Credit Risk
Analysis
Using Logistic Regression Modeling

Introduction
A loan officer at a bank wants to be able to identify characteristics that are indicative of people who are likely
to default on loans, and then use those characteristics to discriminate between good and bad credit
risks.
Sample
This case study uses information on 850 past and prospective customers to execute a
Logistic Regression
Analysis. Of these, 717 cases are customers who were previously given
loans. We will use a random sample of 513 of these 717 customers to create a risk model. We will
set aside the remaining 204 customers as a holdout or validation sample on which to test the creditrisk model;
then use the model to classify the 133 prospective customers as good or bad credit risks. Binary logistic
regression is an appropriate technique to use on these data because the “dependent” or criterion variable (the
thing we want to predict) is dichotomous (loan default vs. no default).
First we display the crosstabulations below, which confirm our sample
characteristics. The first table shows that we will be using 717 cases for building and validating our model,
holding the 133 prospects aside for later scoring using the model’s coefficients.
The second table shows that we have created a variable called “validate.” Customers have been randomly
assigned one of two values of this variable. The 513 customers who will be used to build the model are
assigned a value of 1. The remaining 204 customers will be assigned a value of zero, and will constitute the
validation sample on which the model will be tested.
The crosstabulations also show that the modeling sample contains 410 customers who did not default on a previous
loan, and 103 who did default. The validation or holdout sample contains 163 customers who did not default,
and 41 who did.
Logistic Regression Analysis
Now we will run a logistic regression modeling analysis and examine the results. Our model will be testing
several candidate predictors, including:

Age

Level of education

Number of years with current employer

Number of years at current address

Household income (in thousands)

Debttoincome ratio

Amount of credit card debt (in thousands).
Our logistic regression modeling analysis will use an automatic stepwise procedure, which begins by selecting the
strongest candidate predictor, then testing additional candidate predictors, one at a time, for inclusion in the
model. At each step, we check to see whether a new candidate predictor will improve the model
significantly. We also check to see whether, if the new predictor is included in the model, any other
predictors already in the model should stay or be removed. If a newly entered predictor does a better job of
explaining loan default behavior, then it is possible for a predictor already in the model to be removed from the
model because it no longer uniquely explains enough. This stepwise procedure continues until all the
candidate predictors have been thoroughly tested for inclusion and removal. When the analysis is finished, we
have the following table that contains various statistics.
For our purposes here, we can focus our attention on the “B” column, the “Sig.” column and the “Exp(B)” column,
explained below.
The table's leftmost column shows that our stepwise modelbuilding process included four steps. In the first
step, a constant as well as the debttoincomeratio predictor variable (“debtinc”) are entered into the
model. At the second step the amount of credit card debt (“creddebt”) is added to the model. The third
step adds number of years at current employer (“employ”). And the final step adds number of years at current
address (“address”).
The “B” column shows the coefficients (called Beta Coefficients, abbreviated with a “B”) associated with each
predictor. We see that number of years at current employer and number of years at current address have
negative coefficients, indicating that customers who have spent less time at either their current employer or their
current address are somewhat more likely to default on a loan. The predictors measuring the debttoincome
ratio and amount of credit card debt both have positive coefficients, indicating that higher debttoincome ratios
or higher amounts of credit card debt are associated with a greater likelihood of defaulting on a
loan.
The “Sig.” column shows the levels of statistical significance associated with the various predictors in the
model. The numbers essentially show us the likelihood that the predictor’s coefficient is spurious. The
numbers are probabilities expressed as decimals. We want these numbers to be small, and they are, giving us
our first indication that we appear to have a good model.
For example, the value of 0.018 associated with number of years at current address indicates that we would expect
our model’s result to deviate significantly from reality only about 18 times out of a thousand if we repeated our
modelbuilding process over and over again on new data samples. The statistical significance levels
associated with the other three predictors are all smaller than 0.001 (one chance in a thousand of a spurious
result), and so they are shown simply as 0.000.
While the “B” column is convenient for testing the usefulness of predictors, the “Exp(B)” column is easier to
interpret. Exp(B) represents the ratiochange in the odds of the event of interest for a oneunit change in
the predictor. For example, Exp(B) for number of years with current employer is equal to 0.769, which means
that the odds of default for a person who has been employed at their current job for two years are just 0.769 times
the odds of default for a person who has been employed at their current job for 1 year, all other things being
equal.
Once a final model is created and validated, the information in the above table can be used to score the individual
cases in a prospect database. This will allow the bank’s marketing department to focus their acquisition
efforts on those prospects that have the lowest modelpredicted probability of defaulting on a loan. It is a
simple matter to generate computerreadable instructions that can be used to quickly do the file
scoring.
After building a logistic regression model, we need to determine whether it reasonably approximates the behavior of
our data. There are usually several alternative models that pass the diagnostic checks, so we need tools to
help us choose between them. Here are three types of tool that help ensure a valid model:
Automated Variable
Selection.
When constructing a model, we generally want to include only predictors that contribute significantly to the model.
The modeling procedure that we used offers several methods for stepwise selection of the "best" predictors to
include in the model, and we used one of these stepwise methods (Forward Selection [Likelihood Ratio]) to
automatically identify our final set of predictors.
However, since we used a stepwise variableselection procedure, the significance levels associated with the model
predictors may be somewhat inaccurate because they are assuming a singlestep process rather than a multistep
process. So we also use additional diagnostics to give us more confidence in our model.
Pseudo RSquared
Statistics.
The wellknown rsquared statistic, which measures the variability in the dependent variable that is explained by a
linear regression model, cannot be computed for logistic regression models because our dependent variable is
dichotomous rather than continuous. So we instead use what are called pseudo rsquared statistics. The pseudo
rsquared statistics are designed to have similar properties to the true rsquared statistic. The table below
shows that as our stepwise procedure moved forward from step one to step four, the pseudo rsquared statistics
became progressively stronger. For those who are familiar with the rsquared statistic from linear
regression, the Nagelkerke statistic in the far righthand column represents a good approximation to that statistic,
having a maximum possible value of 1.00. It shows that approximately 72% of the variation in the dependent
variable is explained by the four predictors in our final model.
We also test the model’s goodnessoffit using additional diagnostics (not shown here), and these additional tests
also confirm that we have a good model.
Classification and Validation.
Crosstabulating observed response categories with predicted categories helps us to determine how well the model
identifies defaulters. Here is a classification and validation crosstabulation table:
The table shows that the model correctly classified about 96% of the modeling sample’s nondefaulters and about 76%
of the modeling sample’s defaulters, for an overall correct classification percentage of about 92%.
Similarly, when applied to the holdout or validation sample, the model correctly identified about 95% of the
nondefaulters and about 81% of the defaulters, for an overall correct classification percentage of about
92%.
The doublepanel graph below provides additional information about the model’s strength. It shows probability
distributions for the probability of defaulting, separately for actual nondefaulters and actual defaulters.
The binary logistic regression model assigns probabilities of defaulting to each customer, ranging from zero to
1.00 (zero to 100%). In this case, it uses a cut point of exactly 0.50 (50% probability) as the dividing line
between predicted nondefaulters and predicted defaulters.
The leftmost graph shows that the modeling process assigned the bulk of the actual nondefaulters very low
probabilities of defaulting, far below the 50% probability cut point. And the righthand graph shows that the
model assigned the bulk of the defaulters very high probabilities of defaulting, far above the 50% cut point.
So this adds more confirmation that we have a good model.
Now that we have a valid predictive model, we can use it to score a prospect file. The graph below shows the
result after we have scored our 133 prospects.
It shows that approximately 67% of the prospects would not be expected to default on a loan. (If we had used
much larger customer and prospect samples, as would typically be the case, then the prospect sample’s results would
more closely resemble the modeling sample’s results.) Note that the separation of prospects into predicted
defaulter and nondefaulter subgroups is not quite as clean as for the modeling sample. Although larger
samples would mitigate this difference, it is typical for a model to deteriorate slightly when applied to a sample
that is different from the one on which the model was built, due to natural sampling error.
A critical issue for loan officers is the cost of what statisticians refer to as Type I and Type II errors. That
is, what is the cost of classifying a defaulter as a nondefaulter (Type I error)? And what is the cost of
classifying a nondefaulter as a defaulter (Type II error)?
If bad debt is the primary concern, then we want to lower our Type I error and maximize our “sensitivity”.
(Sensitivity is the probability that a "positive" case [a defaulter] is correctly classified.) If growing our
customer base is the priority, then we want to lower our Type II error and maximize our “specificity”.
(Specificity is the probability that a "negative" case [a nondefaulter] is correctly classified.)
Usually both are major concerns, so we have to choose a decision rule for classifying customers that gives the best
mix of sensitivity and specificity. In our example we arbitrarily chose a probability cut point of 0.50
(50%). But in practice, depending on our specific objectives, we may want to experiment with various cut
points to see how these affect our models’ sensitivity and specificity by examining the rates of correct
classification for each model.
Summary
We have demonstrated the use of risk
modeling using logistic regression
analysis to identify demographic and behavioral characteristics associated with likelihood
to default on a bank loan. We identified four important influences, and we confirmed the validity of the
model using several diagnostic analytic procedures. We also used the results of the model to score a prospect
sample, and we briefly discussed the importance of examining a model’s sensitivity and specificity in the context
of one’s specific, realworld objectives.
Back to Top
Back to Marketing Analytics page
Back to Operations Research & Analytics
page
The foregoing case study is an edited version of one originally furnished by SPSS, Inc., and is used with their
permission.
