American Finance Association

Credit Granting: A Comparative Analysis of Classification Procedures

Author(s): Venkat Srinivasan and Yong H. Kim
Source: The Journal of Finance, Vol. 42, No. 3, Papers and Proceedings of the Forty-Fifth
Annual Meeting of the American Finance Association, New Orleans, Louisiana, December 28-
30, 1986 (Jul., 1987), pp. 665-681
Published by: Wiley for the American Finance Association
Stable URL: http://www.jstor.org/stable/2328378 .
Accessed: 27/11/2013 14:34

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 [email protected].

Wiley and American Finance Association are collaborating with JSTOR to digitize, preserve and extend access
to The Journal of Finance.

http://www.jstor.org

This content downloaded from 35.8.11.2 on Wed, 27 Nov 2013 14:34:11 PM

All use subject to JSTOR Terms and Conditions
THE JOURNALOF FINANCE * VOL. XLII, NO. 3 * JULY 1987

Credit Granting: A Comparative Analysis of

Classification Procedures
VENKAT SRINIVASAN and YONG H. KIM*

ABSTRACT
Financialclassificationissues, and particularlythe financial distress problem,continue
to be subject to vigorousinvestigation.The corporatecredit grantingprocess has not
receivedas much attention in the literature.This paperexaminesthe relativeeffective-
ness of parametric, nonparametricand judgemental classification procedureson a
sample of corporate credit data. The judgemental model is based on the Analytic
Hierarchy Process. Evidence indicates that (nonparametric)recursive partitioning
methods provide greater informationthan simultaneouspartitioningprocedures.The
judgementalmodel is found to performas well as statistical models. A complementary
relationship is proposed between the statistical and the judgemental models as an
effectiveparadigmfor grantingcredit.

THE CREDITGRANTINGprocess involves a tradeoffbetween the perceiveddefault

risk of the credit applicant and potential returns from granting requestedcredit.
The main objective in credit granting is to determine the optimal amount of
credit to grant. The amount of credit requested along with other financial and
nonfinancial factors influences the assessment of default risk. While default risk
assessment can be appropriatelymodeledusing classificatorymodels, integration
of such risk assessment with potential returns can be accomplished using a
dynamic expected value framework.'
Academicresearchon credit grantingcan be groupedinto two basic categories:2
(i) attempts to apply classification proceduresto customer attribute data and
develop classification models to assign group membershipin the future;and (ii)
attempts to explicitly recognizethe need to integrate such risk assessment with
potential return as well as allow for differing degrees of credit investigation
depending on the costs of such investigation. To our knowledge, this study
representsthe first attempt at examining the relative performanceof classifica-
tion proceduresfor corporatecredit granting. Relative performanceis measured
in terms of how well the models replicate expert judgement. We examine four
statistical models: multiple discriminant analysis (MDA), logistic regression
* NortheasternUniversity and University of Cincinnati, respectively.We
are gratefulto Robert
Eisenbeis and Sangit Chatterjeefor their insightfulcommentson an earlierdraft. We also thank an
anonymousFortune 500 corporationas well as the Credit Research Foundationfor their support.
The usual disclaimerapplies.
1This paperis primarilyconcernedwith the creditgrantingdecisionin an industrial(nonfinancial)
setting. A more detailedversion of this paper is availablefrom the authors.Further,for a reviewof
the variousmotives for corporationsto extend credit, referEmery [10].
2 We do not make any attempts to present a comprehensivereview of the literature.Interested

readersare referredto an excellent reviewby Altman et al. [1].

665

This content downloaded from 35.8.11.2 on Wed, 27 Nov 2013 14:34:11 PM

All use subject to JSTOR Terms and Conditions
666 The Journal of Finance
(logit), goal programming(GP) and recursivepartitioning algorithm (RPA), and
a judgementalmodel based on the Analytic HierarchyProcess (AHP). The data
employedfor the study has been providedby an anonymous Fortune 500 corpo-
ration.
The remainder of this paper is organized as follows. Section I presents the
credit grantingproblem formally.We present the four classificatory models and
the judgementalmodel in Section II. The data and methodologyare summarized
in Section III. Section IV contains a discussion of the results as they pertain to
classificatoryability and the determinationof credit limits. The paper concludes
with a summary.

I. The Corporate Credit Granting Problem

The prescriptionsof normative theory in credit granting (e.g., Altman et al. [1])
can be summarized as follows: (i) credit granting is a multiperiod problem,
implying that granting credit may not only enable the firm to make the current
sale but also to sell in the future to the same customer; (ii) the extent of credit
investigation must be determinedby the tradeoff between the incremental costs
of investigation and the amount of credit involved; (iii) estimate the present
value of benefits and losses from grantingcredit for all the periodsin the planning
horizon,given the customer'saction in the immediatelypriorperiod;(iv) integrate
probabilities of collection and default with benefits and losses each period to
compute the net present value from granting credit; and (v) grant credit, if
computednet present value is positive and reject full credit, otherwise.3
The problem of setting credit limits can be formulatedas a dynamic program
(Bierman and Hausman [2]; Srinivasan and Kim [24]). The dynamic program
considers the influence of past collection experience on default probabilitiesin
the future. It is, however, very difficult (if not impossible) to estimate future
benefits and losses from granting credit other than at a qualitative level. Elimi-
nation of future benefits and losses from the dynamic program reduces it to a
single-period formulation which is, of course, computationally more tractable.
The recurrencerelationshipin the single-periodformulationcan be stated as:
fi(Pli) = Max[(Plik) + (P2ik2), 0] (1)
where
fi(Pii) = maximum discounted expected payoffs from state i to 0, given the
current state variable (Pli) and that optimal decisions are made from
now on (the return function);
Pli = probability of payment in period i, given customer action in period
i- 1;
P2i= probability of default in period i, given customer action in period
i - 1;
k, = present value of immediate benefits in the event of collection; i.e.,
discounted profits from sale or s a-Cl - v a c2;
3Alternatively, the maximumcredit limit to be grantedwill be determinedby equating NPV to
zero.

This content downloaded from 35.8.11.2 on Wed, 27 Nov 2013 14:34:11 PM

All use subject to JSTOR Terms and Conditions
 CreditGranting 667
k2= present value of loss in the event of default or -v a
c 1 = averageselling cycle;
c2 = averagepayables cycle;
a = 1/(1 + d) = appropriate discount factor;
d = appropriatebefore-tax discount rate per period;
s = sales; and
v = variable costs.

In the context of a single period, expression (1) is simply the E(NPV) from
grantingcredit in the first period. Assuming for convenience that variable costs
(v) can be expressed as a proportionof sales (s), formulation (1) can be restated
as follows to determinethe maximumcredit limit:

Plik- P2iCLmax(v/s) = 0 (2)

where, CLmaxis the maximumcredit line that should be grantedto the customer.
The above descriptionimplies that the credit grantingprocess consists of two
stages: (i) estimation of default probabilities; and (ii) integration of default
probabilitiesin (1) or (2) to estimate credit limits. Two basic approacheshave
evolved to facilitate the assessment of default risk: statistical and judgemental
systems. All credit analysis, whether based on judgementalor statistical systems,
operate on similar principles. Further, in estimating default probabilities, the
suggestedmultistage investigation process essentially determinesthe feasible set
of measurements,X, that can be used to estimate the probabilities,Pli and P2i.

II. Classificatory Models for Credit Evaluation

A. StatisticalModels
The two parametric methodologies we employ, MDA and logit, have been
extensively investigated in prior classification studies. Interested readers are
referredto Altman et al. (1981) for a detailed discussion. It will suffice to mention
here that the default probabilitiesfor each observationusing MDA and logit will
be the posteriorprobabilitiesof groupmembershipgeneratedfrom the respective
MDA and logit models. These probabilitiescan then be substituted in (1) or (2)
to determinecredit limits.
Both MDA and logit necessitate restrictiveassumptionsof distributionalform,
beingparametricin nature.In particular,the problemsin using MDA for financial
classificationare well documented(see, e.g., Altman et al. [1]). Several nonpara-
metric classification alternatives have been suggested in the literature. Among
these are the rank transformeddiscriminant analysis (Conover and Iman [6]),
the recursive partitioning algorithm (RPA) (Breiman et al. [2]), the Quinlan
algorithm (Quinlan [17]), and goal programming(GP) (Freed and Glover [12]).
We focus on the GP and the RPA in this study.
MDA, logit and GP can be characterizedas simultaneous partitioning proce-
dures since they consider all the potential discriminatoryvariables simultane-
ously in the classification model. However, models like the RPA can be charac-

This content downloaded from 35.8.11.2 on Wed, 27 Nov 2013 14:34:11 PM

All use subject to JSTOR Terms and Conditions
668 The Journal of Finance
terized as sequential partitioning proceduresthat are combinatorialin nature.4
There is a fundamental reason why repeated partitioning procedures like the
RPA may yield greater informationthan simultaneouspartitioning procedures.5
Simultaneous partitioning procedures implicitly assume convexity of group
spaces in the p-variate measurementspace. This may be unrealisticin the context
of many financial classification issues, particularlyin multigroupclassification
problemswhere there is likely to be considerablenon-convexity in groupspaces.
Further,it is feasible that the underlyingdecision process may be a conditional
and sequential one instead of being a simultaneous process. The data may also
have a bimodal or multimodal distribution. In such cases, greater information
and hence better classification accuracy,may be obtained by examining charac-
teristics in an incremental manner using procedureslike the RPA.

1. MathematicalProgramming(MP)
Recently, the interest in MP-based alternatives to parametric classification
proceduresseems to have been revived by Freed and Glover [12]. Formally,in a
MP context, the discriminant problem can be stated as one of finding a linear
transformationX and boundary value b to categorize group k, given points Ai
and groupsGk. Thus, for a 2-groupproblem,the objective is to find X such that:
AiXs b AiE G1
AiX > b Ai E G2
where b is any positive constant. Since, however,the Ai's may be distributedin
a manner that may not permit complete discrimination, we need to relax the
constraints to ensure feasible solutions. The slack (surplus) variable for the
purpose,a, can take two forms: (i) it can be constrained at the grouplevel, ak; or
(ii) it can be constrained at the observationlevel, ai. Constrainingthe a's at the
grouplevel yields LPI:
Min E2=1 ak

such that
AiX c b + ak i =(1,... Q), i E Gl, k = (1, 2).
AiX > b - ak i= (Q + 1, ... ., R), i E G2, k = (1, 2).
where the X's are unrestricted in sign. Constrainingthe a's at the observation
level, on the other hand, yields LPII:
Min E2=1 ak

such that
AiX <b +ai i =(l, .
Q), i E Gl, k =(1, 2).

4A numberof sequentialpartitioningalgorithmshave evolvedin the patternrecognitionliterature.

For a reviewof some of these procedures,referDiettrich and Michalski [7].
'For details, see Chatterjeeand Srinivasan[5].

This content downloaded from 35.8.11.2 on Wed, 27 Nov 2013 14:34:11 PM

All use subject to JSTOR Terms and Conditions
 CreditGranting 669
where the X's, as before, are unrestricted.6Typically, the b values have been set
at 1 or 10. Estimation of the probability of group membership in the GP
formulations can be approximatedby the relative frequencies of the groups in
each of the two halfplanes. These probabilities can, as before, be substituted in
formulation(1) or (2) to obtain credit limits.

2. RecursivePartitioning

RPA yields a binary classification tree, similar to a spanning tree, that enables
the assignment of objects into one or k known groups.Binary classification trees
are constructed by repeated splits of X, where X is the measurement space
containing all relevant measurement vectors. The construction of the binary
classification tree revolves aroundthe following three elements:
1. the selection of splitting rules;
2. the decision when to declare a node as terminal;and
3. the assignment of each terminal node to a group.

The performanceof the model critically depends on the first two elements. It
turns out that the third element is trivially resolved, given solutions to the first
two elements. The fundamental notion behind each split of a subset is one
obtainingdescendant subsets that are 'purer'than the data in the present subset.
The best splitting rule for the given sample is defined as the one which
maximizes the decrease in the sum of the impurities of the two resulting
subsamples comparedwith the impurity of the parent sample. In order to find
the best splitting rule, the algorithmfirst searches for the best splitting point for
each explanatory variable and then the best of these splits is selected. The
splitting process terminates when further splitting does not lead to any decrease
in the impurityof the currenttree. The currenttree then is referredto as Tfin.
The next step in the process is to search for a tree that is less complex than
Tfir but has a smaller cross-validated classification error rate. This step is
motivated by the observationthat Tfin usually overfit the data and are complex.
The final optimal tree is selected based on a criterion of minimizing cross-
validated resubstitution risk. The cross-validation risk of tree T is computed
using a V-foldcross-validationprocess, where V is the numberof folds. All cases
in the sample are randomlydivided into V groups of approximatelyequal sizes.
Observationsin V - 1 groupsare used to construct the tree correspondingto the
penalty chosen from the range of values of penalty parametersfor which tree T
was optimal, based on all of the observations.The observationsin the groupleft
out are classified by the newly constructed tree. The procedure is repeated V
times, each time with a different group being left out. The resubstitution risks
from all cross-validationtests for a particularcandidate T are averagedto obtain
cross-validatedrisk for T.

6In some exceptional cases, LPI and LPII can be shown to be degenerate (Markowskiand
Markowski[15]). However,simple normalizationscan be undertakento remedythe problem (see,
Freedand Glover[11]). In the data under study, we did not have to resortto any normalizations.

This content downloaded from 35.8.11.2 on Wed, 27 Nov 2013 14:34:11 PM

All use subject to JSTOR Terms and Conditions
670 The Journal of Finance
B. A JudgementalModel Based on the AHP7
AHP is a multiattribute modeling (MAD) approach that has substantial
intuitive appeal and is theoretically rigorous.Like other MAD approaches,AHP
also attempts to resolve conflicts and analyze judgements through a process of
determiningthe relative importanceof a set of activities, players or criteria.8The
AHP starts by decomposingthe principal problem into a hierarchy.Each level
consists of a set of elements and each element, in turn, is broken into sub-
elements for the next level of the hierarchy.The final level consists of the specific
coursesof actions that are being evaluatedfor adoption.Within each hierarchical
level, priorities are established for each of the elements using a measurement
methodology.This methodologicalprocess constitutes the core of the AHP and
determinesthe scope of data collection and analysis.
The basic assumption underlying the measurement methodology in AHP is
that relative dominance can be measuredby pairwise comparisons.Assume that
we wish to conduct pairwise comparisons of a set of n attributes and establish
their relative weights. If we denote the attributes by 01, 02, ..., On and their
weights by wi, w2, . . ., wn, the pairwise comparisonmatrix 0 may be expressed
as the reciprocalmatrix shown below:
01 02 ... n
01 Wi/Wi W1/W2 ..
WllWn
02 W2/W1 W2/W2 ... W2Wn
0=.

On wnlWl WnfW2 ... Wn/Wn

Note that if we post-multiply 0 by the column vector iv, we obtain the vector
nw. To recoverthe weights, w, we can solve the system (O - nI) = 0, which must
have a non-trivial solution, since 0 has unit rank and all the eigenvalues of 0,
(L = 1, 2, ..., n), are zero except one (Lma.). Further, the trace (0) = sum of
the eigenvalues = sum of the diagonal elements of 0 = n.
The matrix 0 satisfies the 'cardinal'consistency property 0ij?>k = Oik. Thus,
given any row of A, we can determine the rest of the entries from this relation.
However,suppose we have estimates of the ratios in the matrix. In this case, the
cardinal consistency relation above need not hold, nor need an 'ordinal'transi-
tivity relation of the form:Oi> Oj, Oj> ?k imply Oi > Ok hold (wherethe Oiare
the rows of 0). It can be shown that in a positive reciprocal matrix, small
perturbations in the coefficients imply small perturbations in the eigenvalues
and hence the eigenvector is insensitive to small changes in judgement [18, p.
192].
'This model is more fully describedin Srinivasanand Kim [23]. Further,note that even though
we have includedthe descriptionof the AHP-basedmodel underclassificationmethods,it is strictly
not a classificationmodel. The model is, however,directlyusable to determinecredit limits.
8AHP has been applied to many varied problems including planning for a national waterway
(Saaty and Vargas [20]) and marketingdecisions (Wind and Saaty [27]). Srinivasanand Kim [22]
illustratethe applicabilityof AHP to manyfinancialdecisions.For an extensive listing of applications
of AHP, see Zahedi [28]. For theoreticaldetails, refer Saaty [18]. For an axiomatictreatmentof the
process,refer Saaty [19].

This content downloaded from 35.8.11.2 on Wed, 27 Nov 2013 14:34:11 PM

All use subject to JSTOR Terms and Conditions
 CreditGranting 671
In general, the weights are estimated from data or from experienceddecision
makers or a group of experts. A 9-point measurement scale is used, where 1
stands for equal importanceof element i over elementj and 9 stands for absolute
importanceof element i over element j. The principaleigenvectorof this matrix
is then derived and weighted by the priority of the element in the higher level
with respect to which the evaluation has been made. Continuingthis process of
eigenvectorextraction and prioritizationby weighting leads to a unidimensional
priority scale for the elements in each of the hierarchicallevels.
The next step in the process is to obtain a measure of the consistency of the
judgementaldata. It can be shown that Lmax? n for all possible states and that
(Lmax- n)/(n - 1) serves as an index measure of consistency of the comparison
data (Saaty [18]). The index (C.I.) indicates the departurefrom consistency of
the comparisonratios, w/lw1, and the ratios are deemed consistent if Lmax= n.
The consistency index is comparedto a level of consistency that can be obtained
merely by chance. Saaty and Mariano (21] have established for different order
random entry reciprocal matrices, an average consistency index which ranges
from 0 for 1 to 2 element matrices to 0.9 for 4 element matrices and to 1.49 for
10 element matrices. In general, a consistency ratio of 10%or less is considered
very good. If consistency is poor, additionaldata or another roundof comparisons
may be required.
In the context of the credit granting problem, the final level in the hierarchy
will consist of two elements whose relative weights can be interpreted as the
subjective probabilities of granting and rejecting full requested credit. These
probabilitiescan again be substituted in (1) or (2) to derive the optimal credit
limits.

III. Data and Methodology

A. Data
The actual credit granting process in the participatingcorporation (hereafter
PC) is reported in detail in Srinivasan and Kim [23]. We only describe the
relevant facts here. The process can be defined as a multistageprocess where the
stages have been established as a function of the size of the credit request.There
are four stages in the process:
Stage I: Less than $5,000
Stage II: $5,001-$20,000
Stage III: $20,001-$50,000
Stage IV: Greaterthan $50,000
The extent of investigation varies across the stages with stage I undergoingvery
little or no investigation and stage IV undergoingextensive investigation. Even
extensive investigation in many cases is limited by the amount of financial
informationthat can be obtained.Many customersareprivatelyheld corporations
and the firm's main source for financial information is the Dun and Bradstreet
Corporation(DBC). DBC, however,only providesbalance sheet information,net
profits and sales for the year. Income statement details are not provided.

This content downloaded from 35.8.11.2 on Wed, 27 Nov 2013 14:34:11 PM

All use subject to JSTOR Terms and Conditions
672 The Journal of Finance
Completefinancial evaluation is done (with available information)in addition
to nonfinancialevaluation (e.g., trade and bank references,pay record,customer
background) only in the case of stage IV customers. Analysis of stage IV
customers is supervised by senior level credit management staff and all credit
decisions relatingto these customersare made by such staff. For purposesof this
study, an appropriatesenior level credit managerwas chosen as the 'expert'.9If
a stage IV customer is perceivedto be 'poor'risk, the customer is classified as a
'high risk' (HR) and one of several actions is taken:
a. Future credit shipments are stopped effectively eliminating the customer's
credit line;
b. Future credit shipments are only undertaken if proper collateral is fur-
nished;
c. All future shipments are stopped. Appropriatelegal action is initiated. This
is often the course of action for customers who file for bankruptcy.
For the purpose of this study, the procedureyields two groups of interest: (i)
customerswhere the firm has taken restrictive action on their credit lines (note
that these customers need not have filed for bankruptcy);and (ii) customers
where the firm has not taken restrictive action on credit lines (a very small
portion of this group are customers who at some prior period belonged to group
(i)).
Data were collected on the two groups of interest as of July 1986.10 Several
conditions were imposed on the data:
1. Only credit lines in excess of $50,000 and not previously analyzed in
Srinivasan and Kim [23] were considered;
2. Complete financial information (DBC) was available for up to within 7
months of July 1986;
3. Restrictive action on credit lines was significant."
The above process yielded a total of 215 customerfiles out of which 39 customers
were HR customers and the remaining 176 customers were non-HR customers.
9Note that the study assesses the ability of statistical and judgementalmodelsto replicateexpert
judgement.In this respect, it is similar to prior studies that have attemptedto predict investment
qualityratings (e.g., Srinivasanet al., [25]) and also to the loan classificationstudy by Maraiset al.
[14]) that attempts to replicate loan officer judgements.Our objective could be arguedto raise a
potentialconcernof determiningwhetheran erroris due to the models or an erroron the part of the
creditmanager.However,we feel that the objectiveis justified underthe circumstancesprevalentin
the PC. The PC is convincedthat its existing system is best suited for its business conditions.It has
experiencedvery negligiblelosses. The interest in this study stems from the desire on the part of the
PC to identifyand capturethe underlyingexpertjudgementprocess so that the entire processcan be
replicatedusing artificialintelligence.Therefore,the appropriateobjectivefor this study is to attempt
to replicatethe expert'sjudgement.Further,note that the classification of interest in the study is
definedby restrictivecredit action initiated by the expert.
10Theoretically,it can be arguedthat the study ignores the populationof customerswho were
denied credit in the first place. However,this argumentis not valid in the PC's case as far as stage
IV is concerned.In the PC's case, customersnormallystart in stage I and graduallymove through
the upper stages over time. The PC very rarelyhas receivedfirst-time credit requeststhat fall into
stage IV. Further,the PC has never denied initial creditto the few first-timestage IV customers.
" There were a few cases (=5) where the firm continued to sell on credit because of marketing
needs even thoughthe customerswere classifiedas HR.

This content downloaded from 35.8.11.2 on Wed, 27 Nov 2013 14:34:11 PM

All use subject to JSTOR Terms and Conditions
 CreditGranting 673
For each customer, available information was used to compute the following
ratios:
a. currentratio (CR);
b. quick ratio (QR);
c. net worth to total debt (NWTD);
d. logarithmof total assets (LOGTA);
e. net income to sales (NIS);
f. net income to total assets (NITA)
In addition to the above financial variables,two nonfinancial variableswere also
included:(i) pay recordwhich is indicativeof the firm'spast collection experience;
and (ii) customer backgroundproxied by the number of years the customer has
been in business. Pay record was treated as a categorical variable with three
possible values (Mehta [16]):
Good: when payments have been consistently made within the prescribed
period of one month;
Fair: when the formal credit period has been frequentlyexceeded, but there
has been no need for stronger than formal payment reminders;or the
customer has given a satisfactory explanation for delinquency;
Poor: when strongerand personal remindershave been necessary in the past,
and credit has been at times exceeded beyond a three-month period
without reasonableor satisfactoryexplanations.
Similarly, customer backgroundwas also treated as a categorical variable with
three values: (i) less than 2 years; (ii) 2 to 5 years;and (iii) greaterthan 5 years.
Exploratorydata analysis was conductedby reviewingthe sampledata behavior
of each variable.Obviously,the two categoricalvariablesare nonnormal.All the
six financial variables also violated univariate normality. The financial ratios
could be transformedto approximateunivariate normality. The data even after
transformations is not likely to be multivariate normal due to the categorical
variables. Moreover, transformations may change the relationships among the
variables and may also affect the relative position of observations. Further,
exploratoryMDA models based on transformeddata did not yield significantly
better results. The remainderof the analysis was, therefore,done only based on
the untransformeddata.

B. Methodology
The real test of the classification models in the context of this study is their
ability to replicatethe expert'sjudgements.Maraiset al. [14] identify three types
of overfittingbias that can result if expected misclassificationlosses are estimated
from the same sample used for model estimation: (i) overfitting in the choice of
explanatory variables; (ii) sensitivity to a particular loss function; and (iii)
statistical overfitting bias in the computation of the expected loss rate.
The first type of bias is clearly not relevant for this study as we do not use any
data reduction procedure. Further, loss functions are typically very hard to
estimate as also admitted by Marais et al. [14]. A variety of methods have been
proposedto deal with the statistical overfitting bias. In most prior classification

This content downloaded from 35.8.11.2 on Wed, 27 Nov 2013 14:34:11 PM

All use subject to JSTOR Terms and Conditions
674 The Journal of Finance
studies, some form of holdout samples has been used. Alternatively, some form
of statistical resampling schemes may be used, e.g., cross-validation,bootstrap-
ping. We use the bootstrap for the MDA, logit and GP. Cross-validationis used
for RPA.
Bootstrappingis a computerintensive technique that is especially appropriate
when the data violate the distributionalassumptions of the statistical procedure
employed. Efron [81]finds various bootstrap estimators to dominate cross-vali-
dation or jackknifing, particularly in small samples. Chatterjee and Chatterjee
[4] suggest a superiorprocedurewhere the model developed from the bootstrap
sample is used to classify observationsnot includedin the bootstrapsample. The
bootstrap procedure adopted in this study can be described in terms of the
followingthree steps:
1. From the sample being studied, S, with a total of Ns observations,draw an
independentsample of size n, each observationbeing drawnwithout replace-
ment with a probability of 1/(Ns - Ki), where Ki is the total number of
observations already selected as of the ith selection, i = {1, 2, .. ., n8j. The
sample thus drawn constitutes the bootstrap sample and is used for esti-
mating the discriminant function.
2. The resubstitution loss rate of the discriminant function is assessed by
applying it to classify all the observations not included in the bootstrap
sample.
3. The above procedure is repeated 25 times. Each replication provides an
estimate of the misclassification probability for both the bootstrap and
validation samples. The average of all the replications is taken as the
bootstrapand validation estimates. The standarddeviation of the estimates
providesan estimate of the standarderrorof misclassificationprobabilities.
Further, for the purposes of this study, n, was determined by drawing from a
uniform distribution such that 50% of the sample in each of the two groups
constituted the bootstrap sample.

IV. Results
A. Classification
The cross validated results of a linear MDA model using proportionalprior
probabilitiesare presentedin Exhibit 1. On the average,the linear MDA correctly
classifies 88.89% of all the customers in the bootstrap sample. The average
classification accuracydrops to 85.05%in the case of the holdout sample. There
is some variation in the classification accuracywithin the two groups:90.9%of
the non-HR customers were classified correctly and only 80% of the HR cus-
tomers were correctly classified. Cross validated classification accuracyis lower
but not appreciably.The overall resubstitution loss rate is about 3.84% but is
more severe in the HR cases.
The homogeneityof dispersionmatriceswas tested using a likelihoodratio test
[16] which is approximatelychi-squaredistributed.The results indicatedthat all
the 25 bootstrapsamples did not meet the homogeneityassumption.A quadratic

This content downloaded from 35.8.11.2 on Wed, 27 Nov 2013 14:34:11 PM

All use subject to JSTOR Terms and Conditions
 CreditGranting 675

Exhibit1
Summary of Classification Results (% Correctly Classified)

LMDA QMDA Logit LP I LP II RPA 1 RPA 2

i. Bootstrap Sample

Non HR 90.90 92.05 93.18 89.77 90.90 95.45 94.32

(1.66) (1.29) (1.73) (1.40) (1.51) (0.99)

HR 80.00 85.00 90.00 80.00 80.00 89.74 90.00

(2.48) (2.32) (4.21) (2.83) (2.45) (0.00)

Total 88.89 90.74 92.59 87.96 88.39 94.44 93.52

(1.66) (1.37) (2.11) (1.59) (1.67) (0.80)

ii. Validation Sample

Non HR 87.50 88.64 89.77 86.36 88.64 93.18 93.18

(1.01) (1.52) (1.22) (1.46) (1.27) (0.34)

HR 73.68 78.95 78.95 73.68 78.95 89.74 89.74

(2.53) (2.46) (2.36) (3.36) (2.58) (2.48)

Total 85.05 86.92 87.85 84.11 86.92 92.56 92.29

(1.08) (1.56) (1.41) (1.64) (1.52) (0.69)

1. All figures are averages over the 25 replications. Figures in parentheses represent the estimated standard error
of percent correctly classified. Details of replication results are available from the authors.

2. Results for the RPA 1 are from using 10-fold cross-validation. RPA models were based on the symmetric Gini
criterion and no linear combinations were used.

discriminant rule was, therefore, applied and the results are summarized in
Exhibit 1. The classification accuracyhas improvedin the case of the bootstrap
sample to 90.74% compared to 88.89% using the linear rule. However, the
resubstitutionloss rates are slightly higher.This may be a result of the overfitting
that accompaniesa quadraticrule. Replication results do not reveal any signifi-
cant variation over the bootstraptrials.'2
Relative influence of variables was ascertainedusing a forwardstepwise selec-
12 Note than even though the data do not satisfy some assumptions of the MDA models, the

bootstrapprocedurecan be arguedto, at least partially,offset this bias.

This content downloaded from 35.8.11.2 on Wed, 27 Nov 2013 14:34:11 PM

All use subject to JSTOR Terms and Conditions
676 The Journal of Finance
tion rule. We found that the QR, NITA and NWTD were consistently significant
in all the 25 samples. Variables that were selected less frequentlywere CR and
PAY. Remainingvariables were not found significant univariate discriminators
in the stepwise selection rule used.
Results using the unorderedlogit model are also presentedin Exhibit 1. Overall
as well as within group classification results are better than the linear MDA
models and comparableto the quadratic MDA. Note the relatively significant
variation in classification results for HR customers in the bootstrap sample.
Analysis of the replication results does not, however, reveal any particular
replicationto have caused the high variation. Logit model results reveal that CR,
QR, NITA and PAY were consistently significant at the 0.05 level in all of the
25 replications.
LPI, which allows deviation only at the group level, yields an overall classifi-
cation accuracy of 87.96% and 84.11% for the bootstrap and holdout samples,
respectively(Exhibit 1). The results are comparableto the linear MDA but not
as goodas the results using logit. LPII does not appreciablyimproveclassification
results implying that the data are relatively well behaved and allowing the
deviationvariableat the observationlevel did not have a great influence. Overall
classificationaccuraciesusing LPII are 88.89%and 86.92%for the bootstrapand
holdout samples, respectively. Variances from both LPI and LPII models are
comparableto other models.
In the case of the RPA, overall classification accuracy for the estimation
samplewas 94.44%(Exhibit 1). The cross-validatedaccuracyusing 10-foldcross-
validationwas 92.52%.These results are slightly superiorto the results obtained
using logit. Sensitivity analysis was conducted to determine if the results were
dependent on the number of folds used in the cross-validation process. Folds
werevariedfrom 1 to 10 and the results are presentedin Exhibit 1. As is apparent,
the results using RPA are not very sensitive to changes in the number of folds
(1 to 10). We present the optimal cross-validationtree, based on 10-fold cross-
validation,in Exhibit 2. The tree has 8 terminal nodes and the node membership
is also indicated for each node. The optimal RPA tree considers the variables
NITA, CR, PAY, NWTD, and LOGTA. Each node has either been assigned to
the HR or the NHR category based on node membership.Note further that the
tree in Exhibit 2 is a prunedtree reflecting the tradeoffbetween tree complexity
and resubstitutionloss. The complex tree, Tfin,will classify all the observations
but will, of course, have less generalizability.
In comparison,RPA appears to be slightly superior as a classification model
to replicate the expert's judgements in this study. The classification results
obtained using RPA are better than logit, MDA and GP. Besides, they appear
robustto changes in the cross-validationscheme used.

B. CreditLimits13
We conducteda judgementalinput elicitation process for stage IV customers.
All judgementswere providedby the senior credit manager('expert') who is also
responsible for classifying stage IV customers as HR customers. After many
13
We present the AHP model first and subsequentlyprovide a comparativeassessment of the
methodologiesin determiningcredit limits.

This content downloaded from 35.8.11.2 on Wed, 27 Nov 2013 14:34:11 PM

All use subject to JSTOR Terms and Conditions
 CreditGranting 677
Exhibit2
Optimal RPATree

NITAs0.03 NTA> 0.03

NWDT- 0.8 NWDT> 0.8 CR s 1.26 CR > 1.26

HR NWTD 0.54 NWTD>0.54

16
3
NHR PAY 2 PAY> 2
18
2
NHR
60
0
NHR
34
0 PAY- 2 PAY> 2

LOGTAs 2.23 LOGTA> 2.23

HR NHR HR NHR
6 22 13 30
3 2 6 0

The numbersbeloweach node indicatenode membership.The top numberis the frequencyof the groupto which
the node has been assigned. The tree is based on a 10-foldcross-validation.

PAY:
HR = High Risk Good = 2
NHR = Non-High Risk Fair = 1
Poor = 0

rounds of deliberations (the process is described in Srinivasan and Kim, [23]),

five factors were considered relevant: customer background,payment record,
geographical location, business potential and frequency and financial sound-
ness."4The pairwise reciprocalmatrix for the five factors is presented in Exhibit
3. Lmax is 5.2146 with a consistency ratio of 0.05. The resulting global weights for
14 Note that within each factor,the expert would considerall criteriathat in the expert's opinion,

were relevant. Not all of the criteria are quantifiableand, therefore,ceterisparibus,we can expect
some divergencebetweenthe results using AHP's subjectiveprobabilitiesand objectiveprobabilities
fromMDA, logit, RPA and GP.

This content downloaded from 35.8.11.2 on Wed, 27 Nov 2013 14:34:11 PM

All use subject to JSTOR Terms and Conditions
678 The Journal of Finance
EXHIBIT 3
AHP MODEL ESTIMATION
Reciprocal Judgement Matrix for Stage IV:

Factor 1 2 3 4 5

1 1 0.33 9 3 .125

2 3 1 9 3 .125

3 0.11 0.11 1 .125 .111

4 0.33 0.33 8 1 .125

5 8 8 9 8 1

Lmax = 5.2146 C.I. = 0.054

Weights: 0.0295 0.1236 0.0016 0.022 0.823

Factor Definition:

1 : Customer Background
2 : Payment Record
3 : Geographical Location
4 : Business Potential
5 : Financial Soundness

each of the factors for stage IV customers is also shown in Exhibit 3. Each
customer was then evaluated on the five dimensions of interest (corresponding
to the five factors) and a set of subjectiveprobabilitieswas developed.The firm
provided us estimates of benefits and losses that are actually used (albeit
subjectively) currently in setting limits. The contribution margin on sales was
estimated at 10% before-tax and the variable costs were estimated at 90% of
sales. Most of the sales by the firm involve a credit period of 15 days only and as
a matter of convenience, we ignoredthe present value effect without loss of any
generality.
Subjectiveprobabilitiesfor each customerwere then integratedalong with the
estimates for benefits and costs in formulation (1) or (2) to determine credit
limits. Similarly, objective probabilities using MDA, logit, RPA, and GP were
also integrated. The results are presented in Exhibit 4. The results reveal that
the AHP limits were the closest to actual limits, which is not totally surprising.
Consistency,for the purposes of comparison,was defined to refer to cases where
the model suggested limits and actual limits were within 10% of each other.
Further, since the operational decision is to grant the credit requested or less,
positive NPV situations are considered consistent if the actual limit granted is
equal to the full amount of credit requested."5Some of the cases (=6) in which
15
If a customer requests for a credit of $100,000, then the operative decision for the firm is to
determineif the customeris worthy of being grantedfull credit. In such cases, it is possible that the
tradeofffunction specified in formulationmay imply a limit in excess of $100,000.However,this is

This content downloaded from 35.8.11.2 on Wed, 27 Nov 2013 14:34:11 PM

All use subject to JSTOR Terms and Conditions
 CreditGranting 679
EXHIBIT 4
CREDIT LIMIT COMPARISON RESULTS

Consistent Exceeds Actual Less Than

in HR cases Actual in
Non-HR Cases

MDA (Linear) 84.65% 5.12% 10.23%

MDA (Quadratid) 87.44 3.72 8.84

Logit 88.37 2.79 8.84

LPI 85.12 4.65 10.23

LPII 86.05 4.19 9.77

RPA 94.42 1.86 3.74

AHP 96.28 0.00 3.72

suggestedlimits are less than actual are the exceptions where extraneous factors
have had an influence on the decision. Elimination of these cases will further
improvethe overall consistency of the results from all the models.
The above results provide some insight on the ability of statistical and judge-
mental models to replicate expert decisions. As stated earlier, we have not
addressed the broader issue of which factors are relevant for corporate credit
granting. In our opinion, this issue motivates a complementary relationship
betweenjudgementaland statistical models. Statistical models like the RPA can
perhaps be used with significant success to isolate from a universe of potential
factors those few variablesthat have a significant bearing on the default risk of
the customer. This information can then be used in a judgemental model, like
the AHP, where objective measurements are considered along with subjective
measurements that are difficult to quantify, to provide an optimal setting for
corporatecredit granting decisions.

V. Summary
The corporatecredit grantingprocess has not received adequateattention in the
literature, perhaps because of the lack of publicly available data. We provide
comparativeanalysis of the ability of statistical classification models to replicate
the decisions of a corporate credit granting expert. It is shown that sequential
partitioning proceduresprovide slightly superior classification results. The su-
periority of sequential partitioning proceduresis likely to be a function of the
complexity of the multivariate data being analyzed. A complementary role is
suggested for statistical classification models in the corporate credit granting
process. Such models can serve the useful purposeof reducingand/or identifying

not to be treatedas an inconsistent case. Formulation(2) only specifies the maximumcredit limit to
grant.

This content downloaded from 35.8.11.2 on Wed, 27 Nov 2013 14:34:11 PM

All use subject to JSTOR Terms and Conditions
680 The Journal of Finance
a feasibleset of objectiveinformationthat the credit managercan analyze.Future
researchis being directedtowardexamining more explicitly the role of sequential
and simultaneous partitioning methodologies in understandingthe underlying
economics of the credit grantingprocess.

REFERENCES
1. E. I. Altman, R. B. Avery, R. A. Eisenbeis and J. F. Sinkey, Jr., Applicationof Classification
Techniquesin Business,Bankingand Finance,CT: JAI Press, Inc., 1981.
2. H. Bierman and W. Hauseman. "The Credit GrantingDecision," ManagementScience (April
1980), 519-532.
3. L. Breiman,J. H. Freidman,R. A. Olshen and C. J. Stone, Classificationand RegressionTrees,
Wadsworth:California,1984.
4. S. Chatterjeeand S. Chatterjee. "Estimation of Misclassification Probabilitiesby Bootstrap
Methods,"Communicationsin Statistics, (1983), 645-656.
5. S. Chatterjeeand V. Srinivasan,"GraphicalAnalysisand FinancialClassification:A Case Study,"
WorkingPaper, NortheasternUniversity, Revised,February1987.
6. W. J. Conoverand R. J. Iman. "The Risk TransformationAs A Method of Discriminationwith
some Examples,"Communicationsin Statistics Theoryand Methods,9, (1980), 465-487.
7. T. G. Diettrichand R. S. Michalski,"Learningand Generalizationof CharacteristicDescriptions:
EvaluationCriteriaand ComparativeReview of Selected Methods,"Sixth InternationalJoint
Conferenceon ArtificialInteUigence,IJCAI,Tokyo, Japan, (1979), 223-231.
8. B. Efron, "Estimatingthe ErrorRate of a PredictionRule: Improvementon Cross-Validation,"
Journalof the AmericanStatisticalAssociation,(June 1983), 316-331.
9. R. A. Eisenbeis. "Selection and Disclosure of Reasons for Adverse Action in Credit Granting
Systems,"FederalReserveBulletin, September,(1980), 727-736.
10. G. Emery,"PositiveTheoriesof Trade Credit,"In Y. H. Kim (Ed.),Advancesin WorkingCapital
Management,Vol. 1, (1987), forthcoming.
11. N. Freedand F. Glover,"ResolvingCertainDifficulties and Improvingthe ClassificationPower
of LP DiscriminantAnalysis Formulations,"DecisionSciences,(Fall (1986), 589-595.
12. N. Freed and F. Glover, "Simple But Powerful Goal ProgrammingModels for Discriminant
Problems,"EuropeanJournalof OperationsResearch,7, (1981), 44-60.
13. H. Frydman, E. I. Altman and D. Kao, "IntroducingRecursive Partitioning for Financial
Classification:The Case of FinancialDistress,"Journalof Finance,XL, (1984), 269-291.
14. M. L. Marais,J. M. Patell and M. A. Wolfson,"TheExperimentalDesign of ClassificationTests:
The Caseof CommercialBank Loan Classification,"JournalofAccountingResearch,22, (1985),
Supplement.
15. E. P. Markowskiand C. A. Markowski,"SomeDifficultiesand Improvementsin ApplyingLinear
ProgrammingFormulationsto the DiscriminantProblem,"DecisionSciences, 16, (1985), 237-
247.
16. D. Mehta. "The Formulationof Credit Policy Models,"ManagementScience, (October 1968),
B30-B50.
17. J. R. Quinlan,"LearningEfficient ClassificationProceduresand Their Applicationto Chess End
Games,"In R. S. Michalski,J. F. Carbonelland T. M. Mitchell (Eds.), MachineLearning,Palo
Alto, CA:Tioga Publishing,1983.
18. T. L. Saaty, The AnalyticalHierarchyProcess,New York:McGraw-Hill,1980.
19. T. L. Saaty, "The Axiomatic Foundation of the Analytic Hierarchy Process," Management
Science,(July 1986),841-855.
20. T. L. Saaty and L. G. Vargas, The Logicof Priorities:Applicationsin Business, Energy,Health
and Transportation, Boston: Kluwer-Nijhoff, 1982.
21. T. L. Saaty and R. S. Mariano, "RationingEnergy to Industries:Priorities and Input-Output
Dependence,"EnergySystemsand Policy, (Winter 1979).
22. V. Srinivasanand Y. H. Kim, "FinancialApplicationsof the Analytic HierarchyProcess,"TIMS
InternationalConference,Australia,(July 1986).
23. V. Srinivasanand Y. H. Kim, "Modelingthe Credit GrantingProcess in a Fortune500 Corpo-

This content downloaded from 35.8.11.2 on Wed, 27 Nov 2013 14:34:11 PM

All use subject to JSTOR Terms and Conditions
 CreditGranting 681
ration,"Symposiumon Cash, Treasuryand WorkingCapitalManagement,Montreal,Canada,
(June 1986).
24. V. Srinivasan and Y. H. Kim, "The Credit GrantingDecision: A Note," ManagementScience,
forthcoming.
25. V. Srinivasan, P. V. Medury,Y. H. Kim and A. F. Thompson, "PredictingBond Ratings: A
ComparativeAnalysis of AlternativeClassificationProcedures,"FinancialManagementAsso-
ciationMeetings,New York, (October1986).
26. StatisticalAnalysisSystem, Version 5 Edition, SAS Institute Inc., Cary:NC, 1985.
27. Y. Wind and T. L. Saaty, "MarketingApplicationsof the Analytic HierarchyProcess,"Manage-
ment Science, (July 1980), 641-658.
28. F. Zahedi, "The Analytic HierarchyProcess-A Survey of the Method and its Applications,"
Interfaces,(July-August 1986), 96-108.

DISCUSSION

ROBERTA. EISENBEIS*:This tutorial paper examines the performanceof six

different classification techniques (linear and quadratic discriminant analysis,
unorderedlogit analysis, goal programing,a recursivepartitioningalgorithm,and
an analytic hierarchy process) as a tools in credit granting problems. The
proceduresare comparedusing data on loans extended by a nonfinancial corpo-
ration.
The first main section sets forth the authors'view of the credit grantingprocess
and the problem of setting credit limits. Following others, Kim and Srinivasan
indicatethat in extending credit,the objectiveshouldbe to maximizethe expected
payoff from making a loan. That payoff depends not only on the returns from
the loan under consideration but also all future credit extensions as well.'
Formulated in this way, credit extension is a dynamic programingproblem.
However,as the authors point out it is essential intractable.2To circumventthis
dead end, they collapse the problem, as others have done before them, into a
single period problem where the period is the maturity of the loan, and the net
future benefits from future transactions are ignored.3Given the purpose of the
paper,the authors should have started with the single period model which would
have simplifiedtheir discussion considerably.Their approachis furthermuddled
by not providing a clear link between the structure of the single period credit
granting problem presented in equation 2 and the problem of setting of credit
limits describedin equation 3.
Setting credit limits is irrelevant for certain types of loans but may be very
important in providing trade credit (or in issuing credit cards to individuals or
* Universityof North Carolinaat ChapelHill
'See, for example,Mehta (1968, 1970), Biermanand Hausman (1970) and Dirickxand Wakeman
(1976).
2 Here the authorsemploy cumbersomeand confusing notation, and the model they specify does
not fully capturethe payment performanceover the life of the loan as comparedwith the value of
futurecredit extensions.
'Even this formulation,however,does not recognizethat different types of credit with different
payment patterns may be requested (eg. single payment v revolving lines, personal v corporate,
mortgagev creditcard,short termworkingcapitalv long termborrowings,etc.), and that the expected
probabilitiesof repaymentmay be functionsof differentcharacteristicsof the firm for differenttypes
of credit.

This content downloaded from 35.8.11.2 on Wed, 27 Nov 2013 14:34:11 PM

All use subject to JSTOR Terms and Conditions