TDOC

Z TA245.7
B873 8-1688
NO.1688 June 1990

Annotated
Bibliography

._\

\

l    
“Mil    Mil UNIVERSITY LIBlt 

The Texas Agricultural Experiment Station, Charles J. Arntzen, Director, The Texas A&M University System, College Station. Texas

[Blank Page in Origaal Builetin] '

A Review of Agricultural Credit Assessment Research
and an Annotated Bibliography

Eustacius Betubiza
and

David J. Leatham*

*Respectively, graduate student and assistant professor, Department of Agricultural Economics, Texas A&M
University, College Station, TX.

KEYWORDS: credit scoring/discriminate analysis/ qualitative choice models

Contents

Introduction .......................................................................................................................................................................... .. 1 6*
Development of Credit Assessment Models .................................................................................................................. .. 1
Choose Credit Classification ...................................................................................................................... ..j-, .................. .. 1
Collect Information on Past Good and Bad Borrowers ......................................................................  .................. .. 1

Identify Descriminating Factors ................................................................................................................................ .. 2 ‘
Weigh Credit Factors and Correspond to Loan Classifications .......................................................................... .. 2
Experience .....................................................................  ............................................................................................. .. 2
Discriminate Analysis ................................................................................................................................................ .. 2
Qualitative Choice Models .......................................................................................................................................... .. 4
A Linear Probability Model ...................................................................................................................................... .. 4
Probit Model .............................................................................................................................................................. .. 4
Logit Model ................................................................................................................................................................ .. 5
Comparison of Discriminant and Qualitative Choice Models .......................................................................... .. 5
Validate the Model .......................................................................................................................................................... .. 6
Institutionalize the Model .............................................................................................................................................. .. 6
Applications of Credit Assessment Models .................................................................................................................. .. 6
Annotated Bibliography .................................................................................................................................................... JP< 7
Areas of Further Research ................................................................................................................................................ .. 8
Conclusions ...............................................................................................  ........................................................................... .. 8

Footnotes ................................................................................................................................................................................ .. 8 __
Literature Cited .................................................................................................................................................................... .. 9
Appendix .............................................................................................................................................................................. .. 10
Table 1. A Summary of Credit Assement Models in Agriculture ...................................................................... .. 16

él llkill Eggkllbfifi  lqq-Q  Quvwé.

Introduction

Credit assessment (scoring) models are exper-
ence-based or statistical-based management tools
used by lenders t0 forecast the outcome of existing
loans and potential loans (loan applications). A
credit score is essentially a forecast of what will
happen to various classes of loans. Credit assessment

zgmodels can be grouped into three categories: (1)

credit-scoring models that are associated with the
decision whether or not to grant credit, (2) loan
review models that monitor the risk levels of
existing loans, (3) bankruptcy-prediction models
that can be used for preliminary credit screening
or loan review but are not credit-scoring or loan
review models per se.

As early as 1941, Durand recognized the import-
ance of credit assessment models but also issued a
warning:

A credit formula is ordinarily regarded as a
supplement to, rather than a substitute for,
judgement and experience. It may enable a
loan officer to appraise an ordinary applicant
fairly quickly and easily; and in large opera-
tions, it may be of service in standardizing
procedure, thus enabling most of the routine
work of investigation to be handled by rather
inexperienced and relatively low-salaried
personnel. A credit formula may not be
satisfactory, however, in the investigation of
extraordinary cases (p. 84).

Similary, Batt and Fowkes (1972) said the
following:

Credit scores, used in the hands of experienced
lending officers, can provide a more accurate
and consistent control of lending than is
possible either by using scores alone or by
relying entirely on experience and judgement
(p. 194).

Credit assessment in agriculture has historically
been made by personal examination of individual
credit applications and past performance records
combined with personal knowledge of an applicant

and his or her operation (Dunn and Frey 1976).
‘However, recent declines in commodity prices and

land prices coupled with high interest rates have
led to an increased frequency of farm failures and

defaulted loans, thus increasing the need for more
“sophisticated aind objective credit assessment

techniques.
Credit assessment models can be an important
tool to manage loan risk. For example, the models

pspan be used to identify loan applications with a

iigh likelihood of default, to identify existing loans
that need to be monitored closely, to price loans

according to the level of risk, and to standardize
loan criteria. The standardization of loan criteria
will become especially important as secondary
markets for agricultural loans develop. It is
important to recognize, however, that credit assess-
ment models are only an input into the overall
credit environment.

This report provides a review of credit assessment
and scoring models reported in the agricultural
economics literature. First, the steps in developing
credit assessment models and a discussion of the
strengths and weaknesses of the commonly used
credit assessment methods are presented. Second,
applications of credit assessment models are
reported. Third, past credit assessment studies in
agriculture are reviewed. Fourth, areas of further
research are identified, and finally, concluding
remarks are made.

Development of Credit .
Assessment Models

Six basic steps can be used to describe the
development process of a credit assessment model
(Alcott 1985; Batt and Fowkes 1972; and Lufburrow
et al. 1984): (1) choose the credit classifications, (2)
collect information of past good and bad borrowers,
(3) identify credit (discriminating) factors, (4)
determine the weights given the discriminating
factor in assigning credit scores, and correspond
the credit scores to the loan classification scheme,
(5) validate the model, and (6) institutionalize the
model.

Choose Credit Classification

Choosing the credit classification establishes a
point of reference. The classification scheme chosen
is typically tied to a bank’s existing loan classi-
fication scheme. The following are classification
schemes that have been used in the past: (1)
vulnerable or loss problem, and acceptable, (2)
problem and acceptable, (3) prime, base, and
premium, (4) poor risk and good risk, (5) good and
bad, and (6) successful and unsuccessful (Table 1).

Collect Information on Past
Good and Bad Borrowers

A reservoir of past lending experience is essential
to developing a credible credit assessment model.
Theoretical approaches may identify some factors
that may be important when classifying loans;
however, assigning weights to these discriminating
factors requires experience. This reservoir may
originate from credit experts or from the collection
of relevant information on past borrowers. Data

should include financial, production, market/
external conditions, and subjective information.
Production information, however, has frequently
been omitted in credit-scoring schemes. Subjective
information includes considerations of applicant’s
character, management ability, marital status, age,
and loan repayment record.

Identify Discriminating Factors

Credit risk can be traced to many factors. These
factors can be grouped into broad categories such
as liquidity, solvency and collateral position,
profitability, economic efficiency, repayment capac-
ity, and borrower characteristics, including bor-
rower’s management ability. Table 1 shows the
credit factors within each group that have been
used in past studies.

Weigh Credit Factors and
Correspond to Loan Classification

Concurrent with identifying credit factors, credit
scores are determined by assigning weights to
credit factors. The correspondence between credit
scores and credit classifications is then determined.
These weights can be assigned according to experi-
ence or statistical procedures. Useful statistical
methods are discriminate analysis and qualitative
choice models like linear probability, Logit, and
Probit models. The procedures for estimating credit
scores in past studies are reported in Table 1.

Experience

The choice and weights attached to credit factors
can be based on the experience of the developer or
the consensus of opinion of loan officers and execu-
tive officers of the lending institution.‘ Weights are
established by eliciting the participating individ-
uals’ ranking of each credit factor. The average
rank for each credit factor is used as the weight.
The credit score for each loan is the credit factor
value times the weight summed over all credit
factors. Weighted scores are assigned to credit
classifications. An individual application is classi-
fied by computing a total weighted score and
identifying the category in which it falls.

The experience-based credit assessment models
are validated by comparing the classifications of
existing loans with the loan classifications based on
credit scores. The weights are modified until the
model classifies loans accurately. The credit scores
are also compared with actual loan outcomes over
time.

The strength of this procedure is that it incorpo-
rates the past experiences of the developers, thus

increasing its chances of applicability and success.
Furthermore, the users of the model can be in
cluded in the development process, thus increasingz:
credibility and acceptance. The weakness of this
procedure is that it is highly subjective, and little
statistical theory is employed in its generation. The
statistical significance of the credit score cannot be
determined. i,
Discriminate Analysis

The basic objective of linear discriminate analy-
sis, which was introduced by Fisher in 1936 and
first used on credit data (used cars) in 1941 by
Durand, is to form a linear combination of the
discriminating variables with associated weights
that will require the groups of data (acceptable
and unacceptable borrowers) to be as statistically
different as possible.

In its general form, a discriminate function can
be expressed as

where
Y = discriminate value,

Xi = quantifiable and observable
characteristics (i = 1, ..., k),

Yi= discriminate coefficients, (i = 1,  k).

The objective is to find a set of Y ’s so that for thtt,
two populations under consideration (good and bad
loans), the calculated Y’s are as far apart as
possible. The Y ’s are then estimated using a least
squares technique. Briefly, the procedure entails
the following?

(i) With matrix notation and the subscript G
denoting good loans and B denoting bad or
problem loans, the sums of squares and cross
products for the two groups are

(2) XéXG and XigXB

where

XG is of dimension n x k,

XB is of dimension nb x k,

ng is the sample size of good loans,

nb is the sample size of bad loans, and

k is the number of variables.

To get the total sums of squares and the cross

products, the two matrices are added:

(3) XéXc + XQXB = X'X ,3
(ii) Next, generate a k x 1 vector M of differences

between the means of the explanatory variables for
the two groups:

(4) Ylg ' in) I Xi

(l

Xkg - ikb I 

(iii) Then compute the Y ’s:

A (s) v = (X’X)“M

JlIGPG Y is a k x 1 vector of Y i (i = 1, ..., k).

By substituting these estimates into the general
function, we get A A
 Y :'Y1X1 ‘l’ Y2X2 ‘l’  "l" YkXk.

“The average discriminate value of a good loan is

given by

(7) Yg = vlxlg +  +9 kXkgn
Similarly the average discriminate value of a bad
loan is given by

(s) Yb = fix“, +  + ?kxkb.
In both cases, ifs are mean values. A Z statistic
can then be computed for Yg and Yb to determine a
cutoff point between good and bad loans. Let the

cutoff point be an arbitrary value Yc. For good
loans

(9) Zg = (Y, - Yip/sag
where:
Zg = the Z statistic for good loans, and
Sdg = the standard deviation of Yg.
Similarly, for bad loans,
(10) Zb = (Y, - Ybysdb
Jhere:

Zb = the Z statistic for bad loans,_and
Sdb = the standard deviation of Yb.

Let us assume that the probability of rejecting a
good loan and the probability of accepting a bad
loan are of equal significance. By setting Zg= -Zb,
Yc can be solved for as3

(11) Y, = (Sdg Yb + SdbYgJ/(Sdg + Sdb).

Using past records about good and bad loans, a
financial institution can then estimate the "Y
coefficients with equation 5. Similarly, a cutoff
value that discriminates between the good and bad
loans can be computed using equation 11. To
evaluate loan applications, relevant variables (i.e.,
those that correspond to the variables that were
used during estimation) are used to generate the
discriminant value for this particular borrower
according to equation 6. This value is to be com-

‘Apared with the cutoff value that was computed

using past records as discussed previously. If the
borrower’s computed discriminant value is less
than the cutoff value, the loan will probably be a
bad one, and if it falls above the cutoff value, then

Pthat loan will probably be a good one.

In the aforementioned formulation, the error of
turning down a good loan and the error of accepting

a bad loan are assumed to be of equal significance,
which may not be true. This assumption is a
simplification that assists in computing a cutoff
value. Such an assumption, however, may not be
legitimate because it may be less costly to reject a
good loan, thinking it is bad, than to accept a bad
loan, thinking it is good. Many costs may be involved
i_n trying to recover a bad loan. The sample means
Yg and Yb are assumed to be the true mean
discriminant values. However, these values may be
data specific. Leatham’s (1987) comparison of
explanatory variables used to develop agricultural
credit-scoring models seems to suggest the same
conclusion. This is a strong assumption, given that
they are merely sample means. Results generated
using these values should be used with caution.
Linear programming can also be used to solve
discriminant-type problems. The logic followed in
constructing a linear—programming model to solve
the credit-scoring problem is similar to that used
for discriminant analysis. Weights must be derived
for the measurement variables that will separate
scores computed for the two groups as much as
possible. This is accomplished by first establishing
a critical value or cutoff point as the boundary
between the two groups of data. Next, through a
system of constraints (each constraint representing
an observation), weights for the variables are
established that will maximize the deviationiof an
individual score from the critical cutoff valuex‘
Hardy and Adrian (1985) used a linear-program-
ming formulation. They, however, noted that the
lack of statistical measures creates a problem
because options such as using the F-value to
determine which variables should enter the scoring

equation are not available. The selection of vari-
ables to enter is controlled by the intuitive logic of
the analyst and varified through testing. Even
though coefficients from linear-programming
models cannot be tested statistically, the overall
measure of a “good” model, i.e., the portion of
observations correctly classified, is still present.
Another problem with credit-scoring models
generated with linear programming is that they do
not have an intercept term; thus coefficients
assigned to each variable must account for all
variation in classification. Because of this, algebraic
signs of the coefficients in the function will not
always be as expected.

A major advantage noted by Hardy and Adrian
(1985) is that the user does not need a high level of
statistical knowledge to interpret and analyze the
results. In addition, restrictive statistical assump-
tions are not required. Another advantage of the
linear-programming approach is that the weights
or penalties associated with incorrect classification

can be changed. Conservative lending programs
would attach relatively higher costs for misclassify-
ing problem accounts during formulation of the
linear program by judiciously attaching higher
weights.

Qualitative Choice Models

In many situations, the dependent variable in a
regression equation is not continuous but represents
a discrete choice, such as purchasing or not
purchasing a car, voting yes or no on a referendum,
or participating or not participating in the labor
force. Models involving dependent variables of this
kind are called qualitative response models. The
observed occurrence of a given choice is considered
to be an indicator of an underlying, unobservable
continuous variable, which is characterized by the
existence of a threshold (or thresholds); crossing a
threshold means switching from one alternative to
another. As Kmenta (1986) notes, the complexity of
estimation and testing of models with qualitive
dependent variables increases with the number of
alternative choices. The simplest models are those
with only two alternatives involving a binary or
dichotomous dependent variable. In the context of
credit scoring, one wishes to find a relationship
between a set of attributes describing a loan
applicant and the probability that the loan will
turn out to be a good or a bad loan. The dependent
variable is given the value of 1 (when the loan is
good) or 0 (when the loan is bad). The estimation
procedure may be taken one of three ways: as a
linear probability model, as a Probit model, or as a
Logit model. Each of these estimation procedures
is discussed as follows.

Linear Probability Model
The regression form of the model is
 Yi: 0t +  ‘l’ 
where

Yi = 1 if loan is classified as good and 0 if loan is
classified as bad;

a = constant;

B = change in probability of Yi occuring, when
Xi changes by 1 unit;

Xi = value of the attribute, i = 1,  n;
Ei = error term.

The regression equation, which can be estimated
using ordinary least squares, describes the probabil-
ity that an individual’s loan will turn out to be good
(i.e., the individual will not default). When Xi is
fixed, the probability distribution of 6i must be
equivalent to the (binomial) probability distribution
of Yi. Classical statistical tests cannot be applied to

the estimated parameters because the tests depend
on the normality of the errors.

The variance of the error term is given by \:/

<13) Ec?) = a"; = Eomu - E<Y.)1.

which suggests that the error term is heterosce-
dastic (i.e., the variance of the error term is not
constant for all observations). Observations for —~
which Pi (which is equal to E(Yi)) is close to 0 ori
close to 1 will have relatively low variances, whereas
observations with Pi closer to 0.5 will have higher;
variances. The presence of heteroscedasticity results
in a_loss of efficiency but does not in itself result in
either biased or inconsistent parameter estimates
(Pindyck and Rubinfeld 1981). Using weighted
least squares to correct for heteroscedasticity does
not guarantee that the estimated probabilities will
fall between 0 and 1. Adjusting for such an anomaly
by dropping out the responsible variables or by
arbitrarily setting these probabilities that fall
outside the 0 to 1 interval to numbers like 0.01 for a
lower limit and 0.99 for an upper limit may result
in inefficient estimates, particularly for small
samples.

As previously noted, because the linear-probabil-
ity model involves the interpretation of predicted
values of Yi as probabilities, a problem arises when
a predicted value of Yi lies outside the 0 to 1

interval. Constraining the model to this intervas;

might yield unbiased estimates, but the predictions
obtained from the estimation process are clearly
biased.

To overcome the 0 to 1 interval problem, a
cumulative probability function is used to transform
the values of the attribute Xi, which may range in
value over the entire real line, to a probability that
ranges in value from 0 to 1. The resulting
probability distribution may be presented as

(14) Pi = F(<z + BXi) = F(Zi)

where F is a cumulative probability function and
X is stochastic. The two most commonly used
cumulative probability functions are the normal
and the logistic functions.

The cumultaive normal probability function is
used in the Probit model, whereas the cumulative
logistic probability function is used in the Logit
model. I
Probit Model \.

The Probit probability model is associated with-s

the cumulative normal probability function. The
standardized cumulative normal function is written
as

<15) P, = F(Zi) = i} (21) -1/2 EXP (s-z/Z) dswo (zi/
_@ \./
where s is a random variable that is normally“

distributed with mean zero and unit variance.

EXP represents the base 0f natural logarithms,
which is approximately equal to 2.718. The variable

3P, will lie in the 0 t0 1 interval. P; represents the

probability that a loan will turn out to be good. To
obtain an estimate of the index Zi, the inverse of
the aforementioned cumulative normal function is
taken as

(1a) z, = F-1(Pi) = ~= + BXi.

Another similar probabilistic model is the Logit
model discussed next.

Logit Model
The Logit model is based on the cumulative
logistic probability function and is specified as

(17) P, = F(z,) = EXP (z,)/ {i + EXP(Zi)} ,

-w<Zi<oo.

The logistic and Probit formulations are similar,
the only difference being that the Probit function is
similar in form to the cumulative normal function.
The Logit Model, however, is easier to use computa-
tionally and is often substituted for the Probit
model.

For grouped data, a Logit model can be estimated
using OLS as the following formulation:

(18) z, = log (Pi/(i - P,» = XiTB

/"'\ where Isi is the fraction of good loans in each class

or grouping. However, for both Probit and Logit,
the estimator of choice is a maximum likelihood

estimation because this procedure does not require '

that the data be grouped and thus allows for each
observation within the sample to be associated
with a distinct probability. For large samples, all
parameter estimates are consistent and efficient
asymptotically. In addition, all the parameter
estimates are known to be (asymptotically) normal
so that the analog of the regression t test can be
applied. In this case, the ratio of the estimated
coefficient to its estimated standard error follows a
normal distribution. To use a maximum-likelihood
estimation, a log-likelihood function is formulated
as

n1 n2
(19) logL= 2 log Pi'+2 log(1-Pi)
i = 1 i = 1
where n1 refers to the number of observations

when Y, = _1, and n2 refers to the number of
observations when Yi = 0. Furthermore,

(20) P, = 1/{1 + EXP (-XT B )} and
(21) 1-Pi= 1/{1 + EXP (XTB )l.

"\ To obtain the slope estimators fi, the log-likelihood

function is differentiated with respect to B, the

result is set at zero, and the parameters are solved
for. To test for the significance of all or a subset of
the coefficients in the Logit or Probit model when
maximum likelihood is used, a test using chi-
square distribution replaces the usual F test.

Typically, if the estimated probability is greater
than 0.5, then the first alternative is selected
(Amemiya 1981); that is, the loan will be good. On
the other hand, if the estimated probability is less
than 0.5, the second alternative is selected; that is,
the loan will be defaulted. With the current easy
accessibility to computer software, qualitative
choice models are easy to estimate.

Comparison of Discriminant
and Qualitative Choice Models

As noted earlier, a discriminant function helps
classify a loan as good or bad according to borrower
characteristics. The Probit model, which uses a
cumulative normal distribution function, and the
Logit model, which used a logistic function, can
also classify loans using a qualitative dependent
variable.

There are other limitations of discriminate
analysis. Amemiya (1981) concedes that the litera-
ture on discriminate analysis is typically more
concerned with the art of discrimination or class-
ification than with the estimation of the parameters
that characterize the right-hand side of the
discriminate function. Thus, it is not discernible
how statistically significant these variables really
are. On the other hand, the major interest of the
analysis of qualitative models lies in estimation.
Another distinction between a discriminate analysis
model and a qualitative choice model is the fact
that the former model specifies a joint distribution
of Yfs and Xfs. Amemiya (1981) further notes that
in the discriminate analysis model, the statement
that the loan will turn out to be good logically
precedes the determination of X. Therefore it is
more natural to specify the conditional distribution
of X given Y. This assumption is formally stated as

 Xi I Yi = 1~N(Ll1, 21)

Xi l Yi = 0"’N("2» 52)

where Yi = 1 refers to the fact that a loan will be
good and Y; = 0 refers to the fact that the loan will
be bad or unacceptable. The means of the two
populations are given by u 1 and u 2, whereas the
variances are given by 21 and 22. During
estimation, 2 1 and 2 2 are equated. That is, it is
assumed that two populations of multivariate
normal random variables exist where the variables
of the two populations have different means but a
common variance-covariance matrix.

First, the distribution of some of the credit
factors used are not normal (Collins and Green,
1982), and second, the variances of acceptable
(good) and nonacceptable (bad) loan applications
may differ considerably (i.e., 21 i Z2). This may
then create inefficiency in the statistical estimates.
Amemiya (1981) notes that if normality is not
assumed, the discriminate analysis estimator loses
its consistency in general, whereas the Logit
maximum-likelihood estimator (one of the qualita-
tive analysis techniques previously discussed)
retains its consistency. Thus, one would expect the
Logit maximum-likelihood estimator to be more
robust. Amemiya concludes, however, that normal
discriminate analysis (i.e., discriminate analysis
that assumes the joint normal distribution of X1
and Yi) seems to be more robust against non-
normality than one would intuitively expect. He
points out, though, that this robustness may be
conditioned to binary independent variables that
characterized the work he reviewed.

In addition to being able to classify loans,
therefore, qualitative models render themselves to
statistical tests of significance, which makes them
the most preferred models of choice.

Validate the Model

A credit-scoring scheme has to be validated
before it can be implemented with confidence. In
the initial development of a credit-scoring model,
the model can be used to predict the outcome of
out-of-sample historical loans. An outcome of a
high percentage of successful classifications of loans
is the first step in validation. The second step in
validation is to score credit applicants for a time
even though the scores are not used in credit
decisions. These scores are compared with the
credit manager’s decisions. If differences exist,
additional evaluation is needed to determine if the
credit manager’s decisions are out of line or if the
credit-scoring scheme needs to be revised. After
the credit-scoring scheme is implemented, continual
validation is needed to assure that it is correctly
classifying loans.

Institutionalize the Model

Successfully implementing a credit-scoring
scheme in a large lending institution can be a
tremendous undertaking. First, a uniform system
of data collection must be implemented and main
tained. Without uniform data, credit scores will
not be consistent. Second, an agreement by bank
management, credit managers, and loan officers
about a viable credit-scoring scheme must be
reached. Opinions may vary, but in the end, a
general acceptance must be reached or the discord

could undermine the effectiveness of the program.
Third, bank personnel involved in the use and

evaluation of the credit-scoring scheme must bex

well trained in its purpose and in how it can be
used as an aid in credit decisions. Finally, continual
updating of the credit-scoring program is required
as economic conditions change land additional
information is obtained.

Applications of

Credit Assessment Models

Credit assessment models can be applied in a
number of ways (Alcott 1985; Batt and Fowkes
1972; Leatham 1987). The primary application is to
screen loan applications. Loan applications can be
grouped into categories such as weak, strong, and
those that require special attention. The applica-
tions with a high likelihood of being problem loans
(high loan-servicing cost, delinquency, or default)
should be rejected or priced to reflect the additional
risk and servicing costs. Immediately rejecting the
weak loan applications avoids incurring additional
costs of credit checks for unacceptable loan
applications. The applications with a medium
rating can be given special attention. This may
prevent rejection of loans that have an acceptable
likelihood of being of good loan (low loan-servicing
costs and timely payments of principal and interest).

The categorization of loan applications according
to credit scores can be used to allocate farm accounts
among loan officers. Credit applications with strong
scores and those with weak scores could be given to
a less experienced loan officer. Credit applications
with medium scores could be given to a more
experienced officer. Credit scores could be used to
separate loans that require a minimum of servicing
and those that need careful attention. The loans
could be assigned to the officers according to their
experience and capacity.

Credit assessment scores can provide a basis for
pricing loans. The loan price would relate to credit
risks and servicing costs. Credit scores would also
provide more uniform loan-pricing criteria.

Credit assessment models can be used to manage
loan portfolios. First, credit scores can flag existing
loans that need special attention. Second, the overall
level and distribution of risk in a loan portfolio can
be monitored by observing the changes in the level
and distribution of credit scores. This provides a
means by which lenders can monitor the trend in
the soundness of their existing farm loan portfolio.
It also helps monitor competition and changes in
the farm economy as these changes are reflected in
the creditworthiness of loan applications. Third.
credit scores provide a preferential basis for

(l

(\

reduction of bad loans during times of credit
restrictions. The volume of weak loans can be
reduced by adjusting the cutoff score. Fourth,
credit assessment models used over time may be
used to determine the most appropriate credit
limits for the various classes of loan applicants.
Finally, credit scores can be used to measure the
effect of promotional and advertising policies. The
effects of advertisement programs on the quality of
loan applicants can be assessed by monitoring the
average credit scores of all applicants over a period.
An improvement in credit scores provides evidence
that promotional policies effectively increase loan
portfolio quality, and conversely.

Credit assessment models are educational tools.
First, credit assessment models can be used to
teach inexperienced loan officers important credit
factors and the relationships to consider when
making a loan. Moreover, credit scores can be used
as a lending control measure by comparing a
trainee’s credit decision with the credit scores.
This comparison can also lead to some constructive
self-examination and help identify where additional
training is needed to improve credit analysis.
Trainees can also be given applications having
favorable credit scores early in their training and
thus be provided with training in making potential-
ly good loans. Of course, as their training pro-
gresses, they can be given loans with only marginal-
ly favorable scores and unfavorable scores to round
out their training. In all cases, the credit models
should be used only as input to the decision process.
The decisionmaker must ultimately decide whether
or not to make the loan. Second, agricultural credit
assessment models can be used as a means of
communicating to bank officers unfamiliar with
agriculture and the status of their farm loan
portfolio. Finally, the credit assessment model can
be used as a counseling tool, helping to explain to
borrowers the credit decisions.

Credit assessment models can help determine
the loan information that should be collected and
reported. Information collection is costly; therefore,
only the relevant information should be collected.
Furthermore, computers can print out reams of
data, too much to assimilate; thus, only the relevant
information should be reported.

Annotated Bibliography

An annotated bibliography that represents a
literature review of credit assessment (scoring)
models relating specifically to agriculture is
presented in the appendix. Table 1 provides a
summary of this review.

The loan classification most frequently used has
been problem loans or acceptable loans (Johnson

and Hagan 1978; Dunn and Frey 1976; Hardy and
Weed 1980; Hardy and Adrian 1985; and Dunn
1974). Other similar two-category classifications
were poor risk versus good risk (Bauer and Jordan
1972; and Hardy and Patterson 1983), good versus
bad loans (Morris et al. 1980), and successful versus
unsuccessful (Reinsel 1963). However, a four-
category classification was proposed by Kohl (1987).
and a five-category classification by Alcott (1985)
(Table 1).

A number of credit factors have been used in
estimating credit-scoring models. However, almost
every study has used the ratio of total debt to total
assets (Table 1) and has found it significant in
credit scoring. About half the studies found the
ratio of current assets to current liabilities to be a
good predictor of loan quality. Several measures of
farm profitability, such as the return on investment
(Alcott 1985), have been suggested as an important
credit factor, but there has been little agreement
on the most suitable measure. Qualitative variables
like marital status (Bauer and Jordan 1972) and
meanagement ability (Kohl 1987 and Reinsel 1963)
are beginning to be incorporated in credit assess-
ment models, but so far, few studies have made use
of them.

More than 90% of the studies conducted used
data from Production Credit Associations and
Farmers Home Administrations because of the
extensive detail of the data that they carry (Bauer
and Jordan 1972). Only two studies used data from
the Federal Land Bank (Hardy and Adrian 1985
and Hardy and Patterson 1983), whereas Alcott’s
(1985) paper is based upon her experience at Oneida
National Bank, New York, and Tongate’s (1984)
paper" is based upon his experience at the Federal
Intermediate Credit Bank of Louisville, Kentucky.
The term simulation in Table 1 refers to credit-
scoring propositions that are based on hypothetical
data.

Discriminate analysis was the approach most
frequently used (Johnson and Hagan 1978; Dunn
and Frey 1976; Hardy and Weed 1980; Bauer and
Jordan 1972; Hardy and Patterson 1983; and Daniel
1974). Hardy and Adrian (1985) used a linear-
programming form of discriminate analysis and
reported little difference in performance from the
“mainstream” discriminate model. Qualitative
choice models have not yet received widespread
use. However, both Park (1986) and Lufburrow et
al. (1984) have used Probit.

The number of observations used in the studies
ranged from 99 (Dunn and Frey 1976 and Dunn
1974) to 2200 (Hardy and Patterson 1983) (Table
1). The percentage of observations successfully
classified after model estimation have ranged from
a low of 62 (Johnston and Hagan 1983) to a high of

85 (Bauer and Jordan 1972). To our knowledge,
only two credit-scoring models were institution-
alized: St. Louis FICB (Johnson and Hagan 1978)
and Oneida National Bank, New York (Alcott
1985).

Areas of Further Research

Gustafson (1987) observed that agricultural
lenders make only limited use of formal credit-
scoring models because they believe that such
models quickly become outdated, are difficult to
reestimate, lack general robustness, and have less
than 100% accuracy. He argues that more educa-
tional programs need to be developed to help lenders
implement methods of credit scoring.

In addition to more educational programs, a
need exists for more research that improves the
usefulness of credit-scoring models for lender

_ decisionmaking. Gustafson suggests five areas of

needed research. First, existing models can be
validated. Second, models can be fine-tuned. This
calls for developing more adequate measures of
loan quality such as production variables and
qualitative variables including character, manage-
ment ability, and financial goals. Third, procedures
for incorporating the loan portfolio effects in the
credit decision should be developed. The correlation
of a new loan with the existing loan portfolio will
affect systematic risk and the level of loan portfolio
risk. Finally, loan dynamics should be considered.
A loan may be profitable over several years but
may incur losses in the short run. The trade-off
between short-run losses and long-run profitable
customers should be considered in the credit-
granting decision. In addition to the areas of
research suggested by Gustafson, procedures need
to be investigated that incorporate general economic
conditions and specific industry/ commodity outlook
and cycles into the credit assessment models. This
is especially important for term loans.

Credit assessment models can also be used as
input into other credit decision models. First, more
work is needed in developing bootstrapping models
that filter loan officers’ credit decisions. Others
have reported superior performance by models of
the decisionmaker over performance of the decision-
maker himself or herself (Dawes and Corrigan
1974). Credit-scoring models can serve as the basis
for developing and testing bootstrapping models.

Second, with the prospect of secondary markets -
being developed for agricultural loans, effective ~

loan standardization is needed. An important aspect
of loan standardization is the identification of loans
that exceed a minimum credit quality. Credit
assessment models can be used to help identify

credit factors that are good predictors of credit
quality and the critical values of each credit factor.
Finally, in recent years, much activity has beer

directed toward developing expert systems fog;

financial analysis. A major role of an expert system
is to evaluate the financial condition of a firm,
which can be measured by a set of financial criteria
and the weight given to those criteria. Credit-
scoring models can be used to define the relative
importance of different financial factors for analysis
and to test these criteria in the lending and
borrowing activity. A need exists to link more
carefully the current work in expert systems and
the credit-scoring models.

Conclusions

A number of credit-scoring models have been
developed for agriculture. Credit assessment
(scoring) models can be used as a tool to help bank
management (1) screen loan applications, (2)
allocate loan accounts among loan officers according
to qualifications, (3) price loans equitably, (4)
manage loan portfolios, (5) educate bank personnel
and farm customers, and (6) determine the loan
information that should be collected and reported.
However, even with the strong arguments for the
development and use of credit assessment models,
only a few have been implemented by financial

institutions. Perhaps better educational programa;

are needed to substantiate the usefulness of credit
assessment models. Better data also needs to be
collected, and existing models need further valida-
tion and fine-tuning. Finally, further research is
needed to identify credit factors and appropriate
credit factor weights that maximize the likelihood
of correctly predicting the outcome of potential
and existing loans.

Footnotes

lCredit scoring by experience refers to a collec-
tion of nonstatistical methods of loan rating;
development is largely based upon the developers’
own past experience. No common procedure of
assigning weights to credit factors exists. Alcott
(1985), for example, suggested conducting an

opinion poll among the account officers about their Q

perception of the importance of each credit factor.
Weights were then assigned to each factor accord-
ing to its perceived importance.

2This development follows Bauer and Jordan

(1972). ~¢ i

3The negative sign in the equality Zg = -Zb is just
a mathematical relation to account for the fact that
Z and Zb fall on opposite sides of their respective
distributions.

4A linear-programming problem can be expressed
as follows:
Maximize  aiDi -  ciDi

1 =1 1 = 1
Subject to i E‘ WiVi-i - Di + Di’ _>_ Ci
J I

j = number of variables for each acceptable
borrower i,

j gwivii + Di" - Dis Ci
for each unacceptable borrower i, and

l1 Tl
.2 D15. 2 Ci
1 - 1 1 - 1
for Vi; V2 being unrestricted in sign; D+i, Di-Z 0
where

a i, Bi = constants associated with deviational
variables,
Di = positive deviational variable;
Di = negative (“shortfall”) deviational variable;

Wi = constant associated with jth borrower character-

istic;
Vii = jth borrower characteristic associated with
ith borrower;

Ci = cutoff value for the ith borrower.

The objective function attempts to maximize the
summed deviations from the established cutoff
score. The weights assigned to deviational variables,
a: i and Bi, are arbitrary. Hardy and Adrian (1985)
note, however, that Bi values should always be
greater than ¢~ values to indicate the penalty for
not satisfying a constraint and to prevent the
system from being unbounded. They further note
that although the cutoff score C is also arbitrary, it
should be the same for all observations.

5For a more rigorous discussion, refer to Pindyck
and Rubinfeld (1981, pp. 273-315).

Literature Cited

x Alcott, K. W. 1985. An Agricultural Loan Rating System.

Journal of Commercial Bank Lending 65:29-38.

Amemiya, T. 1981. Qualitative Response Models: A
Survey. Journal of Economic Literature 19:1483-1536.

Batt, C. D., and T. R. Fowkes. 1972. The Development
and Use of Credit Scoring Schemes. In Applications
of Management Science in Banking and Finance.
Grower Press.

Bauer, L. L., and J. P. Jordan. 1972. A Statistical
Technique for Classifying Loan Applications. Uni-
versity of Tennessee Agricultural Experiment Station
Bulletin 476:1-16.

Capps, 0., and R. Kramer. 1985. Analysis of Food Stamp
Participation Using Qualitative Choice Models.
American Journal of Agricultural Economics 67:49-59.

Collins, R. A., and R. D. Green. 1982. Statistical Methods
for Bankruptcy Forecasting. Journal of Economics
and Business 34:349-354.

Dawes, R. M., and B. Corrigan. 1974. Linear Models in
Decision Making. Psychololgical Bulletin 81:95-106.

Dunn, D. J ., and T. L. Frey. 1976. Discriminate Analysis
of Loans for Cash-Grain Farms. Agricultural Finance
Review 36:60-66.

Dunn, J . D. 1974. Evaluating Potential Loan Outcomes
Based on New Loan Applications for Illinois Cash-
Grain Farms. Unpublished Master’s thesis, University
of Illinois at Urbana-Champaign, Urbana, Illinois.

Durand, D. 1941. Risk Elements in Consumer Installment
Financing. New York: National Bureau of Economic
Research.

Evans, C. D. 1971. An Analysis of Successful and
Unsuccessful Farm Loans in South Dakota. Economic
Research Service, U.S.D.A.

Fisher, R. A. 1936. The Use of Multiple Measurement in
Taxonomic Problems. Annals of Eugenics. 8:179-88.

Gustafson, C. R. 1987. Promising Future Research
Related to Credit Scoring. Proceedings of Regional
Research Committee NC-161. Denver, Colorado,
October 6-7 . pp. 59-75.

Hardy, W. E., Jr., and J. B. Weed. 1980. Objective
Evaluation for Agricultural Lending. Southern
Journal of Agricultural Economics. pp. 159-164.

Hardy, E. W., and J . E. Patterson. 1983. An Objective
Evaluation of Federal Land Bank Borrowers Farm
Real Estate Debt, Credit Scoring Techniques.
Alabama, Louisiana, Mississippi. Highlights in Agri-
cultural Research, Alabama Agricultural Experiment
Station, Auburn, 30:3.

Hardy, E. W., and J . L. Adrian. 1985. A Linear Program-
ming Alternative to Discriminate Analysis in Credit
Scoring. Agribusiness 1285-282.

Hardy, E. W., Jr., S. R. Spurlock, D. R. Parish, and L. A.
Benoist. 1987. An Analysis of Factors that Affect the
Quality of Federal Land Bank Loans. Southern Journal
of Agricultural Economics 19:175-182. l

Johnson, R. B. 1970. Agricultural Loa11 Evaluation with
Discriminate Analysis. Unpublished Ph.D Disserta-
tion, University of Missouri.

Johnson, R. B.. and A. R. Hagan. 1978. Agricultural
Loan Evaluation with Discriminate Analysis. Southern
Journal of Agricultural Economics 10:57-62.

Kohl, D. M. 1987. Credit Analysis Scorecard. Journal of
Agricultural Lending 1:14-22.

Kmenta, J . 1986. Elements of Econometrics. New York:
MacMillan Publishing Company, Inc.

Leatham, D. J . 1987. An Overview of Credit Assessment
Models. In Proceedings of the Regional Research
Project, NC-161, Annual Meetings. Denver, Colorado,
October 6-7.

Lufburrow, J., P. J. Barry, and B. L. Dixon. 1984. Credit
Scoring for Farm Loan Pricing. Agricultural Finance
Review 44:8-14.

Morris, C. D., R. L. Harwell, and E. H. Kaiser. 1980.
Agricultural Loan Analysis and Agricultural Invest-
ment Analysis for the South Carolina Farmer Home
Administration. Agricultural Economics and Rural
Sociology Report. South Carolina Agricultural Experi-
ment Station, Clemson University, Clemson, South
Carolina, 411:28. .

Park, N. W. 1986. Analysis of Repayment Ability for
Agricultural Loans in Virginia using a Qualitative
Choice Model. Unpublished Master’s thesis, Virginia
Polytechnic Institute and State University.

Pindyck, R. S., and E. L. Rubinfeld. 1981. Econometric
Models and Economic Forecasts. New York: McGraw-
Hill.

Reinsel, I. E. 1963. Discrimination of Agricultural Credit
Risks from Loan Application Data. Unpublished Ph.D.
dissertation, Michigan State University.

Tongate, R. 1984. Risk Indexing: A Valuable Tool for
Today’s Lender. Agri Finance. Marchzl2.

Wee, B. J., and W. E. Hardy, Jr. 1980. Objective Credit
Scoring of Alabama Borrowers. Alabama Agricultural
Experiment Station Circular, Auburn University,
Auburn, Alabama: 249226.

Appendix

Alcott, Kathleen W. “An Agricultural Loan Rating
System.” The Journal 0f Commercial Bank Lending
65(1985):29-38.

The author discusses the importance of establishing a
loan-rating system by agricultural lenders. Such a system
could help management in pricing loans, monitoring
adherence to internal policies, drawing budgets, as well
as forecasting loan losses and net bank yields. This
system makes it easier to review loans and monitor
trends within the industry or a geographical region.

She recommends that lending institutions should
classify borrower accounts into different classes. In her
dairy farm example, she suggests the following perform-
ance criteria: liquidity (debt structure ratio, debt/dollar
sales, debt/cow, debt/income, and cashflow coverage),
solvency (leverage ratio and percentage of equity), and
efficiency (pounds of milk sold per cow, replacement
stock ratio, feed costs per milk income, machinery and
real estate investment per cow, total investment per cow,
total investment per man, and capital turnover).

These ratios are then weighted according to their
perceived importance.* After the scores are totaled, the
borrower is placed in one of the five relevant categories.
The lending institution then takes any necessary action.
Alcott concludes that a rating system forces agricultural
lenders to look beyond their instincts to a more objective
and complete analysis of farm credits.

*Refer t0 the section on “Erperierzee” for iriforrriatiori on hou" these
ireights are established.

1O

Bauer, L. Larry, and John P. Jordan. “A Statistical
Technique for Classifying Loan Applications."
University of Tennessee Agricultural Experimen‘
Station, Bul. 476(1972):1-16.

The authors sampled 43 problem loans and 41 good;

loans from a group of east Tennessee farmers who were
borrowers at Production Credit Associations. Using
stepwise regression, the authors narrowed the number

of independent variables down to (IYdebt-to-asset ratio,

(2) farm value, (3) total liabilities, (4) marital status, (5)

family expenses as a percentage of total expenses, (6)1

current ratio, (7) number of dependents, and (8) expected

income as a percentage of the previous year’s. The

statistical analysis indicated, after application of dis’-
criminant techniques, that the discriminant function
could, with 99% probability, discriminate between the
two groups of data. The analysis further indicates that
the function correctly classified 85% of the loans included
in the two groups, implying that discretion and further
subjective analysis should be applied when the discrim-
inant value is near the cutoff point because there is a 15%
chance in this case of misclassifying an applicant. The
authors caution that this, coupled with the limitation
that qualitative variables like managerial ability were
not included in the model, serves as a tool to supplement
but not to replace the subjective analysis of the loan
officer.

Dunn, Daniel J., and Thomas L. Frey. “Discriminate
Analysis of Loans for Cash-Grain Farms.” Agri-
cultural Finance Review 36(1976):60-66.

The goal of the study was to determine which character-
istics could be used to distinguish between loans that
become problems and those that remain acceptabl

several years after the original loan is granted. Accepkz;

able loans are defined by the Production Credit
Association (PCA) as loans that are highest in quality
ranging down to and including loans that have significant
credit weaknesses. “Problem” loans are defined by PCA
as those that have serious credit weaknesses and need
more than normal supervision, but are believed to be
collectible in full. Loans classified as “vulnerable” or
“loss” were not included in the study. The study con-
centrated on predicting successful loans from data
available on the original application.

Study data were from loans made to PCA cash-grain
farmers in central Illinois. These farmers obtained their
first PCA loans during 1964-68 and were still members
in 1971, when the study was conducted. Sixteen financial
ratio characteristics and six nonratio characteristics
that are potentially significant measures for classifying
“acceptable” and “problem” loans were used. Multiple
discriminant analysis was used to determine which
groups of ratios best discriminated between loans that
remain acceptable and those that become problems.
Stepwise discriminant analysis was first applied to the
original 22 characteristics to cut down on the number of
variables to be used in the final analysis. Only four
characteristics met the 95% significance level for being
included in the discriminant function: (1) the ratio of
total liabilities to total assets, (2) the amount of credit life
on the applicant, (3) the amount of note (original PCA
loan) as a proportion of net cash farm income, and (4) the
number of acres owned. The joint significance level
exceeded 99%. The ratio of total liabilities to total asse"

was by far the most important. The correlation matrifil’

for the four characteristics showed a low level of

correlation among the characteristics, reflecting the
additional discriminatory information added to the
function by each of the characteristics. In the test, the
model correctly classified 75% of the loans. Lenders
Tvithout the model correctly classified 50% of the test
oans.

Dunn, Daniel J. “Evaluating Potential Loan Outcomes
Based on New Loan Applications for Illinois Cash-
Grain Farms.” Unpublished Master’s Thesis, Uni-
versity of Illinois at Urbana-Champaign, Urbana,
Illinois, 1974.

The purpose of_this study was to analyze financial and
nonfinancial data from the borrower’s original loan
applications to see what information available at that
time could be used to discriminate between potentially
acceptable and unacceptable loans. The data were
collected from four central Illinois PCAs: Bloomington,
Champaign, Charleston, and Decatur. Sixty acceptable
loans and 39 problem loans were selected randomly to be
studied. Information was collected from the original
loan application of new members in 196431968.

Sixteen financial ratio characteristics and six nonratio
characteristics were used. By using multiple discriminant
analysis, it was found that data coming from different
years did not matter. When the data were analyzed
using stepwise discriminant analysis, only two ratios
were significant: (1) the ratio of total assets to total
liabilities and (2) the amount of note-to-net-cash farm
income. This model classified 65% of the acceptable loans
and 55% of the problem loans correctly. It correctly
classified 50% of all loans.

»-\_ Stepwise discriminant analysis was again carried out,

this time on all the 22 ratio and nonratio characteristics.
Only four were found significant: (1) total liabilities-to-
total assets, (2) amount of loan insurance, (3) amount of
note-to-net-cash farm income, and (4) acres owned. This
model classified 90% of the acceptable loans and 60% of
the problem loans correctly. It correctly classified 75% of
all loans. Hence, this four-variable model performed
better than the original two-variable model.

Evans, Carson D. “An Analysis of Successful and
Unsuccessful Farm Loans in South Dakota.”
Economic Research Service, U.S.D.A., Feb. 1971.

Evans reported on successful and unsuccessful farm
loans in South Dakota. He concentrated on existing farm
operating loans and tried to identify borrower character-
istics that showed developing unsatisfactory loan situa-
tions. He studied the differences between successful and
unsuccessful loans of 100 PCA members and 100 Farmers
Home Administration (FmHA) borrowers. All the
farmers were creditworthy at the time of their original

fapplications between 1955 and 1964. By 1964-65, however.
half of the loans were unsatisfactory in terms of
repayment. The study was concerned with developments
after the first loan was made but not with the correct
evaluation of available data by the lender at the time of

“the first loan application. The loans had to be successful

or unsuccessful for at least 2 years before the study date.
No distinctions were made among farm types, which
differed from 5,000-acre ranches to 80-acre crop farms.
The data used for analysis came from original loan

r-sapplications and from the last year’s loan applications.

Evans used discriminant analysis to test 15 character-
istics from the last year’s loan applications and 23
characteristics from the first year’s applications. He

‘l1

found that the 23 characteristics of the first year’s loan
applications were not significant. He, however, did find
significant differences in the characteristics of the last
year’s loan applications.

This study was concerned mainly with the loan
characteristics that determine the deterioration of a
loan. The five most significant characteristics of un-
successful PCA loans were (1) the high ratio of debt to
assets owned, (2) high cost of operations, (3) poor
production record,_(4) high ratio of debt to net worth,
and (5) the large size of the borrowers household. The
five most significant characteristics of unsuccessful
FmHA loans were (1) the ratio of FmHA loans to poor
production record, (2) the high cost of operation, (3) the
high ratio of nonreal estate debt to total debt, (4) the high
ratio of non-real estate debt to value of non-real estate
assets, and (5) a low ratio of net worth to total assets
owned.

Hardy, E. William, and James E. Patterson. “An
Objective Evaluation of Federal Land Bank
Borrowers Farm Real Estate Debt, Credit Scoring
Techniques, Alabama, Louisiana, Mississippi.”
Highlights in Agricultural Research. Alabama Agri-
cultural Experiment Station, Auburn, 30(1983):3-i11.

To analyze characteristics of real estate borrowers in
the Fifth Farm Credit District (Alabama, Louisiana,
and Mississippi), data were collected from the Federal
Land Bank of New Orleans by Alabama Agricultural
Experiment Station researchers. Data provided for the
analysis included more than 22,000 loans made during
the 5-year period from 1974 to 1978. A 10% sample was
drawn from the data for use in the statistical model.

Of all variables considered, two appeared to possess
significant discriminating power to distinguish between
good and bad loans. These variables were total debt
divided by total assets and loan commitment divided by
net worth. Approximately 71% of the loans in the sample
were classified correctly. In addition, 68% of the loans in
a separate test sample were classified correctly.

Hardy, E. William, and John L. Adrian. “A Linear
Programming Alternative to Discriminate Analy-
sis in Credit Scoring.” Agribusiness 1(1985):285-282.

The purpose of the study was to present a linear-
programming formulation for solving discriminant type
problems. Through a system of constraints (each con-
straint representing an observation), weights for the
variables are established that will maximize the deviation
of any individual score from the discriminant critical
cutoff value. A sample of 1984 loan accounts from the
Federal Land Bank of New Orleans was used as the
basis for the analysis. The sample included 1764 accounts
that were classified as acceptable and 216 that had
problems in meeting repayment obligations. Two vari-
ables were used: a ratio of total debt to total assets and a
ratio of loan commitment to net worth. The linear-
programming model correctly classified 82.4% of the
total borrower sample. The ordinary discriminant model
correctly classified 70.6% of the total borrower sample.
As the weight given to misclassifying problem borrowers
is increased relative to that given to acceptable loans. the
portion of acceptable loans classified correctly declines
and the portion of problem loans classified correctly
increases. Increasing the weight associated with mis-
classifying problem loans would yield a more conservative
credit-scoring model.

Hardy, E. William, J r., Stanley R. Spurlock, Donnie
R. Parish, and Lee A. Benoist. “An Analysis of
Factors that Affect the Quality of Federal Land
Bank Loans.” Southern Journal of Agricultural
Economics 19(1987):175—182.

The purpose 0f the research presented in this paper
was t0 examine the agricultural real estate credit market
and determine which loan, borrower, and farm business
characteristics are most important in discriminating
between loans that are good (borrowers are able to meet
repayment obligations) and those that have deteriorated
to the level of foreclosure. It was hypothesized that
certain borrowers’ loan and farm business characteristics
would be significantly different between borrowers who
are making their payments and those who had suffered
foreclosure. An additional justification for the analysis
was the need to determine if the most important dis-
criminating characteristics in the current (i.e., at the
time the research was done) financial market for agri-
culture was different from that existing in the past.

Data for the analysis were taken from the loan files of
the Federal Land Bank (FLB) in the Fifth Farm Credit
District, Jackson, Mississippi, in spring 1985. The data
represented a sample of loans closed between January 1,
1979, and December 31, 1981, in Alabama, Louisiana,
and Mississippi. Data from those years were selected
because they represented relatively recent history.
Furthermore, the loans were of sufficient age to provide
some indication of whether the borrower would be able
to meet loan payment obligations.

A stratified random sample of loan accounts was taken
so that observations would lie at both extremes of the
performance scale. Good loans were those that were
having no problems with repayment (not classified as
problem, vulnerable, or loss), and bad loans were those
that had already suffered foreclosure. A total of 68
observations were classified as good, and 76 were from
foreclosed accounts.

Discriminant analysis was used to estimate a credit
assessment model. Four variables proved to be important
in discriminating between good loans and those that had
been foreclosed. The variables were the ratio of total
debt service to total income, the ratio of acres on security
to acres owned, the ratio of loan amount to appraised
value, and the ratio of debts to assets.

The model correctly classified 82.6% of the sample
data. To statistically verify the validity of the dis-
criminant function, the U-method was preferred on
grounds that it is a particularly appropriate technique
when sample sizes are relatively small, as was the case in
their analysis. With this method, one observation at a
time is deleted from the sample, and the discriminant
classification function is derived using the remaining
observations. The deleted observation is then classified
with the new function. This process is continued until n
classification functions, each using n-1 observations,
have been derived where n is the number of observations.
The “test of goodness” is the measure of the portion of the
individual observations that are classified correctly. For
the data used in this research, the U—method correctly
classified 79.9% of the observations. Since this level of
correct classification is relatively close to that achieved
by the initial function, 82.6%, it could be assumed that
the estimation error rate of about 17.4% in the original
model was valid. This error rate would be associated
with the classification of extreme cases (good and
foreclosed accounts). Variables found to be important in

12

the study are similar to those found by previous
researchers.

Johnson, R. Bruce. “Agricultural Loan Evaluation<y

with Discriminate Analysis.” Unpublished Ph.D
Dissertation, University of Missouri, 1970.
This study was conducted in pursuit of the following
objectives: (1) identify factors that influence the financial
success of selected agricultural borrowers; (2) develop a a

credit-scoring model for evaluating loan applications for 1v

selected Missouri farm operators; and (3) test the
accuracy of the credit-scoring model on loan data from
several PCAs.

The study’s mafjor concern was to develop a computer l

model for classi ying PCA loans into two categories:
acceptable loans and problem loans. Data for the study
were acquired from loan applications in three Missouri
PCAs. Twenty-four counties were represented by the
389 loan examples selected for the analysis. A representa-
tive sample of loan applicants was taken from each
category of borrowers.

Four discriminant models were developed for testing
on the relevant financial data. Models 1 and 2 were
composed of three variables. Both models included the
repayment index and the ratio of current assets to
current debt. The third variable for Model 1 was the
ratio of total debt to total assets. The final ratio in Model
2 was the ratio of net worth to total debt. The coefficients
of these two models were applied to the relevant financial
data of both the secured and unsecured loans.

Models 3 and 4 were developed for application to the
secured loans only. These two models were analogous to
Models 1 and 2 but included a fourth variable, namely,

the ratio of total underlying security to the total PCAé/I

commitment.

Following considerable testing of the. discriminant
functions on data from the original sample and on data
selected from loans of the Pittsfield, Illinois, PCA, Model
1 was selected as the most effective classification function
for all production credit loans. More than 60% of the
loans were correctly classified by Model 1. Moreover,
none of the problem or loss loans were classified as
acceptable loans. Thus, the study indicates that a function
can be developed that will discriminate effectively
between acceptable and problem production credit loans
on the basis of data taken from loan application forms.

Johnson, R. Bruce, and Albert R. Hagan. “Agri-
cultural Loan Evaluation with Discriminate
Analysis.” Soilthern Journal of A gricultural E conom-
ics 10(1978):57-62.

In this study, the authors classified loans into accept-
able loans and problem loans by using discriminant
analysis techniques. Data were collected from loan
applications of borrowers at three PCAs in central and
northwestern Missouri. The variables selected for use in
the model were (1) repayment indexzihe amount of loans

actually repaid each year plus the value of the marketable o
crops and livestock not sold during the year, expressed O

as a percentage of the amount expected to be repaid; (2)
current ratio: current assets to current liabilities; and (3)
debt-to-asset ratio: total debts to total assets. The
discriminant function correctly classified 92% of the
borrowers. This method of loan classification would be

suitable if the consequences of the two possible classifica-qj

tion errors were of equal significance.

T0 reduce the probability 0f misclassifying problem

ex, loans into the acceptable loan group, a new cutoff score

/P\

that had a 0.1 probability for misclassifying problem
loans was calculated. Using the tabulated Z value that
corresponds to a 1% misclassification, Bruce and Hagan
solved for a new critical cutoff value. When data from
the Mississippi Valley PCA were used for verification,
61.6% of the 378 loans were correctly classified.

Kohl, David M. “Credit Analysis Scorecard.” Jouruul
ofAg/riculfurul Lending] 1(1987):14-22.

The author proposes a classification of agricultural
financial analysis criteria by considering repayment
ability. credit management, financial position, level of
management performance and profitability. and farm
resources and individuals. The analysis centered on an
illustration of a farmer who desired to add an enterprise
that required additional capital. The author presents an
agricultural credit scorecard format, guidelines, and
yardsticks based upon previous research and experience
in the agricultural credit analysis area.

Repayment ability was measured by (1) cash flow
coverage ratio (cash earnings to annual debt payment).
(2) debt service ratio (total annual debt payments to
earnings), and (3) operating-expenses-to-earning ratio.
Financial condition measures included (1) current ratio.
(2) percentage of equity (net worth to total assets). and
(3) borrowing capacity and reserve (collateral). Credit
management was measured by (1) credit lines (number
of credit sources) and (2) supplier and creditor accounts
(bill payment status). Production management and
profitability measures were (1) high production and

\ efficiency in the top 20% of managers and (2) returns to

investment. The last category concerned individual
characteristics and farm goals.

A score of 0 to 3 was given to each subcategory. and a
total score was computed from a possible 36. A score of
28-36 points implied that the loan was very serviceable
and would most likely require minimal supervision.
whereas a 22-27 score implied that the loan was service-
able, requiring regular supervision. A questionable loan
scored 16-21 points, and if made, the loan would require
close supervision. A score below 16 implied that the loan
application should be rejected. or if the loan was already
out, a special supervision plan should be made. The
author concludes that the scorecard could act as a guide
by removing some of the subjectivity that frequents
lending analysis, but the author warns that this technique
should not replace the agricultural lender’s judgement.

Lufburrow, Jean, Peter J. Barry, and Bruce L.
Dixon. “Credit Scoring for Farm Loan Pricing.”
Agricultural Finance Review 44(1984):8-14.

This study reports the results of a credit-scoring

l technique for pricing loans to individual farm borrowers.

The credit-scoring model was developed using 1982 data
for a sample of borrowers from five PCAs in Illinois that
classify their borrowers into three risk categories for
pricing purposes-z Class I, prime (lowest risk); Class II,
base (intermediate risk); and Class III, premium (highest
risk).

To estimate and validate the model, a testing procedure
was used that divided the sample of 241 borrowers into a
model-generating sample of 202 and a test sample of 39.
The test sample was chosen randomly from the complete
sample. The coefficients of the credit-scoring model
were estimated from the model-generating sample and

13

tested for predictive accuracy using the test sample.
Additional tests also occurred on the model-generating
sample itself. as well as on coefficients estimated from
the total sample.

The Probit model with an ordinarily ranked limited
dependent variable (OLDV) representing the risk classes
was used. The independent variables used were liquidity,
leverage, profitability, collateral, tenure, repayment
ability, and repayment history. Profitability and tenure
were insignificant and thus omitted from the model.

A comparison of the model’s classification of borrowers
to the PCAs classification based on probabilities of being
in the respective classes indicated that 66% of the model-
generating sample had been classified correctly. The
greatest accuracy occurred for Class I (94%) and Class
III (91%). The accuracy was only 13% for Class II.
According to comparisons of credit scores and threshold
values, 79% of the test sample was classified correctly:
100% of Class I, 62% of Class II, and 77% of Class III. The
authors suggest that the estimation procedure could be
tailored to the characteristics of specific lenders, loca-
tions, and types of borrowers.

Morris, C. David, R. Lynn Harwell, and Eddie H.
Kaiser. “Agricultural Loan Analysis and Agri-
cultural Investment Analysis for the South
Carolina Farmers Home Administration.” Agri-
cultural Economics and Rtliflll Sociology Report, South
Carolina Agricultural Experiment Station, Clemson
University, Clemson, South Carolina, Feb. 1980 (411).

This handbook is designed to assist FmHA county
officers in making consistent loan decisions. Loan analysis
is discussed in terms of general borrower characteristics,
financial ratios, and credit scores using these ratios and
danger signals. The methods of credit scoring presented
were developed from an analysis of FmHA loans in five
counties of South Carolina.

Three financial ratios were used to develop a credit-
scoring system: ( 1) annual debt payment to net income,
(2) net worth to FmHA operating credit, (3) total debt to
net worth. These ratios are categorized into three ranges:
low, medium, and high, each range having a specific
score. A debt—payment-to-net-income ratio more than
2.44 has a score of 0, whereas a ratio of 2.44-1.76 has a
score of 166, and a ratio less than 1.76 has a score of 333.
However, a net-worth-to-FmHA-credit ratio greater than
2.78 has a score of 166; 2.78-1.73 scores 83, and a ratio
less than 1.73 scores 0. A total debt-to-net-worth ratio
more than 2.56 scores 0; 2.56-1.13 scores 249, and a ratio
less than 1.13 scores 499. When this system was applied
to a five-county sample of South Carolina FmHA loans
(actual sample size not given), the average score of bad
loans was 793, and the average good loan score was 855.
The midpoint between these scores, 824, was then
considered as the dividing point between good and bad
loans. Using this system, a loan officer can classify each
loan application.

Park, N. William. “Analysis of Repayment Ability
for Agricultural Loans in Virginia Using a
Qualitative Choice Model. Unpublished Master's
Thesis, Virginia Polytechnic Institute and State
University, May 1986.

The primary objective of this research was to determine
which factors most significantly predict delinquency of
agricultural loan accounts. The study analyzed 382 loans
issued by various commercial banks, FmHAs, and P(‘.+\.<

throughout Virginia for agricultural production opera-
tions during the years 1980-1985. Approximately 20% of
73 loan accounts had become delinquent, whereas the
remaining 209 had not. The study examined the
probability that a loan applicant, given a particular set
of financial ratio and operation characteristics, will
become delinquent in repayment of the loan. The
dependent variable for the model was a 0-1 dummy
variable, with a value of 1 given for a case of delinquency,
and 0 otherwise.

The author used a Probit model, which was estimated
according to maximum likelihood techniques. With a
90% level of confidence, the following variables were
found significant: (1) current debt ratio measured as the
ratio of current debt to total debt, (2) percentage of
equity, measured as the ratio of net worth to total assets,
(3) cash flow coverage ratio, measured as the ratio of
cash available for risk, uncertainty, and new investment
to total annual debt payment, principal and interest on
them, and operating debt, (4) cash-expense-to-receipt
ratio, computed as the ratio of cash expenses (excluding
interest) to cash receipts, (5) credit lines (number of
creditors applicant has), (6) diversification level, (7)
whether or not loan is secured by Farm Credit System,
(8) whether or not loan is secured by FmHA, and (9)
gross farm income.

Efron’s R2 and McFadden’s R2 procedures were used
as measures of goodness of fit. The former is the squared
correlation coefficient between the binary dependent
variable and the predicted probabilities and was 0.23 for
the model. McFadden’s R2 was calculated as 1-[Log
L(Q)/Log 14:30)], where the second term is the likelihood
ratio index and was 0.25 for the model. The two measures
indicated that a significant amount of variation in the
dependent variable was explained by the independent
variables in the model. These two measures were
considered excellent indicators of goodness of fit for the
model because the maximum R2 possible for uniform
distribution (number of delinquent accounts = number of
nondelinquent accounts) would be 0.33. The model
correctly classified 83% of the loan accounts.

Reinsel, Ignatius Edward. “Discrimination of Agri-
cultural Credit Risks from Loan Application Data.”
Unpublished Ph.D. Dissertation, Michigan State
University, 1963.

This study was conducted to accomplish the following
objectives: (1) to evaluate the importance of various
borrower characteristics in discriminating “successful”
from “unsuccessful” loan applicants, (2) to develop a
model that can aid in discriminating “successful” from
“unsuccessful” loan applicants based on information
available at the time the loan is under consideration, and
(3) to evaluate the effectiveness of present loan applica-
tions as sources of data for predicting the outcome of
loans.

Three offices of agricultural lenders provided data.
The lenders were asked to select a dichotomous sample
of “successful” and “unsuccessful” borrowers. The FmHA
and a PCA were used as data sources because these
lenders had more complete information on their bor-
rowers than did other agricultural lenders. Prediction
models were developed independently for each of the
samples used in the analysis. The form of these functions
was much alike although the importance of the different
variables did change.

14

Factors that seemed to be important for the PCA
borrowers were farm ownership, experience on the
particular farm, and the relationship between non-real
estate debts and total debts. Individuals who were able
to make annual gains in their net worth by taking risks
appeared to be discriminated against by the PCA.

Analysis of the Ingham County FmHA sample
produced evidence that the relationship between the
firm and the household needs to be given more considera-
tion for these borrowers. Other factors that seemed
important were the attitudes toward insurance, the
relationship between non-real estate and total debts, and
the planned debt repayment. The ability to make annual
increases in net worth before the loan seemed, in the case
of these borrowers, to be an indicator that the borrower
would succeed.

For the Eaton County FmHA sample, the past level of
living was an indicator of potential future capacity of the
farm to generate needed income. Factors such as the
relationship between debt repayment and income,
non-real estate debts, and total debts and the intensity of
the farmer’s crop program also appeared to be important.

Tongate, Ron. “Risk Indexing: A Valuable Tool for
Today’s Lender.” Agri Finance (March 1984):12.

In October 1983, PCA’s in the Louisville District
began using a risk index to identify the level of risk in a
particular loan. Tongate’s article summarizes the back-
ground and use of the program. Within the Farm Credit
System, annual credit reviews classify a loan as accept-
able, problem, vulnerable, or as a loss. However, this is
an after-the-fact rather than an early warning procedure.
Tongate and his colleagues were concerned that several
institutions were using loan-scoring or loan-indexing
systems to aid in classifying loans, not to identify risk;
i.e., they were developed to determine where a loan or
portfolio was rather than where it might go. They then
identified some 60 to 70 factors contributing to risk.
These factors were further sorted into two groups: those
related to the environment and those unique to the loan.

The researchers tested their formula using as much as
6 years of historical data on both a sample of active loans
and loans charged off during 1982. Five factors were
used in the evaluation: (1) owner’s equity, (2) collateral,
(3) repayment capacity, (4) value of farm production to
debt, and (5) loan size. Owner's equity was assigned a
maximum score of 40 points; collateral, 20 points;
repayment capacity, 15 points; value of farm production
to debt, 15 points; and loan size, 10 points. Three
categories were established: low risk, moderate risk, and
high risk. For example, for owner’s equity, a ratio of 60%
or more was categorized as low risk (assigned a score of
0), a ratio of 40%-60% as moderate risk (assigned a score
of 20), and a ratio of 40% or less as high risk (assigned a
score of 40). Different but similar ranges were established
for the other four variables.

When using this model, data from an applicant are
subjected to this analysis, and if his/her total score falls
in the range of 0-29, he/she is classified as a low risk;
30-50 is a moderate risk; and 50-100 is a high risk.
Researchers propose that this indexing could be used to
(1) sort loans for diferential handling, (2) help the loan
officer avoid the “good man = good loan” syndrome, (3)
point out specific areas of weaknesses in the loan where
the loan officer can reduce the level of risk, (4) make sure
the loan remains within the lender's risk/return criteria

t‘

~44

0n an ongoing basis, (5) determine how the loan fits into

»\_ the lender’s overall portfolio objectives, and (6) aid as a

counseling tool with borrowers. Tongate warns, however.
that loan indexing should not replace personal judgment.

Weed, B. J ohno, and William E. Hardy, Jr. “Objective
Credit Scoring of Alabama Borrowers.” Circular.

Agricultural Experiment Station, Auburn University.

Auburn, Alabama, 249(1980):26.

The specific objective of the research presented in this
report was to develop a quantitative financial analysis
system that would aid Alabama PCAs and the Federal
Intermediate Credit Bank of New Orleans in discriminat-
ing between loan applicants that would be acceptable
and those that would be weak or have problems in
repayment. Data used in this study were obtained from
loan applications or borrowers at the eight PCAs in
Alabama. All eight associations in the state were sampled
because their locations serve every county in Alabama
and thus provided a cross-sectional sample of the
Alabama farm borrower. Data on 220 loan accounts
were received from participating PCAs. A subsample of
25 problem loans and 25 acceptable loans was randomly
drawn from 220 operable accounts to be used as a test
sample for verifying the classification function developed
in the analysis. The remaining sample of 170 loan
applications contained 52 problem loans and 118 accept-
able loans.

1R

Financial and nonfinancial borrower characteristics
from original loan applications were used to identify
ratio and nonratio variables. By using stepwise dis-
criminant analysis, total liabilities divided by total assets
and annual anticipated loan repayment divided by total
assets were found to be the variables most significant in
the study and were hence used to construct a credit-
scoring function. Total liabilities divided by total assets
was found to be the most significant and was three times
as important in the function as was the other variable.

The developed discriminating function was tested
against a holdout of 25 acceptable loan and 25 problem
loans. The function correctly classified 88% of the loan
applications. It classified 84% of the acceptable loans
correctly and 92% of the problem loans correctly. It also
correctly classified 81% of the original sample.

By modifying the original function, an application
technique was developed that could be used by Alabama
PCAs and the Federal Intermediate Credit Bank of New
Orleans for classification of loan applications and existing
loans. Through the application of the derived table of
cutoff values for different acceptable percentages of
problem loan misclassification, a cutoff value could be
selected that meets management requirements for correct
classification of problem and acceptable loans.

TABLE 1. A summary of credit assessment models in agriculture.

ltem

Johnson Dunn Hardy
and Hagan and Frey and Weed

Lufburrow,
Barry, and

Dixon

Hardy
Spuflock
et al.

Alcott

Kohl

Bauer
and Jordan

A. Classification:

1.

7.
8.

Vulnerable or loss,
problem, and
acceptable

. Problem and

acceptable

. Prime, base, and

premium
I-V

. l-lV
. Poor risk and

good risk
Good and bad

Successful and
unsuccessful

B. Performance criteria

1.

Liquidity
a. Current assets/
current liabilities

b. Current debt/
total debt

c. Note-to-net-cash
income

d. Cash expenses/
cash receipts

e. Number of credit
sources

. Solvency and

collateral position

a. Total debt/
total assets

b. Net worth/total
assets

c. Net worth/
FmHA credit

d. Total debt/
net worth

e. Non-real estate
debt/total debt

f. Total liabilities

g. Loan commit-
ment/ net worth

h. Collateral

i. Loan secured by

farm credit
system

j. Loan secured by
FmHA

k. Loan amount/
appraised value

l. Supplier and
creditor accounts

16

w)

TABLE 1. (continued)
‘x

Lutburrow, Hardy
Johnson Dunn Hardy Barry, and Spurlock Bauer
Item and Hagan and Frey and Weed Dixon et al. Alcott Kohl and Jordan

3. Profitability
4 a. Return on
investment
b. Return on equity X
c. Expected in-
come as % of
previous year X
d. Gross farm
income
4. Economic
efficiency
a. Operating ex-
penses/earnings X
5. Repayment
capacity

a. Actual loan pay-
ment/ expected
loan payment X

b. Debt payment/
netincome

c. Planned debt
repayment

d. Total cash/
total debt

e. Total debt/
total income X

f. Cash earnings/
annual debt _
payment X
g. Repayment
history X X
6. Borrower's '
characteristics
a. Size of family
b. Marital status X

c. Number of
dependents X

d. Attitude toward
insurance

e. Credit insurance X
7. Management ability
a. Management
ability X
b. Tenure X
c. Diversification
level

><
><

><

C. Data

1. Commercial banks X

2. FICB, PCAs/FmHA X X X X X X
3. Simulation X
4. Federal land bank

Z5

TABLE 1. (continued)

Lufburrow, Hardy l‘
Johnson Dunn Hardy Barry, and Spurlock Bauer ‘J
Item and Hagan and Frey and Weed Dixon et al. Alcott Kohl and Jordan

D. Estimate of weights
1. Based on a
experience X X a
2. Discriminate
analysis X X X X X

3. Qualitative choice
models
a. Linear prob-
ability models
b. Logit
c. Probit X
4. Linear
programming
5. Number of
observations 272 99 220 241 144 — — 84
E. Validate percentage a
successfully classified
(%) 62 75 81 71 82.6 — —- 85

F. Institutionalized St. Louis — -— — — Oneida — —
FICB Nat’l. Bank,
N.Y.

Morris, Harwell, Hardy and Hardy and
Item and Kaiser Adrian Tongate Reinsel Patterson Dunn Park

A. Classification:

1. Vulnerable or loss,
problem, and
acceptable X

2. Problem and
acceptable X X X

3. Prime, base, and
premium

4. I-V -
5. I-IV
6. Poor risk and
good risk
7. Good and bad X X

8. Successful and
unsuccessful t X

B. Performance criteria
1. Liquidity

a. Current assets/
current liabilities X

b. Current debt/
total debt y X a

c. Note-to-net-cash
income X

d. Cash expenses/
cash receipts X

e. Number of credit @

SOUTCGS

18

TABLE 1. (continued)

Morris, Harwell, Hardy and Hardy and
Item and Kaiser Adrian Tongate Reinsel, Patterson Dunn Park

2. Solvency and
collateral position

a. Total debt/
total assets X X X X
b. Net worth/total -
assets
c. Net worth/
FmHA credit X
d. Total debt/
net worth X
e. Non-real estate
debt/total debt X
f. Total liabilities
g. Loan commit-
ment/ net worth
h. Collateral
i. Loan secured by
farm credit
system X
j. Loan secured by
FmHA X
k. Loan amount/
appraised value
I. Supplier and
creditor accounts
3. Profitability
a. Return on
investment
b. Return on equity
4. Economic
efficiency
a. Operating ex-
penses/earnings
5. Repayment
capacity
a. Actual loan pay-
ment/ expected
loan payment
b. Debt paymentl
net income X
c. Planned debt
repayment X
d. Total cash/
total debt X
e. Total debt/
total income

f. Cash earnings/
annual debt
payment

g. Repayment
history

19

TABLE 1. (continued)

Morris, Harwell,
Item and Kaiser

Hardy and
Adrian

Tongate

Hardy and

Reinsel Patterson

Dunn

Park

6. Borrower's
characteristics

a. Size of family
b. Marital status

c. Number of
dependents

d. Attitude toward
insurance

e. Credit insurance
7. Management ability
a. Management
ability
b. Tenure

c. Diversification
level

C. Data

1. Commercial banks

2. FICB, PCAs/FmHA X
3. Simulation

4. Federal land bank

D. Estimate of weights

1. Based on
experience X

2. Discriminate
analysis

3. Qualitative choice
models

a. Linear prob-
ability models

b. Logit

c. Prcbit
4. Linear

programming X
5. Number of

observations — 9180

E. Validate percentage
successfully classified
(%) -— 82.4

F. Institutionalized -— —

— 2200

99

75

382

83

Mention of a trademark or a proprietary product does not constitute a guarantee or a warranty of the product by The Texas Agriculturv

Experiment Station and does not imply its approval to the exclusion of other products that also may be suitable.

All programs and information ofThe Texas Agricultural Experiment Station are available to everyone without regard to race, color, religion, sex,

age, handicap, or national origin.

1.2M—6-9O