system for customer targeting: a data mining ap-
proach, Decision Support Systems 37 (2004) 215228.
[20] N. Levin, J. Zahavi, Predictive modeling using segmentation, Journal of Interactive
Marketing 15 (2) (2001) 222.
[21] G. Linoff, M. Berry, Data Mining Techniques, Wiley, Indianapolis, 2011.
[22] J. Little, Decision support systems for marketing managers, Journal of Marketing
43 (3) (1979) 926.
[23] N. Mastrogiannis, B. Boutsinas, I. Giannikos, A method for improving the accuracy
of data mining classication algorithms, Computers and Operations Research 36
(10) (2009) 28292839.
[24] J. McCarthy, M. Hastak, Segmentation approaches in data-mining: a comparison
of RFM, CHAID, and logistic regression, Journal of Business Research 60 (6)
(2007) 656662.
[25] R. McDaniel, Management strategies for complex adaptive systems, Performance
Improvement Quarterly 20 (2) (2007) 2142.
[26] B. McKelvey, Avoiding complexity catastrophe in coevolutionary pockets: strategies
for rugged landscape, Organization Science 10 (3) (1999) 294321.
[27] J. Miglautsch, Application of RFM principles: what to do with 1-1-1 customers?
Journal of Database Marketing 9 (4) (2002) 319324.
[28] P. Naik, M. Hagerty, C. Tsai, A new dimension reduction approach for data-rich
marketing environments, Journal of Marketing Research 37 (1) (2000) 88101.
[29] E.W.T. Ngai, L. Xiu, D.C.K. Chau, Application of data mining techniques in custom-
er relationship management: a literature review and classication, Expert Sys-
tems with Applications 36 (2) (2009) 25922602.
[30] C. O'Reilly III, J. Harreld, M. Tushman, Organizational ambidexterity: IBMand emerg-
ing business opportunities, California Management Review 51 (4) (2009) 7599.
[31] D. Olson, Q. Cao, C. Gu, D. Lee, Comparison of customer response models, Service
Business 3 (2) (2009) 117130.
[32] D.L. Olson, D. Delen, Advanced Data Mining Techniques, Springer, Heidelberg, 2008.
[33] P. Rossi, R. McCulloch, G. Allenby, The value of purchase history data in target
marketing, Marketing Science 15 (4) (1996) 321340.
[34] G. Seni, J. Elder, Ensemble Methods in Data Mining: Improving Accuracy Through
Combining Predictions, Morgan & Claypool, 2010.
[35] M. Shaw, C. Subramaniam, G. Tan, M. Welge, Knowledge management and data
mining for marketing, Decision Support Systems 31 (2001) 127137.
[36] K.A. Smith, J. Gupta, Neural networks in business: techniques and applications for
the operations researcher, Computers and Operations Research 27 (11/12)
(2000) 10231044.
[37] E. Turban, J. Aronson, T. Liang, Decision Support Systems and Intelligent Systems,
7th ed. Prentice Hall, Upper Saddle River, NJ, 2004.
[38] E. Turban, R. Sharda, D. Delen, D. King, Business Intelligence: AManagerial Approach,
2nd ed. Prentice Hall, New York, 2011.
[39] B. Van den Berg, T. Breur, Merits of interactive decision tree building: part 1,
Journal of Targeting, Measurement and Analysis for Marketing 15 (3)
(2007) 137145.
[40] B. Van den Berg, T. Breur, Merits of interactive decision tree building: part 2: how
to do it journal of targeting, Measurement & Analysis for Marketing 15 (4) (2007)
201209.
[41] S. Vargo, R. Lusch, Evolving to a new dominant logic for marketing, Journal of
Marketing 68 (2004) 117.
[42] G. Verhaert, D. Poel, Empathy as added value in predicting donation behavior, Jour-
nal of Business Research 64 (2011) 12881295.
[43] P. Verhoef, P. Spring, J. Hoekstra, P. Leeang, The commercial use of segmentation
and predictive modeling techniques for database marketing in the Netherlands,
Decision Support Systems 34 (2002) 471481.
[44] B. Wierenga, The past, the present and the future of marketing decision models,
In: in: B. Wierenga (Ed.), Handbook of Marketing Decision Models, Springer,
2008, pp. 320.
[45] A. Yang, How to develop new approaches to RFM segmentation, Journal of
Targeting, Measurement and Analysis for Marketing 13 (1) (2004) 5060.
[46] H. Yun, D. Ha, B. Hwang, K. Ryu, Mining association rules on signicant rare data
using relative support, Journal of Systems and Software 67 (181191) (2003).
[47] J. Zahavi, N. Levin, Applying neural computing to target marketing, Journal of Di-
rect Marketing 11 (1) (1997) 522.
David L. Olson is the James & H.K. Stuart Professor in MIS and Chancellor's Professor at
the University of Nebraska. He has published research in over 100 refereed journal ar-
ticles, primarily on the topic of multiple objective decision-making and information
technology. He has authored 17 books, is associate editor of Service Business and
co-editor in chief of International Journal of Services Sciences. He has made over 100 pre-
sentations at international and national conferences on research topics. He is a mem-
ber of the Decision Sciences Institute, the Institute for Operations Research and
Management Sciences, and the Multiple Criteria Decision Making Society. He was a
Lowry Mays endowed Professor at Texas A&M University from 1999 to 2001. He was
named the Raymond E. Miles Distinguished Scholar award for 2002, and was a James
C. and Rhonda Seacrest Fellow from 2005 to 2006. He was named Best Enterprise Infor-
mation Systems Educator by IFIP in 2006. He is a Fellow of the Decision Sciences Insti-
tute.
Bongsug (Kevin) Chae is Associate Professor in Information & Operations Manage-
ment at Kansas State University. He has published papers in such areas as business an-
alytics, supply chain management, service innovation, and knowledge discovery. He
has made presentations in universities and global companies in several countries, pri-
marily on the topic of business analytics and intelligence, supply chain management,
and service innovation. He is a recipient of the Ralph Reitz Teaching Award and a nom-
inee of several other teaching awards at Kansas State University.
451 D.L. Olson, B.(K.) Chae / Decision Support Systems 54 (2012) 443451