You are on page 1of 8

European Journal of Operational Research 183 (2007) 1513–1520

www.elsevier.com/locate/ejor

To ask or not to ask, that is the question


Hsin-Vonn Seow *, Lyn C. Thomas
School of Management, University of Southampton, Highfield SO17 1BJ, UK

Received 1 January 2005; accepted 1 August 2006


Available online 22 January 2007

Abstract

Applicants for credit have to provide information for the risk assessment process. In the current conditions of a satu-
rated consumer lending market, and hence falling ‘‘take’’ rates, can such information be used to assess the probability of a
customer accepting the offer?
With the advent of internet broking pages, which allow borrowers to ‘‘apply’’ to a number of different companies at the
same time, this ‘‘take’’ problem will increase. In some mortgage markets, it is quite common for more than 50% of those
offered credit to reject it. In some cases, this is because the sale falls through but often it is because a relatively better retai-
ler has offered a more suitable product to the borrower.
Lenders do not want to make the application process too complicated, and with the growth in adaptive marketing chan-
nels like the Internet and the telephone, they can make the questions they ask depend on the previous answers. We inves-
tigate how one could develop such ‘‘adaptive’’ application forms; which would assess acceptance probabilities.
Ó 2006 Elsevier B.V. All rights reserved.

Keywords: Data mining; Classification trees; Question selection

1. Introduction This competition among the different banking


and financial institutions means that ‘‘take’’ – the
The personal financial market is becoming a probability a customer will accept the product
buyer’s market rather than a lender’s market [6], offered – is now a major problem in personal
as customers can choose between a number of com- finance. One way of maximizing the ‘‘take’’ of offers
peting products. New organizations like internet is by customizing the product offered for each cus-
brokers mean that the customer is able to apply to tomer. There have been several areas where Opera-
a number of financial institutions. They can choose tional Research approaches are useful in this
to take the offer that appeals most to their needs and process of customization as is indicated in Murthi
hence lenders need to try and make appropriate and Sarkar 5. Such research focuses on the follow-
offers. ing question: What offer, acceptable to a lender,
should be made to each population segment so
*
as to maximise the chance it will be taken. Thus
Corresponding author. Tel.: +44 23 8059 3557; fax: +44 23
there are two problems here – how to segment
8059 3844.
E-mail addresses: yseow@yahoo.com (H.-V. Seow), L.Tho- the population and which offer to make to each
mas@soton.ac.uk (L.C. Thomas). segment.

0377-2217/$ - see front matter Ó 2006 Elsevier B.V. All rights reserved.
doi:10.1016/j.ejor.2006.08.061
1514 H.-V. Seow, L.C. Thomas / European Journal of Operational Research 183 (2007) 1513–1520

The telephone and the internet are increasingly ment that is both cost conscious and appropriate in
being used in personal financial services as ways of the treatment of diabetic retinopathy [3], and to
applying for products. Such channels of communica- detect possible complications for pregnancies [4].
tion are both private and interactive and thus allow The strength of classification trees is that they
different variants of a product to be offered to differ- deal well with interactions among the characteristics
ent customers in real time. Of course before any that are being used to classify. They identify what
offers can be made, some information on the cus- are the significant characteristics in the data and
tomer has to be collected and compared with infor- what combinations of the attributes of these charac-
mation on what offers previous customers accepted teristics give the best forecasts of the outcomes of
or rejected, see for example [7]. This is done by ask- the target variable. To construct a tree, one starts
ing the customer some questions, and to ensure that by splitting the population in the data set into two
the customer is not put off by the application pro- subpopulations which are as different as possible
cess, it is important that the number of questions in the frequency of the outcomes of the target vari-
posed is kept to a minimum. In the case of loan prod- able. This is done by looking at each characteristic
ucts, like credit cards, the customer will be expecting in turn and identifying which split of its attributes
to answer some questions as part of the credit scor- is optimal in differentiating between the outcomes
ing risk assessment requirements, but they will not of the target variable and then choosing the charac-
want to answer too many questions in total. teristic which has the best optimal split. This process
The interactive nature of the new application is then repeated on the daughter subpopulations so
channels means that the questions asked can depend formed and subsequently on their daughter subpop-
of the responses to earlier questions. So the exercise ulations until one ends up with subpopulations
is akin to an ‘‘adaptive questionnaire’’. Bearing in which meet some end node conditions. Data that
mind that only a limited number of questions are satisfies the conditions that make up each end node
to be asked, it is important to identify which ques- are then classified as having the target outcome
tions will result in the most relevant data to be which is in a majority in that end node. Note that
collected. one can easily modify classification trees so that at
So, which questions should be asked? Then what a particular node the search is only over a subset
variant of the product should be offered in the light of the characteristics. This property will be
of the responses to the questions asked? In this exploited in the next section to develop a methodol-
paper, we propose using classification and regression ogy that allows one to segment the population and
trees (CART) to answer these two questions. Section choose the best offers for each segment.
2 reviews classification trees and Section 3 describes
TAROT, a way of modifying that methodology so as 3. TAROT – Top application characteristics
to answer these questions. Section 4 gives an exam- remainder offer characteristics tree
ple of this approach to acceptance scoring using data
from a Fantasy Student bank account and some con- The classification tree that is proposed consists of
clusions are drawn in section 5. segmenting the population using the characteristics
of the applicants and then classifying the offer char-
2. Classification and regression trees (CART) acteristics so as to decide which offer to make to
each segment in order to maximise the ‘‘take’’ prob-
Classification trees developed by Breiman et al. ability. We refer to such a tree as TAROT (top
[2] remains a powerful classification tool, especially application characteristics remainder offer charac-
with the new advances like random forests in its use teristics tree).
[1]. In classification one has a set of data, where A sample of previous offers, the characteristics of
each data point consists of the values (the attri- the applicants who were made offers and the take/
butes) of a number of characteristics variables not take decisions of the applicants to those offers
together with the value of a target variable, which is used to build the tree. The target variable is
is often binary. The aim is to identify which combi- whether the offer is accepted by the applicant. If
nations of attributes lead to the different target out- one built a classification tree without any constraint
comes. It is used widely in many areas apart from then it is likely that there will be offer characteristics
credit scoring and marketing [8]. In health for exam- O1 which are higher up the tree than applicant char-
ple, CART analysis is being used to decide on treat- acteristics A1 when one is trying to identify the
H.-V. Seow, L.C. Thomas / European Journal of Operational Research 183 (2007) 1513–1520 1515

accepting group. From this, one can say very little. in terms of the take/not take decision and look at
One cannot say that O1 is the best offer to make to what happens if these types of offers are made. This
someone with A1 characteristics as there might be will restrict the number in the sample who will be
something else that appeals to that segment more considered at each step but the usual policy of defin-
but is less attractive to the whole population. All ing a node as an end node of the tree if there are
the tree indicates is the difference between having insufficient cases should prevent analysis being done
O1 in the offer or not having it in, is greater on on too small samples. One can use the following
the whole population that any other offer. iterative methodology to build the TAROT tree.
A (m, n)-TAROT assumes that there is a maxi- We do not assume it has an optimal segmentation
mum of m levels of applicant characteristics and n and subsequent offers in any sense but it is a good
levels of offer characteristics in the tree. Such a tree compromise between finding ‘‘locally optimal’’ solu-
would lead to an offer process where one uses at tions and offsetting the bias that occurs in such
most m questions in order to segment the popula- ‘‘local optimal’’ solutions because they do not rec-
tion and there are at most n features of the product ognise that the data consists of different offers to
that will be varied from the standard as part of the the different customers.
offer. It is necessary to define a standard value for
each of the features of the offer (usually the least 3.1. Building TAROT
expensive to the lender) and it is assumed that this
is what is offered if that offer characteristic does (a) Build an ‘‘up to m-level offer characteristic
not occur in the tree. Fig. 1 is an example of a only tree’’ on take/not take decision. From
(1, 1)-TAROT in the context of a student current this, one can identify the combination of fea-
account application, where the question is on the tures that gives the highest acceptance rate
course the student is undertaking. If the student is and which has been offered to sufficient
taking Arts or Other, then the offer will include free applicants.
(i.e., no fee) foreign exchange purchases, while if (b) Use this highest offer combination at the top
they were taking Economics or Education, they of a tree and add a last level of the best split
should be extended an offer for a credit card. of the application characteristics.
The difficulty in using a classification tree in the (c) Now take this best ‘‘bottom level’’ application
normal way, even with the first level split being characteristic split and force it in at the top of
forced to be an application characteristic is that a classification tree and allow up to m lower
these splits are made without considering that there levels of offer characteristics splits.
may be a bias in the take/not take decision because (d) Take the branch of the m-level offer charac-
different offers were made to different applicants. So teristics splits that lead to the highest accep-
we need to restrict this bias in some way even if we tance rate and is applied to a sufficiently
cannot remove it entirely. Thus we try to identify large proportion of the population – one
what are the most important features of the offer branch from each of the subsets arising from

Fig. 1. An example of a (1, 1)-TAROT.


1516 H.-V. Seow, L.C. Thomas / European Journal of Operational Research 183 (2007) 1513–1520

the application characteristic splits – and add ered from a fantasy student account which was
a bottom level application characteristic split. started in 2001. This is a web based application form
The choice of what percentage of the popula- of a fantasy student bank account which was made
tion ‘‘is sufficiently large’’ is a parameter of available to the first year students at the University
the methodology. of Southampton. There were 21 applicant character-
(e) Force in these ‘‘bottom level’’ application istics that can be obtained from this information.
splits below the current application levels in They are sex, status, number of children (Num_Chil-
the tree and then allow up to m levels of offer dren), number of credit cards (Num_Cards), whether
characteristic splits. there is any income from wage, loan, or parental con-
(f) Repeat steps (d) and (e) until the n levels of tributions, other sources of income, place of educa-
application characteristics have been intro- tion (establishment), campus, course taken (course)
duced or the characteristic split in (d) is not and interests: sport, travel, music, clubbing, cinema,
significant. cars, Do-It-Yourself activities (DIY), gardening,
beer and country western (see Table 1).
When the tree is completed the top level applica- The website then made the students an offer of a
tion splits give the segmentation of the application specific account, where the offer characteristics of
population, while the lower level offer splits describe overdraft limit, interest when in credit, commis-
the different changes that have to be made to the sion-free travel money, and whether a credit card
standard offer for each of these segments. The phi- was also offered varied from student to student.
losophy in the tree is, at each stage, to concentrate The offer characteristics were chosen randomly in
on the offers with the highest acceptance rates when 25% of the cases and followed a decision tree which
deciding which further segmentation of the applica- was constructed to ensure a variety of offers was
tion characteristic to make. being made in the other 75% of the cases. Further
Taking into account the requirements of the details can be found in Thomas et al. [9].
algorithm, there was a need for a statistical software After pre-processing the initial data, 300 entries
package that had the ability to force splits on sub- were used to build the classification tree. There is
sets of the variables. SAS 9.1.3. has Enterprise no need for a validation sample since the tree is of
Miner 4.3 that allows the user to experiment with fixed size. We are aware that this is not a large sam-
forced variables in the classification trees and is very ple and hence limit the TAROT by only building a
user friendly and presents results well in the forms (2, 1)-TAROT.
of charts and graphs. There is a choice of splitting As one might expect, the data is difficult to clas-
rules to choose from consisting of features like sify well since one is trying to predict a difficult tar-
chi-squared test and entropy to show how one var- get variable like acceptance, especially in the case of
iable is more significant than another. the acceptance of a hypothetical product. Moreover
our aim is to establish a methodology for this type
4. Using TAROT on fantasy student account data of problem, not to derive the features since this data
is for a hypothetical, not a real product.
To the best of our knowledge, there is no publicly The results of building an unrestricted classifica-
available data on the offer strategies of the personal tion tree on this data using all 25 applicants and offer
financial sector (see [9]). So we utilise the data gath- characteristic variables are given in Fig. 2. Note that

Table 1
A sample of the data from the fantasy student account
pkUserNumber Sex Status Num_Children Num_Cards Wage Loan
1 Male Single 0 1 False True
2 Male Divorced 0 3 True True
6 Male Married 0 1 True False
7 Female Single 0 6 True True
8 Male Single 0 1 True True
9 Female Other 0 0 True True
10 Female Single 0 1 False True
11 Female Single 0 0 False True
12 Female Single 0 0 False True
H.-V. Seow, L.C. Thomas / European Journal of Operational Research 183 (2007) 1513–1520
Fig. 2. Applicant and offer characteristics classification tree (with no ‘‘initial classification using applicant characteristic’’ rule).

1517
1518 H.-V. Seow, L.C. Thomas / European Journal of Operational Research 183 (2007) 1513–1520

this tree as indicated earlier does not help decide important characteristic. We then look at those
which offer to make to which applicant segment as who were offered travel money and find the split
one can have offer characteristics above applicant on application characteristics that maximises the
ones. In that case of the same applicant characteristic difference in acceptance rate. This turns out to be
follows two offer characteristics, which offer should whether the course undertaken by the applicant,
you make to that group? For example, from Fig. 2, whether it is Arts or Others, or Economics or Edu-
should males with more than three credit cards be cation (see Fig. 4). We choose Course as the first
given the free travel money (fee-free foreign applicant split and place it at the top of the tree.
exchange) offer or not since they appear as the best Then, we find the best offer only split for each of
group under both the offer ‘‘fee-free’’ travel money the two subsets. This gives the (1, 1)-TAROT, dis-
branch and under the not offer branch as well. played in Fig. 1, where free foreign exchange is
The initial step in a (m, n)-TAROT is to identify given to those who are studying Arts or Others
which n-level offers give the best acceptance rate and an offer of credit cards to those studying Eco-
on the whole population. The standard offer is to nomics or Education.
offer an overdraft of £1000 and no further enhance- To obtain the (2, 1)-TAROT, repeat the analysis
ments like fee-free foreign exchange (travel money), for a (1, 1)-TAROT initially that is take the best
no credit card packaged with the account, no inter- offer split from Fig. 3 which is TravelMoney. We
est when the account is in credit and no gift given on next split each of the subsets of the TravelMoney
opening the account. We build an offer-only tree to splits on applicant characteristics (see Fig. 4). The
see which of these features should be amended and best applicant split chosen is Course. We ignore
get the result as shown in Fig. 3 for the case n = 2. the split from no travel money as it does not have
Fig. 3 suggests that the best offer even at the n = 2 as high acceptance rate.
case is just to offer travel money (68% acceptance). We then place the applicant variable Course at
This is also the offer one would make if n = 1. the top of the tree. So for the branch leading from
To build a (1, 1)-TAROT, we use this one-level studying Arts or Others, the offer characteristic of
offer tree in Fig. 3 to identify travel money as the giving fee-free travel money is the best offer split.

Fig. 3. Offer characteristics only classification tree.

Fig. 4. Finding the first applicant split for the (2, 1)-TAROT.
H.-V. Seow, L.C. Thomas / European Journal of Operational Research 183 (2007) 1513–1520 1519

Then we find the best applicant split for this corre- allow one level of any offer characteristic split for
sponding subset is whether the applicant has two or the Number of Cards and Wage subsets. The result
less credit cards. is presented as a complete (2, 1)-TAROT shown in
As for the branch leading from the course taken Fig. 6.
is Economics and Education, the offer characteristic This says that the best adjustment to the standard
of a credit card has a higher acceptance rate. The offer for those who are studying Arts or Others and
best applicant split corresponding to this is Wage have two or less credit cards is to offer fee-free foreign
(whether the students earn wages or not) (see exchange. For those in the same Course set but have
Fig. 5). By placing both the Number of Cards and three or more credit cards, the best offer split will be
the Wage splits under the Course split, we then an overdraft of £1400. For those studying Economics

Fig. 5. Step (d) of (2, 1)-TAROT.

Fig. 6. (2, 1)-TAROT.


1520 H.-V. Seow, L.C. Thomas / European Journal of Operational Research 183 (2007) 1513–1520

or Education, who have a wage should be offered Although not yet widely used, this idea of tailoring
interest of 1% when in credit, while those who do an offer to increase the chance of acceptance is
not should be offered an offer that brings the optimal becoming more important in selling financial prod-
amount of profit to the lender as they are predicted to ucts through the internet and telephone application
accept any offer extended to them. channels.
In this tree we end up with subsets with small
numbers of applicants in them, which can lead to References
over fitting. One way of increasing the size of the
data set is to build a regression model of the chance [1] L. Breiman, Random forests, Machine Learning 45 (1) (2001)
of accepting an offer as a linear function of the offer 5–32.
and applicant characteristics. This would allow esti- [2] L. Breiman, J. Friedman, R. Olshen, C. Stone, Classification
and Regression Trees, Wadsworth, Monterey, CA, 1984.
mation of the probability of acceptance of any com-
[3] P.R. Harper, M.G. Sayyad, V. de Senna, A.K. Shahani, C.S.
bination of offer characteristics by any individual in Yajnik, K.M. Shelgikar, A systems modelling approach
the application sample – not just whether they for the prevention and treatment of diabetic retinopathy,
accepted the one offer they were made. Clearly the European Journal of Operational Research 150 (1) (2003)
probabilities are only an approximation but at least 81–91.
[4] P.R. Harper, D.J. Winslett, Classification trees: A possible
when there is a split on an offer characteristic, there
method for maternity risk grouping, European Journal of
now would be no loss in population size. Cluster the Operational Research 169 (1) (2006) 146–156.
population into segments so that appropriate offers [5] B.P.S. Murthi, S. Sarkar, The role of the management sciences
can be made to each segment. Here appropriate in research on personalization, Management Science 49 (10)
means the offers which are the highest acceptance (2003) 1344–1362.
[6] K. Purang, Datamonitor Report, London, 2005.
rate for that segment.
[7] T.S. Raghu, P.K. Kannan, H.R. Rao, A.B. Whinston,
Dynamic profiling of consumers for customized offerings over
5. Conclusions the internet: A model and analysis, Decision Support Systems
32 (2001) 117–134.
We have shown that the TAROT classification [8] L.C. Thomas, D.B. Edelman, J.N. Crook, Credit Scoring and
Its Applications, Society for Industrial and Applied Mathe-
trees can be used to advise on which offer to extend
matics, SIAM, Philadelphia, 2002.
to each cluster formed from a classification using [9] L.C. Thomas, K.M. Jung, S.D. Thomas, Y. Wu, Modelling
applicant characteristics. This paper shows the fea- consumer acceptance probabilities, Expert Systems and their
sibility of this approach to acceptance scoring. Applications 30 (2006) 499–506.

You might also like