Professional Documents
Culture Documents
10 2307@4102184 PDF
10 2307@4102184 PDF
10 2307@4102184 PDF
REFERENCES
Linked references are available on JSTOR for this article:
http://www.jstor.org/stable/4102184?seq=1&cid=pdf-reference#references_tab_contents
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at http://www.jstor.org/page/
info/about/policies/terms.jsp
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 support@jstor.org.
Operational Research Society and Palgrave Macmillan Journals are collaborating with JSTOR to digitize, preserve and extend
access to The Journal of the Operational Research Society.
http://www.jstor.org
This content downloaded from 141.211.4.224 on Sat, 09 Jan 2016 05:32:42 UTC
All use subject to JSTOR Terms and Conditions
Journal of the Operational Research Society (2005), 836-843 k 2005 OperationalResearchSociety Ltd.All rightsreserved.0160-5682/05 $30.00
www.palgrave-journals.com/jors
credit;scoring;scorecard;
Keywords: screening
This content downloaded from 141.211.4.224 on Sat, 09 Jan 2016 05:32:42 UTC
All use subject to JSTOR Terms and Conditions
H-TTsaiet a/-Aneconomic
model
forcredit
assessment
problem837
using discrete scale and individual cut point for each cannot be measured directly. Variable Y is linked to the
variable,has some limitationsand is not easily extendedto observed binary variable Y* by the following index
the case with more than two variables. function:
From the scorecard users point of view, the issue of
combining all possible credit information is taken into 1 if Y>k
considerationto predict defaultingprobabilityand then to 1- 0 if Y (1)
maximize profit. In this article, screeningproceduresusing where k, is the upper specification•k or threshold of Y. An
individual cut and linear cut approaches are proposed to
applicantwith Y> k, will default and the others will not, so
deal with the above issue. Furthermore, an economic the overall default rate is given by i = P[Y> ky].
screening model based on default rate is proposed and Without loss of generality,let (Y, X1, X2, ..., Xk) have the
optimal cutoff points are determined so that the better standardmultivariatenormal distributionwith a correlation
applicantscould be selectedby maximizingtotal profit. matrix
1 Poi P02 . . . POk
Screening technology Pol 1 P12 . . . Plk
P02 P12 . . P2k 1
Screeninghas been wildly used in quality control, psychol- F-12 (2)
ogy, education, and medicine and so forth. It is a way of -21
122
This content downloaded from 141.211.4.224 on Sat, 09 Jan 2016 05:32:42 UTC
All use subject to JSTOR Terms and Conditions
838 Journal
oftheOperational
Research Vol.56, No.7
Society
Selected Selected
region Sregion
x = k2
X2
C
alxl+a2X2=
\
X x =
k1
X, X,
Therefore, the object is to determine the optimal cutoff Note that the conditionaldistributionof Y given X, = xl and
points of xl >kl and x2 > k2 so that the overall default X2 - X2 is also normally distributedwith mean a xI + a2x2
rate afterscreeningis minimized.Namely, an optimal(k1,k2) and varianceUa (Graybill,16p 106).
satisfiesthe following nonlinearprogrammingproblem: Therefore,the objectis to determinean acceptanceregion
of alx1 + a2x2< c to minimize the overall default rate after
Minimize rl, (5) screening. The screening procedure using the linear cut
Subjectto f-I constant
approachhas the following steps:
Clearly,if h1is prespecified,then (5) turns out to be a single (1) Collect and compute the values of X1 and X2.
parameterproblem. (2) An applicantis accepted if alxl + a2x2 c, otherwiseis
rejected.The acceptancerate is
(b) Multiple screening variables
The effect of prediction will be improved when the #2 - P[alXi + a2X2 <c] (8)
number of screening variables is increased. Suppose
there are k screening variables available, (5) would turn (3) An applicantwho has been acceptedwill be
out to be a (k-1)-parameter problem that is more difficult (i) bad, if y > ky;
to solve. (ii) good, if y <k,
This content downloaded from 141.211.4.224 on Sat, 09 Jan 2016 05:32:42 UTC
All use subject to JSTOR Terms and Conditions
forcredit
model
H-TTsaieta/-Aneconomic problem839
assessment
Q-i
1 (12)Numerical example
Now a numericalexample simulatedfrom the example in
Now, the problem has been reduced to the case of one O'Conner6is provided to illustratethe use of the proposed
screeningvariableand v/i is called 'equivalencecorrelation'. model. A financialcompanyis promotinga loan programme
However, the equations of acceptance rate and overall in which some better customers will be selected from
default rate after screeningare similarto (8) and (9) except N-= 10000 applicantsand certain amount of credit will be
that the acceptanceregion is substitutedby grantedto them. In this programme,a good customerwill
E••laiXic.
generate average net revenue of $1000 per person while a
defaulting person will cause an average loss of $5000 per
Economic screening model person. The company'sobject is to maximizetotal profit of
In the previous section, the default rate is minimizedfor a the programmeby selectingbettercustomers.
In order to assess the applicant's credit, the company
given acceptance rate f/. However, in many practical
situations, the company might be more interested in bought a creditscore and individualdefaultrecordfrom 'X'
credit bureau and treated it as a single variable.Experien-
maximizingthe total profit of a certainprojectwith/without
the constraintof p. Hence, an economic screeningmodel for tially, the income ratio, one of the applicationdata provided
credit assessmentproblemsis proposed to incorporatewith by all applicants, was an additional potential predictive
different screeningproceduresin this section. To facilitate variable.Hence, both the credit score (XI) and the income
the presentationsome notations are definedbelow: ratio (X2) are jointly used as two screening variables to
predictthe default rate.
N: total numberof customers Let Y denote an unobservableperformancevariablethat
r: averagenet revenueper non-defaultingperson describesthe likelihood of a customer'sdefaulting.Suppose
d: averageloss per defaultingperson thereis a cutoff limit k., a customerwith Y> k, will default,
f/: acceptancerate while one with
Y<<k,
will not default.Furthermore,assume
q: overall defaultingrate after screening that (Y, X1, X2) have a standard trivariate normal
distributionwith the covariance(or correlation)matrix
The total profit is equal to total net revenue minus the
total loss due to defaulting,namely, 1 -0.6 -0.4
S- -0.6 1 0.1
7E= [r(l - q) - dq]/N = [r- (r + d)]lPN (13) -0.4 0.1 1
Note in (13) that maximizingtotal profit is equivalent to where Poi -0.6, P02 --0.4, and Pl2 = 0.1. This matrix is
minimizingdefault rate only when # is prefixed. For the obtained from the two-dimensionalcross table in O'Conner
individualcut approach,(13) is in a nonlinearform such that exampleusing the estimationmethod describedfor (2).
a numericalmethod is requiredto solve for optimal cutoff From the company'sexperiencethe overalldefault rate i
points (kl, k2, ...) by maximizing the total profit. The is assumedto be 20%, which means that 80% of customers
subroutineE04JAF in Numerical Analysis Group (NAG) are good. That is, when p[Y>k,] = 0.2, we can obtain k,
Library can be applied directly. If / is prespecified, it
=-1(1-0.2)= 0.84162. The company plans to select the
becomes a one parameter problem. For the linear cut 40% best customers least likely to default and give them
approachwithout any constraint,an optimalcutoff point c* entireadvance,otherwiserejectingthem.Now, four different
can be obtainedin a closed form by maximizing(13) namely screeningproceduresare compared for a fixed acceptance
rate / = 0.4 as shown below.
(a) Single screeningvariableX1 (the creditscore). From
r+ (14) acceptanceratep[X1 > k]= 0.4, we can determinethe cutoff
This content downloaded from 141.211.4.224 on Sat, 09 Jan 2016 05:32:42 UTC
All use subject to JSTOR Terms and Conditions
840 Journal
oftheOperational
Research Vol.56, No.7
Society
This content downloaded from 141.211.4.224 on Sat, 09 Jan 2016 05:32:42 UTC
All use subject to JSTOR Terms and Conditions
model
Tsaiet al-Aneconomic
H-T problem841
assessment
forcredit
example, the optimal total profit is $3 100600 with position among the four screeningprocedures.Clearly, the
Pol=-0.6 while it is $1717 000 with 02= -0.4. For the linear cut approachwith two screeningvariablesuniformly
individual cut approach with two screeningvariables, the outperformsthe other three proceduresin terms of default
optimal total profitis equal to $3 314900, which is in second rate and total profit, and it has less computational
complexitythan the individualcut approach.
Table 2 Screening effects for various correlation P12 given po = -0.6 and p02 =-0.4
(1) X1-X2 individualcut (2) X1-X2 linear cut
This content downloaded from 141.211.4.224 on Sat, 09 Jan 2016 05:32:42 UTC
All use subject to JSTOR Terms and Conditions
842 Journal
oftheOperational
Research Vol.56, No.7
Society
also showed that the total profit increasesas the equivalent The impactof correlationbetweentwo screeningvariables
correlationincreasesfor the linear cut approach,while it is is also examinedin Table 2. When the correlationis at either
not true for the individualcut approach.For example,when end of the constraint interval, the total profit reaches the
V/ = 1 the linearcut approachcan exactlyidentifyall good largestvalue for the linearcut approach,while it is relatively
customers(80%)with zero defaultrate and consequentlythe small for the individual cut approach. Therefore, the
total profit is equal to $8 000 000. The individual cut screeningprocedureusing the linearcut approachis strongly
approach has two cases of /i = 1. When P12=0.9732, it recommendedfor credit assessmentproblems.
can only select 54.49%of customerswith 0.0718 defaultrate A commercial scorecard is normally constructed using
and total profit of $3 101000; when P12= -0.4932, it can severalvariablesand is costly, and here we do not intend to
only select 51.19%of customerswith 0.0231 defaultrate and discussits development,revision, or maintenance.However,
total profit of $4 410 600. Clearly, for large equivalence when more useful predictivevariables are considered,and
correlations,the individualcut approachcannot improvethe the originaldata of all variablesin the existing scorecardare
screeningeffect as much as the linearcut approachcan. available, then the linear cut approach could serve as one
Furthermore,as the equivalent correlation is 0.6 when possible method to update the existing scorecard.For the
P12 = 0.6667 (Pol/PO2 or P02/P01), it means that two screening mailing credit assessment problem, the individual cut
variableshave the same screeningeffect as only one of the approach was adopted by Tsai and Yeh14 instead of the
two screeningvariables;the two approacheshave exactlythe matrix approachused in Bennett et al.7 In future research,
same screeningeffects, since both have the total profit of the linearcut approachcould be applied and be expectedto
$3 101000. lead to significantimprovement.
This content downloaded from 141.211.4.224 on Sat, 09 Jan 2016 05:32:42 UTC
All use subject to JSTOR Terms and Conditions
model
Tsaiet a/-Aneconomic
H-T forcredit problem843
assessment
13 MoskowitzH, Plante R and Tsai HT (1993). Multistage Using Equation (Al), the above equation can be writtenas
screeningmodelfor evaluationand controlof misclassifica-
tion error in the detection of hypertension. Mngt Sci 39:
307-321. NNr(D ) - N(r + d)
14 Tsai HT and Yeh HC (1999).A two-stagescreeningproce-
dure for mailing credit assessment.IMA J Math Appl Bus Ind c/,/5
10: 1-15. +
x -k( (A3)
15 Long JS (1997). RegressionModelsfor Categoricaland Limited )dx
-00 D/'(x) -vx
DependentVariables.Sage Publications:London.
16 GraybillFA (1976). TheoryandApplicationof the LinearModel. The first partialderivativeof it with respectto c is
DuxburyPress:Massachusetts.
17 Owen DB (1980). A table of normal integrals. CommunStat- + d),
- N (r
07r
SimulC 9: 389-419. N rDc
=N
-k+
(A4)
Appendix: Proof of (14) (~-)
Referringto Owen,17 one of the resultsis settingOci/c equal to zero. Consequently,the optimalcutoff
point of acceptanceregion is given by
I(x)( dx = BvN(h, k;p) (Al)
f px + v-V- (A5)
c* k' (r +d)
where D(.) denotes the standard normal distribution The second derivativecan be shown that
function, and BvN denotes the standard bivariate normal
distributionfunction. 02' cE-k7 '( ' + c
<0
(r +d)N (A6)
Rearrange(13), the total profit is C2-
-
This content downloaded from 141.211.4.224 on Sat, 09 Jan 2016 05:32:42 UTC
All use subject to JSTOR Terms and Conditions