Propensity Models
Propensity Models
Propensity Models
REFERENCES
Linked references are available on JSTOR for this article:
https://www.jstor.org/stable/4102197?seq=1&cid=pdf-reference#references_tab_contents
You may need to log in to JSTOR to access the linked references.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact [email protected].
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms
Operational Research Society, Palgrave Macmillan Journals are collaborating with JSTOR
to digitize, preserve and extend access to The Journal of the Operational Research Society
This content downloaded from 223.182.2.32 on Mon, 01 Jun 2020 08:23:51 UTC
All use subject to https://about.jstor.org/terms
journal of the Operational Research Society (2005) 56, 104 1-1050 ? 2005 Operational Research Society Ltd. All rights reserved. 0160-5682/05 $30.00
www.palgrave-journals.com/jors
We investigate the incremental roles of information that becomes available only after a revolving loan has been granted
in explaining and predicting the time taken until the borrower makes a second purchase. Using data relating to a store
card, granted around the time of first purchase and used in Belgium, we find that characteristics of a first purchase and
remaining credit available for use enhance the explanatory and predictive power of application characteristics. The
relationship differs between good and poor payers.
Journal of the Operational Research Society (2005) 56, 1041-1050. doi:10.1057/palgrave.jors.2601933
Published online 16 February 2005
Introduction swiftly if they were to return. Hence, once past a certain time
they were unlikely to return. Older and more sophisticated
Historically, credit scoring has concerned itself with asses-
customers tended to continue to respond over a longer
sing which individuals are a good risk and which are a poor
period and there was no obvious cutoff point.
risk, primarily focusing on the probability of default at any
The objective of the current study is to develop a model
point within a given time period after receiving a loan. A
that explains and predicts the time taken by the holder of a
number of authors now appreciate that a second aspect of
revolving credit product to make a second purchase. The
the risk of default, which is relevant to profitability, is the
analysis relates to a store card and its use in Belgium. The
time between the initial granting of credit and the time to
card is normally taken out at the time of a first purchase.
default."'2 Some high-risk applicants can generate a sig-
This has two benefits for the store: a purchase is made and a
nificant profit if they use the credit product actively and pay
relationship can develop. The credit relationship may also be
interest and charges for long enough before going into
attractive to the store and the lender. Clearly, the store will
default. On the contrary, low-risk applicants may pay the
be keen that the customer will return to make further
full balance every month, thus keeping the revenues from
purchases and this may be in the interest of the lender. The
such 'good' accounts low.
issues then are which individuals return and when they
This observation has led many to use techniques devel-
return. Such knowledge will give the opportunity to plan a
oped in survival analysis to predict the time to default.3-8 It
strategy to enhance the relationship and to gain mutual
has been found that survival analysis, especially Cox
benefit from the relationship for the store, the lender and the
Regression,9 compares well to logistic regression in provid- customer.
ing an ability to predict default, but it also gives insights into
In this study, we are particularly interested in the
the time to default. This can be particularly helpful in
explanatory and predictive power of information that
determining the profitability of a client.
becomes available after the card has been issued but which
Only a small number of studies have used a survival
is received in time to make future predictions. Stepanova an
analysis to predict the time to the next purchase of a
Thomas5 found that the importance of application and
product. Ansell et allo considered the behaviour of clients of
behavioural data in predicting time to default varied
an insurance company. The aim was to characterize their
throughout the age of the loan. In the current study, we
behaviour to decide on the marketing strategy for the
find that the combination of application, purchase and
individuals. The approach was to segment the population
behavioural characteristics predicts time to second purchase
into defined groups and then explore the behaviour of these
well, with behavioural characteristics becoming most im-
segments. They found that generally the less sophisticated
portant over time. Our analysis has to be read with slight
and younger segments tended to come back reasonably caution since it is based on a limited period.
A noteworthy finding is that the difference between the
*Correspondence. G Andreeva, Credit Research Centre, University of
Edinburgh, 50 George Square, Edinburgh EH8 9JY, UK. credit limit and the outstanding balance, therefore the
E-mail: [email protected] available credit to spend, has a major effect on the time until
This content downloaded from 223.182.2.32 on Mon, 01 Jun 2020 08:23:51 UTC
All use subject to https://about.jstor.org/terms
1042 Journal of the Operational Research Society Vol. 56, No. 9
a second purchase. We also find that future defaulters are Our research strategy is to model the hazard function
less likely to make use of the card than non-defaulters. using predictor variables that are available at each sequential
The structure of the paper is as follows. The next section stage of a customer's behaviour. First, we consider as
presents the basic model. The following three sections predictor variables only information available to the lender
present the effects of introducing information sequentially at the time of application. Second, the predictor variables are
gained by a lender during and after the application process. those available at the time the customer makes a first
Then we consider differences between good and poor payers. purchase, and third the predictor variables are those
The subsequent section examines the predictive performance available after the first purchase has been made.
of the variables and the final section concludes. The proportional hazards model that we use can be
described as follows. Let T be the time until a customer
Basic model makes a second purchase. The hazard function can then be
defined as
One can consider the information available to predict a
P(t < T < t + Atl TT >t)
customer's behaviour as being revealed in a series of h(t) At
sequential stages. First, there is the application for credit,
which provides the application data. At this stage bureau
That is, it is the instantaneous potential fo
data may also be available. At the next stage, there is extra
second purchase-to occur in the next instant
information in terms of the nature of the purchase and the
that it has not previously occurred. The prop
type of agreement entered into. This provides further insight
approach assumes
into the potential behaviour of the customer. Finally, there is
the customer's behaviour after he/she has been granted the h(t) = eP'Xho(t)
credit. In this study, our interest is focused on furtherwhere P is a vector of parameters to be estim
purchases and primarily the second purchase. baseline hazard. As was shown by Cox (1972)
Each individual in the analysis applied for a storecard,
can be estimated without knowledge of h
was given the card and has made the first purchase. The which the event did not occur within the obs
customer may then, within the observed time, either make a censored and the parameters can only be
are
second purchase, or make no further purchases and/or
non-censored cases.
default. Both 'Second Purchase' and 'Default' can be well The data used relate to a store card that is used in three
defined. Our definition of default is '2 consecutive missed
European countries, but our analysis considers its use only in
payments'. Figure 1 displays the behaviour of five typicalBelgium. After data cleaning, 25792 observations were
customers. Customer A makes a purchase within the study available for analysis. All card holders in the sample had
period and then continues on potentially to make furthersuccessfully applied for a card within a 14-month period in
purchases. Customer B defaults and does not continue.the late 1990s. The observation period ranged from 12 to 25
Customer C does not make a further purchase within themonths after the month of first purchase. The behaviour of
study period. Customer D defaults but then makes repay-some cardholders was not observed for the full 25 months
ments and so can make a further purchase before the end ofbecause they defaulted or they were issued with a card
the study period. Customer E makes the second purchase between periods 2 and 14 or because they closed their
and defaults afterwards. The simplest model would be to account within the observation period. The composition of
consider time to the second purchase ignoring the impact of
the sample between those whose observed purchase beha-
default. In this formulation, there is a single form ofviour was censored and those for whom it was not is shown
censoring by time. in Table 1. The predictor variables available at each of the
three stages described above are shown in Table 2. Their
values/levels were coarse-classified according to similarity in
Customer A - Purchase -- pi, where pj denotes the probability that those cases within a
coarse category, j, make a second purchase, and were
Customer B I Default
transformed into binary dummy variables, see Thomas
Customer C et al." Model parameters were estimated using a randomly
selected training sample (18040 cases) and the predictive
Customer D ------+ Defau
performances of the estimated models were tested on the
Customer E 0- Purchase
remaining 7752 cases. -*
This content downloaded from 223.182.2.32 on Mon, 01 Jun 2020 08:23:51 UTC
All use subject to https://about.jstor.org/terms
G Andreeva et a/-Modelling the purchase propensity 1043
Table 1 Sample used in the analysis status. The self-employed are less likely to purchase than
other groups and those working part-time are more likely
Performance Total 2nd purchase Censored % Censored
than others. One may say, however, that those renting and
Good 22 708 16 648 6060 26.69 those only recently at the address are more likely to purchase
Bad 3084 1634 1450 47.02
on the card. These might be explained in terms of the
Total 25792 18282 7510 29.12
individuals' financial context, that is, these customers can be
credit constrained.
The baseline survival curve in Figure 2 shows that 50% of
Table 2 Variables used the
in the have
customers analysis
made a second purchase within 7 months
and about 30% have not made a second purchase after 25
Application variables
Home telephone Time in employment months. These percentages will be affected by those
Residential status Type of business customers who have defaulted and those who have not been
(manufacturing, banking,
observed over the whole period.
catering, etc)
From a marketing view it would be helpful to know
Marital status Employer's phone
whether
Occupation (full-time, part-time, Spouse age either of these two groups could be differentiated.
self-employed, etc) The former group is making use of the card; the latter might
Age Number of dependents be deemed non-users.
Time at address since 18 years old
This content downloaded from 223.182.2.32 on Mon, 01 Jun 2020 08:23:51 UTC
All use subject to https://about.jstor.org/terms
Table 3 Parameter estimates of propor
Model
App
App + App + App + Purchase + From 2nd Purchas
App
Application + Purchase
Purchase + Purchase
ATS,=f ATS_ A + ATSt_
TSt-2 I + period-m
variables Goods
Variable 1 2 3 4 5 6 7
Price
10 000 < 10 000
Price < 16BEF 0.630
000 BEF 0.208
0.487 0.085 0.487 0.706
0.367 0.529 0.695
0.471 0.110 0
16 000 <Price < 20 000 BEF 0.328 0.224 0.316 0.274
20 000 < Price < 40 000 BEF 0.205 0.153 0.246 0.227 0.058
Contract type-budget 0.726 0.854 0.775 0.452 0.675 0.878
ATSt-,--over credit limit -1.095 -3.867 -4.607 -2.540 -1.107
ATSt_ 1 = 0 -0.235 0.574 0.624 1.223 -0.171 -0
ATS,_1 5000 -0.423 -0.318 -1.156 -0.451 -0.359 -
This content downloaded from 223.182.2.32 on Mon, 01 Jun 2020 08:23:51 UTC
All use subject to https://about.jstor.org/terms
Table 4 Hazard ratios of proportional hazard models wit
Model
This content downloaded from 223.182.2.32 on Mon, 01 Jun 2020 08:23:51 UTC
All use subject to https://about.jstor.org/terms
1046 Journal of the Operational Research Society Vol. 56, No. 9
1.0
e 0.8
- 0.6
0.2
0.0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 I 15s 16 17I 18 19 20 21 22 23 24 25 6
Time
Figure 2 Baseline survival distribution function with 95% confidence intervals, application data.
Table 5 Log Likelihood statistics for models with different information levels
This content downloaded from 223.182.2.32 on Mon, 01 Jun 2020 08:23:51 UTC
All use subject to https://about.jstor.org/terms
G Andreeva et ai-Modelling the purchase propensity 1047
The results are presented in Tables 3 and 4, column 4. The dynamics. Therefore, additional information was included
improvement in Log Likelihood over the no-covariate model into the analysis. For each period of time, three more
is 2281, emphasizing the importance of recent ATS overvariables were added: an indicator of one missed payment,
initial ATS. Marital status and Credit insurance leave the
an indicator of two or more missed payments (our definition
model, and while other variables are affected by inclusion of
of default) and the percentage of the outstanding balance
ATS,_1, their coefficients change only slightly. that was repaid during the period preceding the purchase.
To check the character of ATS, the model was fitted withFigure 4 shows a slow decline in the impact of the
ATS two periods before the purchase, ATSt-2, andapplication
both score. This is expected since the information
ATS 1 and 2 periods before the purchase. Table 6 compares
becomes more historic. This highlights the difference
the Log Likelihood statistics for the model with ATS lagged
between this investigation and previous work on defaults
and not lagged, applied to the sample of cardholders onthat
fixed terms loans (see Stepanova and Thomass). The
make the second purchase, starting from period 3. behavioural aspects become of greater importance over time.
The inclusion of both lagged and current ATS results Percent
in a repaid becomes significant at period 4 but loses its
quite notable increase in the amount of variation accounted
significance temporarily at period 7. Delinquency status,
for, and both variables are significant (Tables 3 and
both 4,
1 and 2 periods, remains significant through all 10
column 5). This suggests that ATS cannot be used simply in Period 10 was the last period for which a model was
periods.
the model as a first-order markovian variable. estimated because the risk set became too small for
subsequent periods.
Investigating the dynamics of A TS The model with time-dependent behavioural variables w
fitted to the risk set of cardholders who make a second
The importance of behavioural variables in describing the
time to the second purchase called for closer investigation of
purchase at period 3 and beyond. The parameter estimates
and hazard ratios are shown in Tables 3 and 4, column 6. It
the dynamics of ATS. A dynamic model was developed,
where parameters were re-estimated for each period of time.5 is interesting to note that Percent repaid was not selected by
The model that was estimated can be written as the stepwise procedure. This reflects its unstable behaviour
in the dynamic model where it was initially not significant
hs(t) = eP'(s)x+Y(s)zt hs(t) and later became significant.
Table 6 Log Likelihood statistics for the ATS lagged and not lagged
This content downloaded from 223.182.2.32 on Mon, 01 Jun 2020 08:23:51 UTC
All use subject to https://about.jstor.org/terms
1048 Journal of the Operational Research Society Vol. 56, No. 9
001-10000 BE10,001-20000 BE
a,
L
0i -e PLOT No amount to spend ----5000 BEF
-3
I 2 3 ' I 6 7 8 9 10
PERIOD
-2
15
2 3 I
6 5 6
9 7 810
9 10
PERIOD
similar
were small, the estimation was carried out using the to that previously seen. For Bads, however, th
training
and hold-out samples combined. models appear with the entry of whether a person h
phone 7-9.
The results are presented in Tables 3 and 4, columns for both categories of Bads, while Number of ch
There are differences between the models arising appears
from thefor those who default before purchase.
Residential
samples. As might be expected, the model for the Good is status, Industrial sector, Card and Credit
This content downloaded from 223.182.2.32 on Mon, 01 Jun 2020 08:23:51 UTC
All use subject to https://about.jstor.org/terms
G Andreeva et al-Modelling the purchase propensity 1049
1.0
0.8
0.6
LL
L.............
0,0r
S 3 5 6 7 8 9 10 11 12 13 1I 15 16 1I 18 19 20 21 22 23 24 25 e6
Ti u(e
Bad ---.Good
Figure 5 Baseline survival distribution function for good/bad with 95% confidence intervals.
Table 7 Area under the ROC-curve for models with different information levels
Hold-out
Application 0.553
Application + product 0.689
Application + product + ATS at period 1 0.715 0.652 0.626 0.608 0.596 0.586
Application + product + ATS time-varying 0.710 0.662 0.638 0.621 0.610 0.600
Application + product + all behavioural info time-varying 0.643 0.638 0.658 0.647 0.648
Dynamic model, ATS,_1 0.670 0.654 0.649 0.641 0.635
Dynamic model, all beh. variables 0.669 0.658 0.654 0.654 0.650
This content downloaded from 223.182.2.32 on Mon, 01 Jun 2020 08:23:51 UTC
All use subject to https://about.jstor.org/terms
1050 Journal of the Operational Research Society Vol. 56, No. 9
make the second purchase immediately in the next period. The use of the card also indicates the likely behaviour of
As time progresses, behavioural information gives a lot of the individual. Defaulters who may be credit constrained are
additional predictive power. For example, the difference in less likely to use the card than non-defaulters. There seems to
the area under the ROC curve between each application, be differentiation between models for prediction for the
product and ATS model and, for the same time period, the Good and for the Bad, hence again indicating different
same variables plus behavioural information, increases as we behaviour. For the Bad, there are also differences between
consider future time periods. Thus for hold-out 2, the those who default before the second purchase and those who
difference between AUROC for the model including default after second purchase. Such differences may also
application, product and all behavioural variables and help in identifying patterns so that action can be taken to
AUROC for the model including application, product and protect the lender.
initial ATS is (0.643-0.652)= -0.009, which means that
adding behavioural variables at this stage does not improve
the prediction, but makes the model less parsimonious. Acknowledgements-GA is grateful to the University of Edinburgh
and Credit Research Centre for financial support during this work, and
Whereas for hold-out 6 the same difference is
to ESRC for dissemination funding (PTA-026-27-0216). We are
(0.648-0.586) =0.062, which means that a probability
grateful to the of
referees for their useful and constructive comments
on the
randomly selecting a good account with a higher orpaper.
equal
ranking as compared to a randomly selected bad account,
has increased by 6.2%. The best results can be achieved by
incorporating all behavioural information dynamically. The
References
model with the largest area under the ROC curve for each
hold-out is identified in bold. Four out of five of these occur
1 Hopper MA and Lewis EM (1992). Development and use of
when a dynamic model is used. This reinforces the view that credit profit measures for account management. IMA J Math
over time the behavioural information becomes more Appl Business Ind 4: 3-17.
important in determining actions. 2 Leonard KJ (1997). Behavior scores to predict profitability. In:
Thomas LC, Crook JN and Edelman DB (eds). Proceedings of
the Credit Scoring and Credit Control V Conference. Credit
Research Centre: University of Edinburgh.
Conclusion 3 Narain B (1992). Survival analysis and the credit granting
decision. In: Thomas LC, Crook JN and Edelman DB (eds).
This paper presents an exploration of customer behaviour in Credit Scoring and Credit Control. Oxford University Press,
using a retail card. Modelling the behaviour of the customer Oxford, 109-122.
4 Banasik J, Crook JN and Thomas LC (1999). Not if but when
will help to establish the customer's profitability and hence
will borrowers default. J Opl Res Soc 50: 1185-1190.
the benefit for the lender. The paper has demonstrated how
5 Stepanova M and Thomas LC (2001). PHAB scores: propor-
the data, which are available at various early stages of the tional hazards analysis behavioural scores. J Opl Res Soc 52:
life cycle of a card, can be used. The results obtained are 1007-1016.
similar in a large part to those seen in using survival6 Stepanova M and Thomas LC (2002). Survival analysis
techniques for predicting default. There are differences that methods for personal loan data. Opns Res 50: 277-289.
reflect the need for certain customers to use the card to 7 Stepanova M (2001). Using survival analysis methods to
build credit scoring models. PhD thesis, University of
enhance their spending power. Southampton.
The new information at each stage enhances the modelling8 Hand DJ and Kelly MG (2001). Lookahead scorecards for new
with a greater volume of variation accounted at each stage. fixed term credit. J Opl Res Soc 52: 989-996.
It is also notable that the coefficients are generally only 9 Cox DR and Oakes D (1984). Analysis of Survival Data.
Chapman & Hall: London.
slightly modified. Thus, it is important that a lender builds
10 Ansell JI, Archibald T and Harrison T (2001). Lifestage, lifestyle
subsequent models after the application model because the and the impact on client propensity in financial services,
incorporation of subsequent data would enhance the Working Paper 01/3, Credit Research Centre: University of
predictive performance of the model. Edinburgh.
The variables that seem to have a major effect are the 11 Thomas LC, Edelman DB and Crook JN (2002). Credit Scoring
contract type and ATS. In this paper it has been seen that and its Applications, Society for Industrial and Applied
Mathematics: Philadelphia, PA.
ATS does not have a first-order markovian property with12 Hand DJ (1998). Consumer credit and statistics. In: Hand DJ
the most recent value accounting for the whole effect of the and Jacka SD (eds). Statistics in Finance. Edward Arnold
variable. While the customer requires a positive credit London, pp 69-81.
13 Bamber D (1975). The area above the ordinal dominance graph
balance before repeatedly using the card, it is also important
and the area below the receiver operating characteristic graph.
to know how the customer got to this stage and if they were
J Math Psychol 12: 387-415.
generally in surplus before. It is important to note that, 14 Hanley JA and McNeil BJ (1982). The meaning and use of the
increasingly throughout the period, behavioural aspects have area under a receiver operating characteristic (ROC) curve.
major impacts on the behaviour of a customer. Radiology 143: 29-36.
This content downloaded from 223.182.2.32 on Mon, 01 Jun 2020 08:23:51 UTC
All use subject to https://about.jstor.org/terms