Development and Validation of Credit-Scoring

Models1

Dennis Glennon2 Nicholas M. Kiefer3 C. Erik Larson4

Hwan-sik Choi5

July 9, 2008

1 Disclaimer: The statements made and views expressed herein are solely those of the
authors and do not necessarily represent o¢ cial policies, statements, or views of the Of-
…ce of the Comptroller of the Currency or its sta¤ . Acknowledgement: We are grateful to
our colleagues for many helpful comments and discussions, and especially to Regina Vil-
lasmil, curator of the OCC/RAD consumer credit database. Please address comments and
questions to Dennis Glennon, OCC, Risk Analysis Division, Third and E Streets, SW,
Washington, DC 20219, email: [email protected].
2 U.S. Department of the Treasury, O¢ ce of the Comptroller of the Currency, Risk

Analysis Division.
3 Cornell University, Departments of Economics and Statistical Sciences; US Department

of the Treasury, O¢ ce of the Comptroller of the Currency, Risk Analysis Division, and
CREATES, University of Aarhus, Denmark.
4 Promontory Financial Group and ceriklarson.com
5 Texas A&M University, Department of Economics

Electronic copy available at: http://ssrn.com/abstract=1180302

 Abstract

Accurate credit-granting decisions are crucial to the e¢ ciency of the decentralized

capital allocation mechanisms in modern market economies. Credit bureaus and
many …nancial institutions have developed and used credit-scoring models to stan-
dardize and automate, to the extent possible, credit decisions. We build credit scoring
models for bankcard markets using the O¢ ce of the Comptroller of the Currency,
Risk Analysis Division (OCC/RAD) consumer credit database (CCDB). This unusu-
ally rich data set allows us to evaluate a number of methods in common practice. We
introduce, estimate, and validate our models, using both out-of-sample contempora-
neous and future validation data sets. Model performance is compared using both
separation and accuracy measures. A vendor-developed generic bureau-based score
is also included in the model performance comparisons. Our results indicate that
current industry practices, when carefully applied, can produce models that robustly
rank-order potential borrowers both at the time of development and through the near
future. However, these same methodologies are likely to fail when the the objective
is to accurately estimate future rates of delinquency or probabilities of default for
individual or groups of borrowers.
JEL Classi…cation: C13, C14, C52, G11, G32
Keywords: Logistic regression, CHAID, speci…cation testing, risk management,
nonparametrics, validation.

Electronic copy available at: http://ssrn.com/abstract=1180302

1 Introduction
The consumer credit market in the United States has grown rapidly over the last two
decades. According to the Federal Reserve Board’s Statistical Release on Consumer
Credit (FRB (2006)), the total outstanding revolving consumer credit in the United
States was $860:5 billion and increasing at an annual rate of 4:9 percent as of Sep-
tember 2006. Of course, the lion’s share of this total represents debt in the form of
credit card balances carried by consumers. More than 1 billion credit cards are in
circulation in the United States; fully 74:9 percent of all families have credit cards,
and 58 percent of them carry a balance. The Federal Reserve’s triennial Survey of
Consumer Finances in 2004 showed the average and median credit card balance of
those carrying a balance was $5; 100 and $2; 200 respectively (see Bucks, Kennickell,
and Moore (2006)). Given recent trends in the consumer credit market, e¢ cient de-
cision making is more important than ever both socially (for e¢ ciency) and privately
(for pro…tability).
Financial institutions have been pressed to develop tools and models to help
standardize, automate and improve credit decisions. From an economic point of
view, increasing the e¢ ciency of credit allocation has the e¤ect of directing resources
toward their most productive applications, increasing productivity, output, growth
and fairness. From the …nancial institution’s point of view, a small improvement in
credit decisions can provide a competitive edge in a …ercely contested market, and
lead to increased pro…ts and increased probability of survival. Further, retail credit
decisions are numerous and individually small, and it is costly to devote the time of
loan o¢ cers to each application.
For these reasons, there are clear advantages to developing automated decision
tools that consistently and reliably classify credits by their overall credit quality.
Credit scoring models,or scorecards, are typically developed for this purpose and are
used to predict expected performance over a …xed-length horizon, conditional on
borrowers’past behavior, current status and other risk factors.that a¤ect willingness
and ability to repay.
Kiefer and Larson (2006) provide an overview of conceptual and statistical is-

3
sues that arise during the process of developing credit-scoring models. Bierman and
Hausman (1970); Dirickx and Wakeman (1976); Srinivasan and Kim (1987); Thomas,
Crook, and Edelman (1992); Thomas, Edelman, and Crook (2002); Hand (1997); and
others, outline the development of scorecards using a range of di¤erent mathematical
and statistical techniques. A recent research conference with industrial, academic and
supervisory participants sponsored by the O¢ ce of the Comptroller of the Currency
(OCC), the primary supervisor of nationally chartered banks in the United States,
had a full program of papers on speci…cation and evaluation of credit-scoring models.
This literature re‡ects substantial advances but not consensus on best practices in
credit scoring.
In this paper, we demonstrate a range of techniques commonly employed by prac-
titioners to build and validate credit scoring models using the OCC Risk Analysis
Division (OCC/RAD) consumer credit database (CCDB). We compare the mod-
els with each other and with a commercially developed generic bureau-based credit
score. Our model development process illustrates several aspects of common indus-
try practices. We provide a framework in which to compare and contrast alternative
modeling approaches, and we demonstrate the strengths and weaknesses of alter-
native modeling techniques commonly used to develop a scoring model. We focus
on a limited number of sample and modeling issues that typically arise during the
model-development process and that are likely to have signi…cant impacts on the
accuracy and reliability of a model.1 Speci…cally we …nd that accuracy in predicting
default probabilities can deteriorate substantially as forecasts move away from the
development time frame. We attribute this at least in part to the di¤erential e¤ects
of changing macroeconomic conditions on the di¤erent credit categories. Higher-risk
of default groups are considerably more a¤ected by small changes in the economy
than low-default risk groups. This …nding points out robustness issues that can guide
future research and applications. On the other hand, although the accuracy deteri-
1
There are other legitimate ways of addressing issues of sample design, model selection, and
validation beyond those outlined below. Moreover, we believe newer and better techniques continue
to be developed in the statistical and econometric literature. For those reasons, we emphasize that
there are alternatives to the processes outlined below that can and, under certain circumstances,
should be used as part of a well-developed and comprehensive model development process.

4
orates, the ranking or separation quality is largely maintained. The models remain
useful, but they have their weaknesses which must be realized.
One signi…cant objective of our work is to illustrate aspects of model validation
that can and we believe should be employed at the time of model develop-
ment. Model validation is a process that is comprised of three general types of
activities: (1) the collection of evidence in support of the model’s design, estimation,
and evaluation at the time of development; (2) the establishment of on-going mon-
itoring and benchmarking methods by which to evaluate model performance during
implementation and use, and (3) the evaluation of a model’s performance utilizing
outcomes-based measures and the establishment of feedback processes which ensure
that unexpected performance is acted upon. The focus of this paper is on the …rst
of these activities: the compilation of developmental evidence in support of a model.
However, as a natural part of the model development process, which involves bench-
marking alternative models and identifying of appropriate outcomes-based measures
of performance, we do touch upon some of the post-development validation activities
noted in (2) and (3). Finally, we show that there are limitations to the application
of a model developed using a static sample design as a risk measurement tool. A
model that performs well at ranking the population by expected performance may
still perform poorly at generating valid default probabilities required for pricing and
pro…tability analysis.
Section 2 describes the data development process employed to create the OCC/RAD
consumer credit database. The CCDB is unique in many ways. It contains both
tradeline (account) and summary information for individuals obtained from a recog-
nized national credit bureau, and it is su¢ ciently large to allow us to construct both
a holdout sample drawn from the population at the time of development and several
out-of-sample and out-of-time validation samples. The database also allows for one
to observe the longitudinal performance of individual borrowers and individual ac-
counts; however, models exploiting this type of dynamic structure generally have not
been developed or used by lenders and other practitioners. Such dynamic models are
consequently not within the scope of this paper.
Section 3 outlines the methods used to specify and estimate our suite of models

5
and the calibration process used to construct our scores. Section 4 describes methods
that we employ to benchmark and compare the performance of the scores within the
development sample and in various validation samples from periods subsequent to
that of the development sample. Section 5 summarizes our …ndings.

2 Sample Design of The OCC/RAD Consumer

Credit Database
Each of the three major U.S. credit bureaus Equifax, Experian, and TransUnion
maintain credit …les for about 200 million individuals. Approximately four and
a half billion pieces of data are reported to the bureaus each month by grantors
of consumer credit and collectors of public records. The bureaus are faced with
the daunting task of collecting this information on an ongoing basis and using it to
update the consumer credit histories in their repositories.
As the primary supervisor of nationally chartered banks, the OCC has a broad
set of interests and issues that it would like to analyze using data on consumer
credit performance. These include evaluating various credit scoring methods in use
by banks, developing new methods, and identifying and documenting national or
regional trends in credit utilization by product type. To this end, the OCC purchased
a large multi-year extract of individual and tradeline data from one of the three
national credit bureaus and used it to contruct the CCDB.

2.1 Bureau-Based versus Institution-Speci…c Models

Practitioners and researchers alike typically base their analysis and modeling on
samples of data drawn from one or more of the credit bureaus, historical data drawn
from their own portfolio, or a combination of both. Sample designs will vary with
the intended use of the data; however, the primary consideration in the speci…cation
of any sample design will be the population to which the results of the modeling of
analysis are to be applied.

6
 Most …nancial institutions that purchase research samples of credit bureau data
do so in order to analyze and build models that describe the credit behavior of
their current or likely future customers. In these cases, the sample design might be
limited to selecting a sample of the bank’s current or prior customers, or alternately
to selecting a sample of individuals with a generic credit score greater than some
pre-speci…ed value (under the assumption that future customers will look like those
from the past.) In contrast, large-scale developers of generic bureau-based credit
scores are interested in having these scoring tools robustly predict performance for a
broad spectrum of the consumer credit-using population and consequently will want
a broader, more nationally representative sample on which to base their work. In
many ways, the design of the CCDB and the development of the models in this paper
more closely parallel that of the later group.

2.2 Unit of Analysis

For our models, the unit of analysis is the behavior of an individual rather than that
of any one tradeline. This re‡ects a common industry practice of using bureau data
to construct credit scores for individuals rather than to develop tradeline-speci…c
scores for each of an individual’s accounts (a more common application of custom
scorecards). In credit-scoring model building, it is also commonplace to develop
summary measures of an individual’s credit pro…le across tradelines for example,
the construction of a variable measuring aggregate bankcard balance or the compu-
tation of a generic credit score and to use this attribute data in custom scorecard
construction.
As a result, the existing CCDB consists of unnecessary tradeline level data and
attribute data for sampled individuals. While some common attribute data were
obtained directly from the bureau at the time of sampling, we have the ability to
use the tradeline data to construct additional attributes as necessary. It is useful to
think of the CCDB as consisting of two component databases: an individual-level
database with attribute information, and a matching tradeline-level database with
detailed account information for every account of each sampled individual.

7
2.3 Temporal Coverage
Sample designs di¤er in their breadth and unit of analysis and in terms of their
temporal coverage. Common modeling practice in the development of credit scoring
tools has historically utilized cross-sectional sampling designs, when a selection of
consumer credit histories is observed at time t, and payment behavior is tracked over
k future time periods (k is often typically de…ned as 24 months). Scoring models
are developed to predict performance over the interval [t; t + k] as a function of
characteristics observed at time t.
In contrast, the study of the dynamic behavior of credit quality requires obser-
vations over multiple periods of time for a …xed set of analysis units that have been
sampled in a base year (i.e., a longitudinal or panel data design). In both instances,
data has to be extracted with su¢ cient detail to allow the tracking of performance,
balances, line increases, etc., by tradelines (i.e., by lender) for each unit over time.
Under a longitudinal sample design, annual extracts represent updated (or re-
freshed) observations for each of the observations in the sample. To facilitate the
objectives of illustrating existing cross sectional methods and allowing for experi-
mentation with longitudinal-based analysis, the CCDB has a unique structure. The
database has been constructed so as to incorporate a “rolling”set of panels, as well
as an annual sequence of random cross sectional samples. Rather than simply iden-
tifying a base period sample and then tracking the same individuals though time, as
might be the case in a classic panel, the CCDB seeks to maintain the representative
nature of the longitudinal data by introducing supplemental parallel structure indi-
viduals at various points in time, and by developing weights relating the panel to
the population at any point in time. Further details are presented in the following
sections.

2.3.1 Cross-Sectional Sampling

The initial sample consists of 1; 000; 000 randomly selected individual credit reports
as of June 30, 1999. Nine hundred …fty thousand of these individuals were ran-
domly sampled from the sub-population of individuals for whom the value of a

8
generic, bureau-based score (GBS) could be computed (the scoreable population),
while 50; 000 individuals were sampled from the unscoreable population. The allo-
cation of the sample between scoreable and unscoreable populations was chosen in
order to track some initially unscoreable observations longitudinally though subse-
quent time periods. Because the unscoreable segment represents roughly 25 percent
of the credit bureau population, a purely random sampling from the main credit
bureau database would have yielded too many unscoreable individuals.2

2.3.2 Longitudinal Sampling

Given the required cross-sectional size and the need to observe future performance
when developing a model, it was also determined that the sample should include
performance information through June 30, 2004 the terminal date of our data
set. The 1,000,000 observations from the June 30, 1999 sample make up the initial
“core”set of observations under our panel data design. The panel is constructed
by updating the credit pro…le of each observation in the core on June 30th of each
subsequent year. In Figure 1 we illustrate the general sampling and matching strat-
egy using the 1999 and 2000 data; counts of sampled and matched individuals are
presented in Tables 1 and 2.
In general, the match rate from one year’s sample to the following year’s bu-
reau master …le is high. Some of the scoreable individuals sampled in 1999 became
unscoreable in 2000, again due to death or inactivity, and some of the previously
unscoreable became scoreable in 2000 (for instance, if they had acquired enough
credit history). Of the 1; 000; 000 individuals sampled in 1999, 949; 790 individuals
were found to be scoreable as of June 30, 2000. As indicated in Table 2, this change
resulted from 17,339 individuals moving from scoreable to unscoreable or missing,
while 17; 129 individuals moved from unscoreable to scoreable.
Over time, the credit quality of a …xed sample of observations (i.e., the core) is
likely to diverge from that of a growing population. For that reason, we update the
core each year by sampling additional individuals from the general population and
2
Unscoreable individuals include those who are deceased or who have only public records or very
thin credit tradline experience.

9
then developing “rebalanced”sampling weights which allow for comparison between
the updated core and the current population. For example, we update the core
in 2000 by comparing the GBS distribution of the 949; 790 individuals from the
1999-2000 matched sample (tabulated using 10-point score buckets from 300 to 900,
the range of the GBS) to a similarly constructed GBS distribution for an additional
950; 000 individuals randomly sampled from the credit bureau’s master …le as of June
30, 2000. The relative di¤erence in frequency by bucket between the two distributions
was then used to identify the size of an “update sample”of individuals to add to the
1999-2000 matched sample. The minimum of these bucket-level frequency changes
(i.e., the maximum decrease rate in relative frequency) was then used as a sampling
proportion to determine the number of additional individuals that would be randomly
sampled from the June 30, 2000, scoreable population and added to the core data
set (i.e., the 1999-2000 matched …le). For 2000, the “updating proportion” was
determined to be 7 percent, resulting in the addition of 66,500 individuals from
the 2000 scoreable population to the 1999-2000 matched scoreable sample on the
CCDB. Use of this updating strategy ensures that the precision with which one might
estimate characteristics at the GBS bucket level in a given year does not diminish
due to drift in the credit quality of those individuals sampled in earlier years.
Sampling for years 2001-2004 proceeded along similar lines, with the results re-
ported again in Tables 1 and 2. The individuals who were members of the CCDB
panel in a previous year (i.e., the core) were matched to a current year’s master …le.
Individuals who were unmatched or remained or became unscoreable in the current
year were dropped from the CCDB panel and then replaced with another draw of
50,000 unscoreable individuals from the current year’s master …le. The GBS distribu-
tion from the panel was compared with that for a random cross section of individuals
drawn from the current master …le and a “updating proportion”was determined and
applied to de…ne an additional fraction of the random cross section to add to and
complete the current-year CCDB panel.

10
3 Scorecard Development
3.1 De…ning Performance and Identifying Risk Drivers
We follow industry-accepted practices to generate a comprehensive risk pro…le for
each individual. We use as a starting point the …ve broadly de…ned categories out-
lined in Fair-Isaac (2006). We summarized our own examples of possible credit
bureau variables that fall within each category and which are obtainable from our
data set; these are presented in Table 3.
Scorecard development attempts to build a segmentation or index that can be
used to classify agents into two or more distinct groups. Econometric methods for
the modeling of limited dependent variables and statistical classi…cation methods
are therefore commonly applied. In order to implement these types of models using
the type of credit information available from bureaus, it is necessary to de…ne a
performance outcome; this is usually, but not necessarily, dichotomous, with classes
generally distinguishing between “good”and “bad”credit histories based upon some
measure of performance.
In this paper, we choose to classify and develop a predictive model for perfor-
mance of good and bad credits based upon their “default”experience. Bad outcomes
correspond to individuals who experience a “default”and “good”outcomes to indi-
viduals who do not. It is our convention to assign a default if an individual becomes
90 days past due (DPD), or worse, on at least one bankcard over a 24-month per-
formance period (for example July 1999 through June 2001). Although regulatory
rules require banks to charge-o¤ credit card loans at 180 DPD, it is not uncommon
among practitioners to use our more conservative de…nition of default (90+ DPD).
We experimented with a de…nition of default based on both a 12- and 18-month
performance period. The results of our analysis are fundamentally the same under
the alternative de…nitions of default.

11
3.2 Construction of the Development and Hold-Out (In-Time
Validation) Samples
We develop our model using a conventional scorecard sample design. The re…nement
process that was applied to the CCDB and that resulted in the development sam-
ples is presented in Figure 2. A randomly selected, cross-section sample of 995; 251
individual credit …les with valid tradeline data representing over 14:5 million trade-
lines is drawn from the CCDB database as of June 30, 1999. The sample includes
733; 820 individuals with at least one open bankcard line of credit that had been
updated during the January through June 1999 time period.3 We drop 19; 122 …les
with a bankcard currently 90+ DPD, choosing to model the performance of accounts
that are no worse than 60 DPD at time of model development. A separate model for
accounts that are currently seriously delinquent (i.e., greater than 60 DPD) could
be developed (although we do not attempt to develop such a model in this paper.)
An additional 37; 436 accounts are deleted because their future performance could
not be reliably observed in our panel, leaving us with a sample of bankcard credit
performance on 677; 262 individual credit records. We split this group randomly into
two samples of approximately equal size and then develop our suite of models using
a sample of 338; 578 individual credit histories. The remaining 338; 684 individuals
are used as a holdout sample for (within-period) validation purposes.
To allow for the more parsimonious modeling of di¤erent risk factors (i.e., charac-
teristics), and possibly di¤erent e¤ects of common risk drivers, it is standard practice
in the industry to segment (or split) the sample prior to model development. We
have implemented a common segmentation by introducing splits based upon the
amount of credit experience and the amount, if any, of prior delinquency. Credit
…les that contain no history of delinquencies are de…ned as clean, and those with a
history of one or more delinquencies are de…ned as dirty.4 Because individuals with
3
A bankcard tradeline is de…ned as a credit card, or other revolving credit account with variable
terms issued by a commercial bank, industrial bank, co-op bank, credit union, savings and loan
company, or …nance company.
4
We de…ne an observation as dirty if the individual has a history of delinquencies greater than
30 DPD ever, a public record, or collections proceedings against him or her.

12
little or no credit experience are expected to perform di¤erently from those with
more experience and thicker …les, we create additional segments within the clean
group made up of individuals with thin credit …les (fewer than 3 tradelines) or credit
…les (more than 2 tradelines). On the other hand, we created two segments within
the dirty group consiting of individuals with no current delinquency and with mild
delinquency (60- DPD). Consequently, we identify four mutually exclusive segments:
clean/thick, clean/thin, dirty/current, and dirty/delinquent.
In Figure 2 we report the number of individuals and the average default rate in
each of the segments. The development sample has an average default rate of 7:19
percent. The clean and dirty segments have a default rate of 3:1 percent and 20:3
percent respectively. Our objective is to model the likelihood of default (i.e., 90+
DPD) for each segment using credit bureau information only.

3.3 Model Forms

There are several analytical modeling techniques that are discussed in the scoring
literature and used in the industry to construct a scoring model. These include
regression-based models (i.e., ordinary least squares, logit procedures), discriminant
analysis, decision trees, neural networks, linear programming methods and other
semiparametric and nonparametric techniques.5 In practice, most scorecards are
developed using a regression-based model.
We consider and illustrate the di¤erences between the three most commonly
employed model forms. First, we consider a logistic regression. Logistic regressions
are a form of generalized linear model characterized by a linear index and a logistic
“link” function. Next, we develop a form of semiparametric model in which we
retain the linear index from the parametric model speci…cation but estimate the
link nonparametrically. Although we generalize the link function from logistic to
5
By design, discriminant analysis, linear programming, and tree methods use a maximum diver-
gence (between good and bad performance) criterion for selecting the best combination of factors
and factor weights for developing classi…cation models. Regression and neural network methods use
an error minimization criterion, which is well suited for constructing prediction models. However,
regression models often perform well over multiple objectives.

13
nonparametric, we retain the assumption that the link function is the same across
segments. That is, we retain the assumption that there is a common relationship
between the value of the index and the default probability, though we no longer
require the logistic functional form. We experiment with further generalizations
to di¤erent link functions across segment;, however, these generalizations are not
especially productive, especially for the segments with smaller sample sizes. Finally,
we compare these two regression forms with a fully nonparametric model developed
using a decision-tree approach. This can be thought of as a further generalization in
which both the index and the link are estimated nonparametrically.

3.3.1 Parametric models

The parametric speci…cation is the logistic regression

0
pi = E(yi jxi ) = 1=(1 + exp( xi )) for each individual i; (1)

where yi 2 f0; 1g is an indicator variable for non-default/default, xi is a vector of

covariates, and is the vector of associated coe¢ cients. The estimates p^i of the
probability of default are derived from the estimated model

p^i = 1=(1 + exp( b0 xi )); (2)

where b is the maximum likelihood estimator of .

If we de…ne the index Z = b0 x, then Z represents the estimated log-odds

Z = ln(^
p=(1 p^)): (3)

3.3.2 Semiparametric Models

The semiparametric models use the estimated (parametric) index function to parti-
tion the sample into relative risk segments. We rank the sample by the estimated
index from the logistic regression and then estimate the link function nonparamet-
rically. Speci…cally, for this model the estimates of the default rate are equal to the

14
empirically observed default rate within each segment.
We follow current industry practice and partition the sample into discrete seg-
ments, chosen so that each band contains the same number of observations, m: Given
the sample size, we create 30 distinct segments. For each segment, the predicted
probability of default is given by
p^i = y Ji ; (4)

where Pn
yk 1fJk = Ji g
y Ji = Pk=1
n : (5)
k=1 1fJk = Ji g

and Ji 2 f1; :::; 30g denotes the segment J to which individual i belongs.

3.3.3 Variable Selection Methods

Variable selection for the parametric and semiparametric forms is accomplished

though application of each of three alternative variable selection methods; we re-
fer to the variable selection methods as Stepwise, Resampling, and Intersection.
Our Stepwise method starts with an intercept-only regression model and then
searches for the set of covariates to …nd the one with the strongest statistical re-
lationship with performance (forward selection). It repeats this process, searching
within a multivariate framework for additional covariates that are predictive of the
performance variable. As each new covariate is added, the algorithm tries to elimi-
nate the least signi…cant variables (backward selection). The forward selection stops
when the remaining covariates fail to reach a level of statistical signi…cance at the 5
percent level.
The Stepwise method, however, may result in over identifying, or over…tting,
the model especially in large samples (Glennon (1998)). To reduce this tendency to
over…t the regression model, our Resampling method is characterized by the repeated
application of a stepwise selection procedure over sub-samples of the data. Covariates
that most frequently enter the model over multiple replications are then combined
into a single model estimated over the full development sample. Speci…cally, we …rst
randomly select !-percent of the data and then run a stepwise regression. Then,

15
we repeat the resampling and stepwise regressor selection k times and choose the
variables that appear most often in the k replications (variables that occurred in
10 or more of the replications). We use k = 20 and experiment with values for
! = f20 percent; 50 percent; 100 percentg. After some experimentation, we use the
results from the 50 percent trial. We applied the stepwise and Resampling methods
separately to each segment.
Finally, we de…ne the Intersection method as the variable selection resulting from
construction of the common set of covariates that appear in the Stepwise and Re-
sampling methods. The Stepwise selection approach generates the largest, and the
Intersection approach the smallest, set of covariates.

3.3.4 A Nonparametric Model

The fully nonparametric model form does not assume a functional form for the
covariates. To implement our nonparametric speci…cation, we use a tree method
called CHAID (Chi-squared Automatic Interaction Detector) to cluster the data into
multiple “nodes” by individual characteristics (attributes). The variable selection
process searches by sequential subdivision for a grouping of the data giving maximal
discrimination subject to limitations on the sizes of the groups (avoiding the best …t
solution of one group per data point). The approach is due to Kass (1980).6 The
CHAID approach splits the data sequentially by performing consecutive Chi-square
tests on all possible splits. It accepts the best split. If all possible splits are rejected,
or if a minimum group size limit is reached, it stops. Each of the …nal nodes is
assigned with predictions that are equal to the empirical default probability, p^n for
node n. By design of the algorithm, individuals within a node are chosen to be as
homogeneous as possible, while individuals in di¤erent nodes are as heterogeneous
as possible (in terms of p^n ), resulting in maximum discrimination. Note that the
splitting of the development sample data into four segments which preceded the
construction of parametric and semiparametric models was not undertaken prior to
implementing the CHAID algorithm.
6
Various re…nements have been made to Kass’s original speci…catgion; we implement CHAID
using the SAS macro %TREEDISC (SAS (1995)).

16
 For the CHAID method we have to specify (1) the candidate variable list, (2)
the transformation of continuous variables into discrete variables, and (3) the mini-
mum size of the …nal nodes. We considered two di¤erent sets of candidate variables.
Initially, we considered all available attributes and kept only those that generated
at least one split. As an alternative, we used only those attributes that were iden-
ti…ed using the Intersection method for variable selection outlined above. In the
latter case, for each model segment (i.e., clean/thick, clean/thin, dirty/current, and
dirty/delinquent), we take the intersection of the variables from the stepwise selec-
tion process with the variables appearing 10+ times in the 20 percent, 50 percent,
and 100 percent Resampling methods, then combine the selected variables across the
model segments by taking the union of those sets of variables.
As the CHAID approach considers all possible splits, it requires the splitting of
continuous variables into discrete ranges. We chose the common and practical ap-
proach of constructing dummy variables to represent each quartile of each continuous
variable. As a validity check on this procedure we also split the continuous variables
into 200 bins. (Note that this process includes all intermediate splits from 4–199 as
special cases).
To prevent nodes from having too few observations or having only one kind of
account (good or bad), we set the minimum of observations in a node to be 1; 000.
The CHAID rejects a split if it produces a node smaller than 1; 000. Therefore the
size of the …nal nodes works as a stopping rule for the CHAID. Since this speci…cation
is rather arbitrary, we experiment with di¤erent node sizes ranging from 100 to 8; 000
observations.

3.4 Explanatory variables

In Tables 4 through 7, we report the variables selected using the Stepwise, Resampling
(50 percent) and Intersection methods for each of the segments. Each table includes
the set of variables selected using the Stepwise method, sorted by variable type (see
Table 3), for that segment of the population. In the third and fourth columns of each
table, we list the subset of variables identi…ed using the Resampling and Intersection

17
methods, respectively. The worst status for open bankcards within the last six months,
the total number of tradelines with 30+ DPD, and the total number of tradelines with
good standing are “individual credit history” variables that consistently show up as
important explanatory variables. Utilization rates for bankcards and for revolving
accounts are the more important “amount-owed” variables. The age of the oldest
bankcard tradeline enters as a relevant measure of the “length of credit history,”
and “new credit activity”is measured using the total number of inquiries within the
last 12 months and the total number of bankcard accounts opened within the last
two years. Finally, the total number of revolving tradelines active was an important
explanatory variable capturing the impact of the “type of credit used.” It is clear
from our results that a fairly small set of variables su¢ ces to capture almost all of
the possible explanatory power. In Table 8, we report the set of “splitting”variables
identi…ed under the CHAID selection method, again sorted by variable type.

3.5 Score Creation Through Model Calibration

We transform the estimated p^ into credit scores, namely Risk Analysis Division
(RAD) Scores. Credit scores are a mapping from the estimates p^ to integers. Scores
contain the same information as p^ estimates but are convenient to use and easy to
interpret. We follow industry convention and calibrate the RAD Scores (S) to a
normalized odds scale using the following rules:

1. S = 700 corresponds to an odds ratio (good:bad) of 20 : 1. Equivalently

(1 p^700 )
= 20; (6)
p^700

where p^700 is the p^ value at S = 700, and

2. Every 20-unit increase in S doubles the odds ratio. The score values, S; are
calibrated using the a¢ ne transformation:

S = 28:8539(Z + 21:2644); (7)

18
 where Z is as given in equation (3). We calculate eight di¤erent RAD scores,
one from each of the three parametric (Stepwise, Resampling, and Intersec-
tion), three semiparametric (Stepwise, Resampling, and Intersection), and two
nonparametric (all variables, and Intersection) models.

We also recalibrate the GBS so as to allow for comparison with the RAD scores.
Since we cannot observe the predicted p^ associated with the GBS, we estimate it
though a linear regression of the empirical log-odds in our sample against the score
values. Data for the regression consists of empirical log-odds estimated for 20 di¤erent
buckets of individuals sorted by the GBS and the associated bucket mean bureau
values.

4 Evaluation of Scoring Model Performance

4.1 Performance Measures
We evaluate our models based upon two primary metrics of interest: discriminatory
power and predictive accuracy. We consider two types of measures by which to as-
sess scorecard performance: separation measures and accuracy measures. These are
widely used in practice. Separation measures give the degree of separation between
good and bad performance, and the accuracy measures gives the degree of di¤erence
between the predicted and realized default rates.
A popular separation measure is the Kolmogorov-Smirnov statistic (K-S value),
de…ned by the maximum di¤erence between two cumulative distribution functions
(CDFs) of good and bad performers.
For the accuracy measure, we consider the Hosmer and Lemeshow Goodness-of-
Fit Test (H-L). It is based on the di¤erence between the realized default (or bad)
rates pj and the average of the predicted default rates, pj , for individuals grouped
into deciles (j = 1; :::; 10) (the deciles are constructed by sorting the sample by
individual predicted default rate p^i ). The H-L statistic is de…ned as

19
 X
10
(pj pj )2
HL = ; (8)
j=1
pj (1 pj )=nj

where nj is the number of observation in each of the j deciles. The H-L statistic is
distributed as Chi-square with 8 degrees of freedom under the null that pj = pj for all
j. Just to be clear, a good model should have a high value of the separation measure,
K-S, but a low value of the accuracy measure H-L. It would be perhaps better to
label the H-L as an an “inaccuracy” measure, as it is a Chi-squared measure of …t,
but the contrary convention is long established.

4.2 In-Time Validation at Development

We …rst compare the performance of models di¤erentiated by variable selection
method (Stepwise, Resampling, Intersection), given model form (parametric, semi-
parametric, nonparametric (CHAID)). Then we compare the performances of the
di¤erent model forms. The performance of scorecard models was measured in the
development samples and is presented in Tables 9 and 10. Table 9 shows the me-
dian RAD Scores by segment and by validation samples, while Table 10 shows K-S
and H-L measures constructed from pooled-across-segment model predictions and
outcomes. The models developed using the Stepwise variable selection method per-
form best at di¤erentiating between good and bad accounts, although the di¤erence
between the parametric and semi parametric approach is very small. Overall, the
nonparametric CHAID approach performed worse on the pooled data. Although all
the models perform well at di¤erentiating between good and bad accounts, none of
them is particularly accurate as re‡ected in the low p-values for the H-L test. All
but the semiparametric model generate predicted values that are statistically di¤er-
ent from the actual default performance. Because the actual (development sample)
performance is used, by design, to predict performance under the semi parametric
approach, the H-L test is not applicable for the development sample and not very
informative for the in-sample, hold-out validation data.
We also evaluate the accuracy and reliability of each model (i.e., by segment and

20
model form) as stand alone models. Table 11 shows K-S and H-L measures from the
parametric and semiparametric models for each segment.7 Individually, the segment-
speci…c models perform well at di¤erentiating between good and bad accounts. As
is commonly observed in practice, credit bureau-based models perform better on the
clean-history segments of the population as re‡ected in the nearly 20 point di¤er-
ence in the K-S values between the clean-history and dirty-history segments across
model form and variable selection procedures. It is interesting to note, however, that
the parametric models are relatively accurate on the development and in-sample,
hold-out data except for the clean-history/thick-…le segment. That latter result is
likely driving the accuracy results in Table 10, given the relative size of the clean-
history/thick-…le segment.8 These results clearly show that a model can perform
well at discriminating between good and bad accounts (i.e., high K-S value), yet per-
form poorly at generating accurate estimates of the default probabilities – a result
that illustrates the importance of considering model purpose (i.e., discrimination or
prediction) in the development and selection of a credit scoring model.
The K-S test evaluates separation at a speci…c point over the full distribution of
outcomes. In Figures 3 and 4, we plot the Gains charts for each of models. The
Gains charts describe the separation ability graphically by showing the CDF for ob-
servations with “bad”outcomes plotted against the CDF for all sample observations
(the 45-degree line serves as a benchmark representing no separation power). The
parametric and semiparametric models and the GBS produce very similar graphs,
while the CHAID models showed much weaker discriminatory power.
In Figures 5 and 6, we plot the empirical log odds by RAD score for each model, for
both the development and hold-out samples, respectively. We compare the empirical
7
By design, the actual performance within each decile (i.e., score band) from the development
sample is used to generate the predicted values under the semiparametric method. For that reason
(as noted above), the H-L test is not well designed for evaluating the accuracy of the semiparametric
models. Therefore, we use the actual performance (i.e., default rate) derived from the pooled-
segment analysis summarized in Table 10 as the predicted values in the calculation of the H-L
values for each of the semiparametric models in Table 11.
8
The more accurate model results for the semi parametric model on the development and in-
sample, hold-out data are likely to be due to the construct of the tests, and therefore, must be
interpreted carefully. Clearly an out-of-sample test will better re‡ect the true accuracy of the
models constructed using this approach.

21
log odds for each model against calibrated target values. The calibration target line
is given in eq. (7). The graphs show that the models preform relatively well for
score values below 750. Although the semiparametric models and the CHAID do not
generate estimates for scores below 600 due to the smoothing nature of the models,
we point out that the parametric model continues to perform well on the score range
below 600. For scores between 760 and 780, the parametric and semiparametric
models slightly overestimate default risk. For the score range over 800, the RAD
models underestimate the default rates. These results suggests that the lack of overall
accuracy of the model is being driven primarily by the imprecision in the estimates
at the higher end (i.e., greater than 750) of the distribution: that portion of the score
distribution, based on the median scores reported in Table 9, heavily populated by
observations from the clean-history (both thick and thin) segments.
It is worth noting that the Resampling and Intersection models generate very
similar levels of separation and accuracy measures using fewer covariates. Those
results hold across both the development and hold-out samples. The decision tree
approach (i.e., CHAID method), however, clearly generates models with lower dis-
criminate power. That result is re‡ected in the …ve-point di¤erence in the K-S values
between the Stepwise parametric model and Intersection CHAID model in Table 10.
It is di¢ cult, however, to interpret the meaning of that result. Instead, we look to the
relationship between the Gains charts in Figure 3. The Gains chart for the Stepwise
parametric model is above the Gains chart for the CHAID-Intersection model. As a
result, at each point on the horizontal axis, the Stepwise parametric model identi…es
a greater percentage of the bad distribution. For example, over the bottom 10 per-
cent of the score distribution, the Stepwise parametric model identi…es roughly 60
percent of the bad accounts, while the CHAID-Intersection identi…es only approxi-
mately 48 percent. At that point, the Stepwise model identi…es nearly 25 percent
(i.e., 12/48) more bad accounts a substantial increase over the CHAID model.
We have estimated the models using the widely, but not universally, accepted
90+ DPD de…nition for the outcome variable. It is interesting to ask whether the
model would also do well at discriminating between good and bad accounts if default
is de…ned at 60+ DPD, or evaluated over a shorter performance horizon (e.g., 18,

22
12, and 6 months). In Table 12, we summarize the observed performance over these
alternative de…nitions of performance. A reliable model should order individuals by
credit quality over a variety of bad de…nitions. In Table 13 we compare the K-S
measures from eight di¤erent RAD models and the GBS, using both the 90+ DPD
and 60+ DPD bad de…nitions. We …nd that a model’s ability to di¤erentiate between
good and bad accounts is virtually the same as re‡ected by the K-S values across
the development and holdout samples for all methods. As expected, the models
perform better under the 90+ DPD de…nition. Nevertheless, the models seem to
order observations well by credit quality for the alternative de…nitions. This topic is
revisited below.

4.3 Out-of-Time Validation (Subsequent to Development)

Given the longitudinal characteristics of the CCDB data set, we are able to track the
out-of-sample performance of our models through 2002. Table 14 shows the sample
sizes and bad rates for the development and out-of-time validation samples. On
a pooled-segment basis, overall default rates increase from 1999 to 2001, and then
decrease in 2002. However, there is a signi…cant improvement in the clean/thin
segment over those years. Table 15 shows the median scores by segment, model
form, and variable selection method for the development and out-of-time validation
samples. To better illustrate the shift in the distributions over time, we report the
box and whisker plot for the RAD Scores, by model form in Figure 7, and by segment
and model form in Figures 8 through 11. There is an obvious upward shift in the full
score distribution over time for all model forms in Figure 7, which re‡ects the general
trend in the median values reported in Table 15. Although the score distributions for
the dirty/current and dirty/delinquent segments are shifting down over time (Figures
8 and 9), the overwhelming shift up in the distributions for clean/thick and clean/thin
(Figures 10 and 11) dominate the overall shift in the distribution of scores.
We update the model separation and accuracy measures reported in Table 10 for
the out-of-sample periods 2000-2002 in Table 16 on a pooled-across-segments basis.
We observe that the H-L measures become very large (and the p-values very small)

23
in the out-of-time validation samples, indicating general lack of statistical …t for
predictive purposes. None of the scoring models developed using conventional indus-
try practices generated accurate predictions over time even though all the models
maintained their ability to di¤erentiate between good and bad accounts. These con-
clusions are supported by the out-of-sample results in Table 17. For each segment,
the K-S values remained relative constant, or improved, over time; however, in all
cases, the H-L statistics increased signi…cantly. The signi…cant increase in the H-L
values across all model segments in Table 17 suggests that our simple cross-section
model is under-speci…ed relative to the factors that re‡ect changes in the economic
environment over time.
As an additional test of the non-parametric approach, we reran the CHAID model
with continuous variables discretized to 200 values, and compared the performance
to the CHAID model based on quartiles. The CHAID based on all variables did
substantially worse in terms of model accuracy in the out-of-time validation samples.
The CHAID based on the Intersection selection performed about the same with 200
values as with quartiles for the 2000 and 2001 samples but substantially worse in
2002 in terms of model accuracy. Thus, there seems to be no real bene…t from
adding splits beyond quartiles for our continuous variables.
Figure 12 compares the empirical log-odds by di¤erent RAD scores for the 2002
validation samples. The plot clearly shows a deterioration in the predicted default
rate over the range 650-750. The actual performance is worse than the predicted,
and the RAD scores underestimate the default rates. The results for other years
were very similar to 2002, and are not shown here.
Overall, out-of-sample analyses show that the separation power of the models is
relatively stable over time; however, model accuracy decreases substantially. This
result, combined with the observed increase in the average default rate over the
full sample period except for the clean/thin segment (Table 14), implies that the
models estimated on a cross section of data from 1999 will underpredict defaults
over future periods. Moreover, it suggests that when the defaults are disaggregated
into buckets, the higher-default buckets will tend to be underpredicted more than
the low default buckets a result observed in Figure 12. These results imply that

24
models aimed at accuracy should be frequently updated, or that dynamic models,
with some dependence on macroeconomic conditions, should be considered.
Figure 13 compares the Gains chart for each of the RAD scoring models using the
2002 validation samples. Other years showed very similar results. As in the develop-
ment samples, the parametric and semiparametric models, and the GBS performed
very similarly, and the CHAID models were worse than the others. Although the
Gains charts for all parametric and semiparametric models are nearly overlapping,
the Stepwise selection method produces models that discriminate slightly better (for
both parametric and semiparametric forms). The Resampling selection method is
nearly as good, followed by the Intersection method.
We compare the Gains charts for the development samples and the validation
samples for each of the “preferred”models (the Resampling-based parametric model,
the Stepwise-based semiparametric model, the CHAID with all variables) and the
calibrated GBS in Figures 14 though 17. For all models and the GBS, the Gains
charts are again nearly overlapping and support the general results of the comparison
of K-S values over time.

4.4 Robustness of separation

As noted above, useful credit scores should be informative about di¤erent credit
related events. Although the RAD scores are developed for the event of 90+DPD
within 24 months, it is expected that they will be relevant for reasonable changes
in the outcome de…nition. As with the within-period validation above, we consider
di¤erent horizons (6, 12, and 18 months) as well as a di¤erent delinquency de…ni-
tion (60+ DPD). If the RAD scores generate reasonable separation for these other
events, we consider them to be robust in terms of separation. For some individuals,
performance data was missing over sub-portions of the 24-month observation pe-
riod. If performance information was missing as of the observation month (e.g., 6th,
12th, or 18th month), the observation was labeled missing in Table 12. As a result,
we excluded individuals with missing observations in shorter horizons in calculating
separation measures.

25
 The results in Table 18 show that the K-S measures for di¤erent de…nitions
of default are relatively consistent over time under the alternative event horizons.
Although the models perform better under a 90+ DPD de…nition of default, they
perform reasonably well under a 60+ DPD de…nition. If we compare across models,
parametric and semiparametric models showed the best separation being slightly
better than the calibrated GBS. The CHAID model consistently performs slightly
worse at separating good from bad accounts. These results show that the RAD scores
are very robust and informative in the separation metric for the delinquency events
we considered.

5 Conclusion
We developed credit-scoring models for bankcard performance using the OCC Risk
Analysis Division consumer credit data base and methods that are often encountered
in the industry. We validated and compared a parametric model, a semiparametric
model, and a popular nonparametric approach (CHAID).
It is worth pointing out that data preparation is crucial. The sample design
issues are important, as discussed, but simple matters such as variable de…nition and
treatment of missing or ambiguous data become critical. This is especially true in
cases where similar credit attributes could be calculated in slightly di¤erent ways.
Evaluating these data issues was one of the most time-consuming components of the
project.
With the data in hand, we …nd that careful statistical analysis will deliver a useful
model, and that, while there are di¤erences across methods, the di¤erences are small.
The parametric and semiparametric models appear to work slightly better than the
CHAID. There is little di¤erence between the parametric and semiparametric models.
We …nd that within-period validation is useful, but out-of-time validation shows
a substantial loss of accuracy. We attribute this to the changing macroeconomic
conditions. These conditions led to a small change in the overall default rate. This
change re‡ects much larger changes in the default rates of the high-default (low-score)
components of the population. This raises robustness issues in default prediction.

26
A practical conclusion is that accurate out-of-time prediction of within-score-group
default rates should be based on models which are frequently updated. The longer-
term response is to develop models which have variables re‡ecting aggregate credit
conditions. On the positive side, the separation properties of the models seem quite
robust in the out-of-time validation samples. This suggests that it is easier to rank
individuals by creditworthiness than to predict actual default rates.
There are many additional models in each of the categories, parametric, semi-
parametric and nonparametric, which could be considered. We have taken a rep-
resentative approach from each category. Our models are similar to those used in
practice. Our results suggest that the performance of models developed using simple
cross sectional techniques may be unreliable in terms of accuracy as macroeconomic
conditions change. The results suggest that increased attention be placed on the
use of longitudinal modeling methods as a means by which to estimate performance
conditional on temporally varying economic factors.

References
Bierman, H., and W. H. Hausman (1970): “The Credit Granting Decision,”
Management Sci., 16, 519–532.

Bucks, B. K., A. B. Kennickell, and K. B. Moore (2006): “Recent Changes

in U.S. Family Finances: Evidence from the 2001 and 2004 Survey of Consumer
Finances,”Federal Reserve Bulletin, 92, A1–A38.

Dirickx, Y. M. I., and L. Wakeman (1976): “An Extension of the Bierman-

Hausman Model for Credit Granting,”Management Sci, 22, 1229–1237.

Fair-Isaac (2006): “Understanding Your Credit Score,”

http://www.my…co.com/CreditEducation/WhatsInYourScore.aspx.

FRB (2006): Statistical Release G. 19. Board of Governers of the Federal Reserve
System.

27
Glennon, D. C. (1998): “Issues in Model Design and Validation,” in Credit Risk
Modeling Design and Application, ed. by E. Mays, chap. 13, pp. 207–221. Glenlake
Publishing.

Hand, D. J. (1997): Construction and Assessment of Classi…cation Rules. John

Wiley, Chichester, U.K.

Kass, G. (1980): “An exploratory technique for investigating large quantities of

categorical data,”Applied statistics, 29:2, 119–127.

Kiefer, N. M., and C. E. Larson (2006): “Speci…cation and Informational Issues

in Credit Scoring,” International Journal of Statistics and Management Systems,
1, forthcoming.

SAS (1995): “%Treedisc Macro for CHAID (Chi-Squared Automatic Interaction

Detection) Algorithm,” Discussion paper, Cary, NC: SAS Institute Inc. August
2005 Revision.

Srinivasan, V., and Y. H. Kim (1987): “The Bierman-Hausman Credit Granting

Model: A Note,”Management Sci, 33, 1361–1362.

Thomas, L., J. Crook, and D. B. Edelman (1992): Credit Scoring and Credit
Control. Oxford University Press, Oxford.

Thomas, L. C., D. B. Edelman, and J. Crook (2002): Credit Scoring and Its
Applications. SIAM.

28
TABLES AND FIGURES

29
Table 1:
CCDB Sampling Design Counts
A B C D E F G H H

Random
Matched Random
Scoreable CCDB
from Unmatched Unscoreable Total Current
Cross Panel
Previous from Cross Updating Year Masterfile
Section Updating
Year Year's Previous Section Random Extracts
From Proportion = B+E+G
Panel Year's Panel From Sample
Current
and and Dropped Current Year = B+C+D+E
Year
Scoreable Masterfile
Masterfile

1999 N.A. N.A. 950,000 50,000 100% 950000 1,000,000 1,000,000

2000 949,790 50,210 950,000 50,000 7% 66500 1,066,290 2,000,000
2001 1,015,469 50,821 883,710 50,000 7% 61860 1,127,329 2,000,000
2002 1,075,669 51,660 822,671 50,000 7% 57587 1,183,256 2,000,000
2003 1,130,033 53,223 766,744 50,000 5% 38337 1,218,370 2,000,000
2004 1,162,722 55,648 731,630 50,000 5% 36582 1,249,304 2,000,000

30
Table 2:
Transition to scoreable and unscoreable states
Individuals Individuals
Transition Base Year Panel Transitioning to Transitioning to Net Transitions
Period Size Unscoreable from Scoreable from to Scoreable
Scoreable Unscoreable
1999~2000 1,000,000 17,129 17,339 210
2000~2001 1,066,290 14,150 14,971 821
2001~2002 1,127,329 11,988 13,648 1,660
2002~2003 1,183,256 9,846 13,069 3,223
2003~2004 1,218,370 8,837 14,485 5,648

31
Table 3:
Variables by Type

Payment History Name

Worst status of open bankcards within 6 months CURR
Total number of tradelines with 90 days past due or worse BAD01
Dummy variable for the existence of tradelines with 90 days past due or worse BAD11
Total number of tradelines with 60 days past due or worse BAD21
Dummy variable for the existence of tradelines with 60 days past due or worse BAD31
Total number of tradelines with 30 days past due or worse BAD41
Dummy variable for the existence of tradelines with 30 days past due or worse BAD51
Total number of bankcard tradelines with 90 days past due or worse within 12 months BK03
Dummy variable for the existence of bankcard tradelines with 90 days past due or worse within 12 months BK13
Dummy variable for the existence of installment tradelines with 90 days past due or worse within 12 months IN13
Dummy variable for the existence of mortgage tradelines with 90 days past due or worse within 12 months MG13
Dummy variable for the existence of retail tradelines with 90 days past due or worse within 12 months RT13
Dummy variable for the existence of revolving retail tradelines with 90 days past due or worse within 12 months RTR13
Dummy variable for the existence of auto lease tradelines with 90 days past due or worse within 12 months AS13
Dummy variable for the existence of auto loan tradelines with 90 days past due or worse within 12 months AL13
Dummy variable for the existence of revolving tradelines with 90 days past due or worse within 12 months RV13
Months since the most recent 60 days past due or worse in bankcard tradelines of which the records were updated within
BK33
12 months
Worst status of bankcard tradelines with 60 days past due or worse and of which the records were updated within 12
BK43
months
Maximum of the balance amount, past due amount, and charged off amount of delinquent bankcard tradelines with 60 days
BK53
past due or worse and of which the records were updated within 12 months
Total number of public records in the DB PU01
Dummy variable for the existence of public records PU11
Total number of tradelines with good standing, positive balance, and of which the records were updated within 12 months GO01
Total number of closed tradelines within 12 months NUM_Closed

Amounts Owed
Aggregate credit amount of bankcard tradelines of which the records were updated within 12 months BK27
Aggregate credit amount of installment tradelines of which the records were updated within 12 months IN27
Aggregate credit amount of mortgage tradelines of which the records were updated within 12 months MG27
Aggregate credit amount of auto loan tradelines of which the records were updated within 12 months AL27
Aggregate credit amount of revolving tradelines of which the records were updated within 12 months RV27
Dummy variable for the positive aggregate credit amount of bankcard tradelines U11
Dummy variable for the positive aggregate credit amount of installment tradelines U12
Dummy variable for the positive aggregate credit amount of mortgage tradelines U13
Dummy variable for the positive aggregate credit amount of auto loan tradelines U17
Dummy variable for the positive aggregate credit amount of revolving tradelines U18
Aggregate balance amount of open bankcard tradelines of which the records were updated within 12 months ABK16
Aggregate balance amount of installment tradelines of which the records were updated within 12 months IN16
Aggregate balance amount of mortgage tradelines of which the records were updated within 12 months MG16
Aggregate balance amount of auto loan tradelines of which the records were updated within 12 months AL16
Aggregate balance amount of finance tradelines of which the records were updated within 12 months ALN08
Aggregate balance amount of retail tradelines of which the records were updated within 12 months ART08
Aggregate balance amount of revolving tradelines of which the records were updated within 12 months RV16
Aggregate balance amount of open home equity tradelines of which the records were updated within 12 months AEQ08
Bankcard utilization rate (Aggregate balance / Aggregate credit amount) BK28
Dummy variable for zero bankcard utilization rate BK28_0

32
Dummy variable for bankcard utilization rate=100% BK28_100
Dummy variable for bankcard utilization rate>100% BK28_101
Installment accounts utilization rate (Aggregate balance / Aggregate credit amount) IN28
Mortgage accounts utilization rate (Aggregate balance / Aggregate credit amount) MG28
Auto loan accounts utilization rate (Aggregate balance / Aggregate credit amount) AL28
Open bankcard utilization rate (Aggregate balance / Aggregate credit amount) ABK18
Revolving accounts utilization rate (Aggregate balance / Aggregate credit amount) RV28
Average credit amount of bankcard tradelines with positive balance and of which the records were updated within 12
BK17
months
Average credit amount of installment tradelines with positive balance and of which the records were updated within 12
IN17
months
Average credit amount of mortgage tradelines with positive balance and of which the records were updated within 12
MG17
months
Average credit amount of retail tradelines with positive balance and of which the records were updated within 12 months RT17
Average credit amount of auto loan tradelines with positive balance and of which the records were updated within 12
AL17
months
Average credit amount of revolving tradelines with positive balance and of which the records were updated within 12 months RV17

Length of credit history

Age of the oldest tradeline (Months) AG04
Age of the oldest bankcard tradeline (Months) BK04
Age of the oldest installment tradeline (Months) IN04
Age of the oldest mortgage tradeline (Months) MG04

New credit
Total number of inquiries within 6 months AIQ01
Total number of inquiries within 12 months IQ12
Total number of bankcard accounts opened within 2 years BK61
Total number of installment accounts opened within 2 years IN61
Total number of mortgage accounts opened within 2 years MG61
Dummy variable for the existence of new accounts within 2 years NUM71

Types of credit in Use

Dummy variable for the existence of installment tradelines within 12 months D1
Dummy variable for the existence of mortgage tradelines within 12 months D2
Dummy variable for the existence of retail tradelines within 12 months D3
Dummy variable for the existence of revolving retail tradelines within 12 months D4
Dummy variable for the existence of auto lease tradelines within 12 months D5
Dummy variable for the existence of auto loan tradelines within 12 months D6
Total number of credit tradelines (excluding inquiries/public records) NUM01
Total number of bankcard tradelines BK01
Total number of installment tradelines IN01
Total number of mortgage tradelines MG01
Total number of retail tradelines RT01
Total number of revolving retail tradelines RTR01
Total number of auto lease tradelines AS01
Total number of auto loan tradelines AL01
Total number of revolving tradelines RV01
Total number of credit tradelines of which the records were updated within 12 months NUM21
Total number of bankcard tradelines of which the records were updated within 12 months BK21
Total number of installment tradelines of which the records were updated within 12 months IN21
Total number of mortgage tradelines of which the records were updated within 12 months MG21

33
Total number of retail tradelines of which the records were updated within 12 months RT21
Total number of revolving retail tradelines of which the records were updated within 12 months RTR21
Total number of auto lease tradelines of which the records were updated within 12 months AS21
Total number of auto loan tradelines of which the records were updated within 12 months AL21
Total number of revolving tradelines of which the records were updated within 12 months RV21
Total number of bankcard tradelines with positive balance and of which the records were updated within 12 months BK31
Total number of installment tradelines with positive balance and of which the records were updated within 12 months IN31
Total number of mortgage tradelines with positive balance and of which the records were updated within 12 months MG31
Total number of retail tradelines with positive balance and of which the records were updated within 12 months RT31
Total number of auto loan tradelines with positive balance and of which the records were updated within 12 months AL31
Total number of revolving tradelines with positive balance and of which the records were updated within 12 months RV31

34
Table 4: Dirty/Delinquent Segment - Explanatory variables selected using Stepwise, Resample, and Intersection methods

Significance Ranking
Variables selected using the Stepwise method Variable in Selection Methods
(sorted by variable type) Names Inter-
Resample Stepwise
section

I. Payment History
Total number of tradelines with good standing, positive balance, and of which the records were updated
GO01 1 1 1
within 12 months
Worst status of open bankcards within 6 months CURR 3 3 3
Total number of closed tradelines within 12 months NUM_Closed 10 6
Dummy variable for the existence of tradelines with 90 days past due or worse BAD11 9 13
Worst status of bankcard tradelines with 60 days past due or worse and of which the records were updated
BK43 15
within 12 months
Maximum of the balance amount, past due amount, and charged off amount of delinquent bankcard
BK53 29
tradelines with 60 days past due or worse and of which the records were updated within 12 months
Dummy variable for the existence of installment tradelines with 90 days past due or worse within 12 months IN13 31

II. Amount Owed

Open bankcard utilization rate (Aggregate balance / Aggregate credit amount) ABK18 4 4 4
Dummy variable for bankcard utilization rate>100% BK28_101 7 8 7
Aggregate balance amount of retail tradelines of which the records were updated within 12 months ART08 9
Aggregate balance amount of revolving tradelines of which the records were updated within 12 months RV16 12
Average credit amount of bankcard tradelines with positive balance and of which the records were updated
BK17 13 19
within 12 months
Auto loan accounts utilization rate (Aggregate balance / Aggregate credit amount) AL28 21
Aggregate balance amount of finance tradelines of which the records were updated within 12 months ALN08 23
Average credit amount of revolving tradelines with positive balance and of which the records were updated
RV17 25
within 12 months
Dummy variable for the positive aggregate credit amount of auto loan tradelines U17 26
Average credit amount of mortgage tradelines with positive balance and of which the records were updated
MG17 27
within 12 months
Dummy variable for bankcard utilization rate=100% BK28_100 30

III. Length of Credit History

Age of the oldest bankcard tradeline (Months) BK04 8 7 8
Age of the oldest installment tradeline (Months) IN04 17

IV. New Credit

Total number of bankcard accounts opened within 2 years BK61 6 6 5

35
Total number of inquiries within 12 months IQ12 11 10
Dummy variable for the existence of new accounts within 2 years NUM71 14 18
Total number of installment accounts opened within 2 years IN61 12 20

V. Type of Credit Used

Total number of revolving tradelines with positive balance and of which the records were updated within 12
RV31 2 2 2
months
Total number of installment tradelines with positive balance and of which the records were updated within
IN31 5 5 11
12 months
Total number of credit tradelines of which the records were updated within 12 months NUM21 14
Total number of mortgage tradelines with positive balance and of which the records were updated within 12
MG31 16
months
Total number of installment tradelines of which the records were updated within 12 months IN21 22
Dummy variable for the existence of revolving retail tradelines within 12 months D4 24
Dummy variable for the existence of retail tradelines within 12 months D3 28

36
Table 5: Dirty/Current Segment - Explanatory variables selected using Stepwise, Resample, and Intersection methods

Significance Ranking
Variables selected using the Stepwise method Variable in Selection Methods
(sorted by variable type) Names Inter-
Resample Stepwise
section

I. Payment History
Total number of tradelines with good standing, positive balance, and of which the records were updated
GO01 3 4 3
within 12 months
Total number of tradelines with 30 days past due or worse BAD41 4 5 4
Total number of tradelines with 90 days past due or worse BAD01 7 9 7
Total number of closed tradelines within 12 months NUM_Closed 12 10 9
Total number of bankcard tradelines with 90 days past due or worse within 12 months BK03 25 13
Dummy variable for the existence of mortgage tradelines with 90 days past due or worse within 12 months MG13 18 21 16
Dummy variable for the existence of installment tradelines with 90 days past due or worse within 12 months IN13 13 13 18
Dummy variable for the existence of bankcard tradelines with 90 days past due or worse within 12 months BK13 24 20
Dummy variable for the existence of retail tradelines with 90 days past due or worse within 12 months RT13 28
Dummy variable for the existence of revolving tradelines with 90 days past due or worse within 12 months RV13 20 11 34
Dummy variable for the existence of revolving retail tradelines with 90 days past due or worse within 12
RTR13 36
months
Dummy variable for the existence of auto loan tradelines with 90 days past due or worse within 12 months AL13 38
Dummy variable for the existence of tradelines with 90 days past due or worse BAD11 43
Maximum of the balance amount, past due amount, and charged off amount of delinquent bankcard
BK53 44
tradelines with 60 days past due or worse and of which the records were updated within 12 months

II. Amount Owed

Open bankcard utilization rate (Aggregate balance / Aggregate credit amount) ABK18 1 2 1
Dummy variable for bankcard utilization rate>100% BK28_101 5 6 6
Average credit amount of bankcard tradelines with positive balance and of which the records were updated
BK17 9 8 10
within 12 months
Average credit amount of mortgage tradelines with positive balance and of which the records were updated
MG17 26 20 14
within 12 months
Aggregate credit amount of bankcard tradelines of which the records were updated within 12 months BK27 14 15 17
Aggregate balance amount of open bankcard tradelines of which the records were updated within 12 months ABK16 22 19 25
Average credit amount of revolving tradelines with positive balance and of which the records were updated
RV17 31
within 12 months
Aggregate balance amount of revolving tradelines of which the records were updated within 12 months RV16 37
Average credit amount of installment tradelines with positive balance and of which the records were updated
IN17 40
within 12 months
Aggregate balance amount of open home equity tradelines of which the records were updated within 12
AEQ08 42
months
Dummy variable for the positive aggregate credit amount of revolving tradelines U18 45

37
Aggregate balance amount of installment tradelines of which the records were updated within 12 months IN16 46
Mortgage accounts utilization rate (Aggregate balance / Aggregate credit amount) MG28 47
Dummy variable for the positive aggregate credit amount of installment tradelines U12 48

III. Length of Credit History

Age of the oldest bankcard tradeline (Months) BK04 6 8
Age of the oldest mortgage tradeline (Months) MG04 21 22
Age of the oldest installment tradeline (Months) IN04 19 14 23
Age of the oldest tradeline (Months) AG04 32

IV. New Credit

Total number of bankcard accounts opened within 2 years BK61 8 3 5
Total number of installment accounts opened within 2 years IN61 16 16 11
Dummy variable for the existence of new accounts within 2 years NUM71 23 18 21
Total number of inquiries within 6 months AIQ01 39
Total number of inquiries within 12 months IQ12 11 41

V. Type of Credit Used

Total number of revolving tradelines with positive balance and of which the records were updated within 12
RV31 2 1 2
months
Dummy variable for the existence of revolving retail tradelines within 12 months D4 17 17 12
Total number of revolving tradelines RV01 10 7 15
Total number of installment tradelines of which the records were updated within 12 months IN21 15 12 19
Total number of installment tradelines IN01 24
Total number of credit tradelines (excluding inquiries/public records) NUM01 26
Total number of bankcard tradelines with positive balance and of which the records were updated within 12
BK31 27 27
months
Total number of installment tradelines with positive balance and of which the records were updated within 12
IN31 29
months
Total number of mortgage tradelines of which the records were updated within 12 months MG21 30
Total number of mortgage tradelines MG01 33
Dummy variable for the existence of installment tradelines within 12 months D1 35

38
Table 6: Clean/Thin Segment - Explanatory variables selected using Stepwise, Resample, and Intersection methods

Significance Ranking
Variables selected using the Stepwise method Variable in Selection Methods
(sorted by variable type) Names Inter
Resample Stepwise
section

I. Payment History
Worst status of open bankcards within 6 months CURR 2 2 4
Total number of tradelines with 30 days past due or worse BAD41 10
Total number of tradelines with good standing, positive balance, and of which the records were updated
GO01 11
within 12 months

II. Amount Owed

Revolving accounts utilization rate (Aggregate balance / Aggregate credit amount) RV28 1 1 1
Dummy variable for bankcard utilization rate=100% BK28_100 6 4 7
Dummy variable for bankcard utilization rate>100% BK28_101 7 8
Open bankcard utilization rate (Aggregate balance / Aggregate credit amount) ABK18 12

III. Length of Credit History

Age of the oldest bankcard tradeline (Months) BK04 5 9

IV. New Credit

Total number of bankcard accounts opened within 2 years BK61 4 2
Total number of inquiries within 12 months IQ12 3 3 3

V. Type of Credit Used

Total number of bankcard tradelines BK01 5
Total number of bankcard tradelines with positive balance and of which the records were updated within 12
BK31 6
months

39
Table 7: Clean/Thick Segment - Explanatory variables selected using Stepwise, Resample, and Intersection methods

Significance Ranking
Variables selected using the Stepwise method Variable in Selection Methods
(sorted by variable type) Names Inter-
Resample Stepwise
section

I. Payment History
Total number of tradelines with 30 days past due or worse BAD41 33
Dummy variable for the existence of tradelines with 30 days past due or worse BAD51 3 5 5
Worst status of open bankcards within 6 months CURR 2 3 3
Total number of tradelines with good standing, positive balance, and of which the records were updated
GO01 6 4
within 12 months
Total number of closed tradelines within 12 months NUM_Closed 19

II. Amount Owed

Aggregate balance amount of open bankcard tradelines of which the records were updated within 12
ABK16 11 10
months
Open bankcard utilization rate (Aggregate balance / Aggregate credit amount) ABK18 13 8 8
Auto loan accounts utilization rate (Aggregate balance / Aggregate credit amount) AL28 36
Aggregate balance amount of finance tradelines of which the records were updated within 12 months ALN08 24
Aggregate balance amount of retail tradelines of which the records were updated within 12 months ART08 27
Average credit amount of bankcard tradelines with positive balance and of which the records were updated
BK17 12 12
within 12 months
Bankcard utilization rate (Aggregate balance / Aggregate credit amount) BK28 35
Dummy variable for bankcard utilization rate=100% BK28_100 14 13
Dummy variable for bankcard utilization rate>100% BK28_101 19 25
Aggregate balance amount of mortgage tradelines of which the records were updated within 12 months MG16 34
Average credit amount of mortgage tradelines with positive balance and of which the records were updated
MG17 15 6 23
within 12 months
Aggregate credit amount of mortgage tradelines of which the records were updated within 12 months MG27 22
Aggregate balance amount of revolving tradelines of which the records were updated within 12 months RV16 20 21
Average credit amount of revolving tradelines with positive balance and of which the records were updated
RV17 29
within 12 months
Revolving accounts utilization rate (Aggregate balance / Aggregate credit amount) RV28 1 2 1

III. Length of Credit History

Age of the oldest tradeline (Months) AG04 16 9 16
Age of the oldest bankcard tradeline (Months) BK04 8 10 9

IV. New Credit

40
Total number of inquiries within 6 months AIQ01 31
Total number of bankcard accounts opened within 2 years BK61 7 7 6
Total number of installment accounts opened within 2 years IN61 10 11
Total number of inquiries within 12 months IQ12 5 4 17
Total number of mortgage accounts opened within 2 years MG61 32
Dummy variable for the existence of new accounts within 2 years NUM71 37

V. Type of Credit Used

Total number of auto loan tradelines AL01 28
Total number of auto loan tradelines with positive balance and of which the records were updated within 12
AL31 20
months
Total number of bankcard tradelines with positive balance and of which the records were updated within 12
BK31 17 30
months
Total number of installment tradelines of which the records were updated within 12 months IN21 18 14
Total number of installment tradelines with positive balance and of which the records were updated within
IN31 9 7
12 months
Total number of credit tradelines (excluding inquiries/public records) NUM01 26
Total number of credit tradelines of which the records were updated within 12 months NUM21 15
Total number of revolving tradelines of which the records were updated within 12 months RV21 18
Total number of revolving tradelines with positive balance and of which the records were updated within 12
RV31 4 1 2
months

41
Table 8: Variables used at least once in CHAID splitting

Variable Names All Inter-

Variables
(sorted by variable type) Attributes section

I. Payment History
Total number of tradelines with 60 days past due or worse BAD21 X X
Dummy variable for the existence of tradelines with 60 days past due or worse BAD31 X
Total number of tradelines with 30 days past due or worse BAD41 X X
Dummy variable for the existence of tradelines with 30 days past due or worse BAD51 X X
Worst status of open bankcards within 6 months CURR X X
Total number of tradelines with good standing, positive balance, and of which the records were updated within 12
GO01 X X
months

II. Amount Owed

Aggregate balance amount of open bankcard tradelines of which the records were updated within 12 months ABK16 X X
Open bankcard utilization rate (Aggregate balance / Aggregate credit amount) ABK18 X X
Average credit amount of bankcard tradelines with positive balance and of which the records were updated within 12
BK17 X X
months
Aggregate credit amount of bankcard tradelines of which the records were updated within 12 months BK27 X
Bankcard utilization rate (Aggregate balance / Aggregate credit amount) BK28 X
Dummy variable for bankcard utilization rate=100% BK28_100 X
Installment accounts utilization rate (Aggregate balance / Aggregate credit amount) IN28 X
Average credit amount of mortgage tradelines with positive balance and of which the records were updated within 12
MG17 X X
months
Average credit amount of retail tradelines with positive balance and of which the records were updated within 12
RT17 X
months
Aggregate balance amount of revolving tradelines of which the records were updated within 12 months RV16 X
Average credit amount of revolving tradelines with positive balance and of which the records were updated within 12
RV17 X
months
Aggregate credit amount of revolving tradelines of which the records were updated within 12 months RV27 X
Revolving accounts utilization rate (Aggregate balance / Aggregate credit amount) RV28 X X
Dummy variable for the positive aggregate credit amount of installment tradelines U12 X
Dummy variable for the positive aggregate credit amount of mortgage tradelines U13 X

III. Length of Credit History

Age of the oldest tradeline (Months) AG04 X X
Age of the oldest bankcard tradeline (Months) BK04 X X
Age of the oldest installment tradeline (Months) IN04 X X
Age of the oldest mortgage tradeline (Months) MG04 X

42
IV. New Credit
IQ06 X
Total number of bankcard accounts opened within 2 years BK61 X X
Total number of installment accounts opened within 2 years IN61 X
Total number of inquiries within 12 months IQ12 X X
Dummy variable for the existence of new accounts within 2 years NUM71 X

V. Type of Credit Used

Total number of bankcard tradelines of which the records were updated within 12 months BK21 X
Total number of bankcard tradelines with positive balance and of which the records were updated within 12 months BK31 X
Dummy variable for the existence of installment tradelines within 12 months D1 X
Dummy variable for the existence of retail tradelines within 12 months D3 X
Dummy variable for the existence of revolving retail tradelines within 12 months D4 X X
Total number of installment tradelines of which the records were updated within 12 months IN21 X
Total number of installment tradelines with positive balance and of which the records were updated within 12 months IN31 X
Total number of mortgage tradelines MG01 X
Total number of credit tradelines (excluding inquiries/public records) NUM01 X X
Total number of retail tradelines with positive balance and of which the records were updated within 12 months RT31 X
Total number of revolving tradelines RV01 X X
Total number of revolving tradelines of which the records were updated within 12 months RV21 X
Total number of revolving tradelines with positive balance and of which the records were updated within 12 months RV31 X X

43
Table 9:
In-Time Validation: Median Scores for Various Models at Development

Scoring Model Sample

Segment Variable 1999 1999
Model Form
Selection Dev Hold-Out
Stepwise 616 616
Parametric Resampling 616 616
Intersection 617 617
Dirty History and Presently Mildly
Delinquent Stepwise 598 598
Semi Parametric Resampling 599 599
Intersection 600 600
Calibrated Generic Bureau Score 636 637
Stepwise 676 676
Parametric Resampling 676 676
Intersection 676 676
Dirty History and
Presently Current Stepwise 673 673
Semi Parametric Resampling 672 672
Intersection 672 672
Calibrated Generic Bureau Score 675 675
Stepwise 729 730
Parametric Resampling 727 727
Intersection 735 735
Clean History and Thin File Stepwise 733 733
Semi Parametric Resampling 727 727
Intersection 735 735
Calibrated Generic Bureau Score 722 723
Stepwise 750 750
Parametric Resampling 751 751
Intersection 750 750
Clean History and Thick File Stepwise 759 759
Semi Parametric Resampling 758 758
Intersection 760 760
Calibrated Generic Bureau Score 747 747
Stepwise 734 734
Parametric Resampling 735 735
Intersection 734 734
Stepwise 738 738
All (Pooled) Semi Parametric Resampling 737 737
Intersection 735 735

Nonparametric All variables 725 725

(CHAID) Intersection 724 724
Calibrated Generic Bureau Score 734 734

44
Table 10:
In-Time Validation: Model Separation and Accuracy Measures at Development (Pooled Across Segments)
(Bad = 90+Days Past Due, or Worse, over the Following 24 Months)

Statistic and Sample

Scoring Model
Kolmogorov-Smirnov (K-S) Hosmer-Lemeshow (H-L)
Variable
Model Form 1999 Dev 1999 Hold-out 1999 Dev 1999 Hold-out
Selection

(value) (p-value)2 (value) (p-value)

Stepwise 64.0 64.0 74.2 (<.0001) 73.7 (<.0001)

Parametric Resampling 63.7 63.8 70.0 (<.0001) 70.2 (<.0001)

Intersection 62.7 62.9 69.3 (<.0001) 72.6 (<.0001)

Stepwise 64.0 63.9 NA1 NA 6.8 .5584

Semi Parametric Resampling 63.8 63.7 NA NA 8.0 .4335

Intersection 62.6 62.8 NA NA 16.7 .0334

All Variables 58.3 57.9 2.1 .9778 360.4 (<.0001)

NonParametric
(CHAID)
Intersection 59.2 58.9 44.8 (<.0001) 145.9 (<.0001)

Calibrated Generic Bureau Score 62.4 62.6 194.1 (<.0001) 2506.3 (<.0001)

1. The H-L test does not apply. By design, the predicted outcomes under the semi parametric approach are equal to the actual outcomes.
2. The p-values are derived under null hypothesis H0: pj=qj for all j (see equation 8) under the assumption that the H-L ~ c2 df=8.

45
Table 11:
In-Time Validation: Separation (K-S) and Accuracy (H-L) Measures for Parametric and Semi
Parametric Models at Development, by Segment
Statistic and Sample
Scoring Model Kolmogorov-
Segment Hosmer-Lemeshow
Smirnov
Variable 1999 1999
Model Form 1999 Dev 1999 Hold-out
Selection Dev Hold-out
(value) (p-value) (value) (p-value)
Stepwise 40.3 38.9 16.5 .0358 22.1 .0047
Dirty History Parametric Resampling 39.3 37.2 12.5 .1303 23.9 .0024
and Presently
Mildly Intersection 37.5 37.0 6.1 .6360 10.8 .2133
Delinquent Stepwise 40.3 38.5 489.8 (<.0001) 548.1 (<.0001)
Semi
1 Resampling 39.1 37.3 526.8 (<.0001) 565.6 (<.0001)
(n=13,302) Parametric
Intersection 39.1 37.3 592.7 (<.0001) 625.6 (<.0001)
Stepwise 42.9 43.2 30.8 .0002 26.9 .0007
Dirty History Parametric Resampling 42.4 43.0 26.0 .0011 27.3 .0006
and Presently Intersection 41.4 42.0 31.1 .0001 32.8 .0001
Current
Stepwise 43.0 43.2 34.7 (<.0001) 25.8 .0011
Semi
(n=67,814) Resampling 42.7 43.2 46.1 (<.0001) 42.5 (<.0001)
Parametric
Intersection 42.2 42.6 52.7 (<.0001) 35.8 (<.0001)
Stepwise 58.2 57.1 8.3 .4047 27.3 .0006
Parametric Resampling 57.4 56.1 11.5 .1750 16.7 .0334
Clean History Intersection 54.3 54.4 48.7 (<.0001) 51.7 (<.0001)
and Thin File
Stepwise 57.6 57.1 7.5 .4837 16.4 .0370
Semi
(n=15,132) Resampling 57.2 56.8 6.1 .6360 16.9 .0312
Parametric
Intersection 55.5 55.2 21.4 .0062 51.8 (<.0001))
Stepwise 60.2 60.1 84.9 (<.0001) 94.8 (<.0001)
Parametric Resampling 60.0 59.9 74.6 (<.0001) 74.5 (<.0001)
Clean History
and Thick File Intersection 58.7 58.9 67.6 (<.0001) 82.2 (<.0001)
Stepwise 60.1 60.2 9.7 .2867 15.4 .0518
(n=242,330) Semi Resampling 60.0 59.9 7.7 .4633 14.4 .0719
Parametric
Intersection 59.1 59.2 6.8 .5584 15.9 .0438

1. The predicted values were derived from the actual default rates in the decile range based on the pooled segment data in Table 10.

46
Table 12:
Empirical Bad Rates for Alternate Bad Definitions on the Development Samples

Bad Definition N Rate

Horizon 1
Event Bad Good Missing Bad Good
(Months)
24 24340 314238 0 7.19% 92.81%
90+ Days Past 18 19665 318757 156 5.81% 94.19%
Due or Worse 12 14516 323679 383 4.29% 95.71%
6 8019 329193 1366 2.38% 97.62%
24 30107 308471 0 8.89% 91.11%
60+ Days Past 18 25204 313218 156 7.45% 92.55%
Due or Worse 12 19478 318717 383 5.76% 94.24%
6 12192 325020 1366 3.62% 96.38%

1. Missing observations were generated if the lender failed to report performance as of the observation date 6, 12, or 18 months forward.

47
Table 13:
K-S separation measures for models built to alternate bad definitions on the development sample
Bad Event Type and Sample
Scoring Model
90+ Days Past Due or Worse 60+ Days Past Due or Worse
Bad
Variable
Model Form Event Dev Hold-Out Dev Hold-Out
Selection
Horizon
24 64.0 64.0 61.5 61.6
18 65.8 65.8 63.1 63.4
Stepwise
12 67.8 67.6 65.4 65.4
6 71.8 72.2 68.6 68.7
24 63.7 63.8 61.4 61.6
18 65.7 65.5 63.0 63.3
Parametric Resampling
12 67.7 67.5 65.3 65.3
6 71.5 72.0 68.4 68.8
24 62.7 62.9 60.5 60.9
18 64.6 64.8 62.2 62.6
Intersection
12 66.7 66.8 64.5 64.8
6 71.0 71.7 67.8 68.2
24 64.0 63.9 61.6 61.7
18 65.8 65.8 63.0 63.3
Stepwise
12 67.8 67.6 65.3 65.3
6 71.7 72.2 68.6 68.7
24 63.8 63.7 61.4 61.6
Semi 18 65.7 65.5 63.0 63.2
Parametric Resampling
12 67.6 67.5 65.2 65.3
6 71.6 72.1 68.4 68.8
24 62.6 62.8 60.5 60.9
18 64.6 64.8 62.1 62.5
Intersection
12 66.6 66.7 64.4 64.8
6 71.0 71.6 67.8 68.2
24 58.3 57.9 56.6 56.4
18 59.9 59.4 57.9 57.7
All variables
12 61.6 61.0 59.6 59.4
Non 6 65.0 64.9 62.3 62.2
Parametric
24 59.2 58.9 57.2 56.8
18 60.9 60.4 58.4 58.1
Intersection
12 62.7 62.0 60.1 59.6
6 65.8 65.7 62.8 62.0
24 62.4 62.6 60.3 60.4
Calibrated Generic Bureau 18 64.6 64.4 61.9 61.9
Score
12 66.5 66.4 63.9 63.8
6 70.2 70.7 67.2 66.7

48
Table 14:
Sample Sizes and Bad Rates For the Development and Out-of-Time Validation Samples
(Bad = 90 Days Past Due, or Worse, over the Following 24 Months)

Sample and Statistic

Segment 1999 2000 2001 2002
Development Validation Validation Validation
Size Bad Rate Size Bad Rate Size Bad Rate Size Bad Rate
Dirty History and
Presently Mildly 13,302 49.27% 29,252 56.25% 35,823 55.42% 39,523 53.79%
Delinquent
Dirty History and
67,814 14.60% 133,399 18.02% 163,647 17.36% 174,290 15.79%
Presently Current

Clean History and

15,132 4.76% 39,032 4.38% 26,732 3.97% 30,984 2.73%
Thin File

Clean History and

242,330 2.96% 549,635 3.61% 584,571 3.44% 611,924 3.27%
Thick File

All 338,578 7.19% 751,318 8.26% 810,773 8.56% 856,721 8.13%

49
Table 15:
Out-of-Time Validation: Median Scores for Various Models Across Validation Samples

Scoring Model Sample

Segment Variable 1999 2000 2001 2002
Model Form
Selection Dev Val Val Val
Stepwise 616 611 610 614
Parametric Resampling 616 612 611 615
Intersection 617 612 611 615
Dirty History and Presently Mildly
Delinquent Stepwise 598 598 598 598
Semi Parametric Resampling 599 599 599 599
Intersection 600 600 600 600
Calibrated Generic Bureau Score 636 631 631 631
Stepwise 676 669 672 675
Parametric Resampling 676 669 672 674
Intersection 676 668 671 673
Dirty History and
Presently Current Stepwise 673 666 673 673
Semi Parametric Resampling 672 666 672 672
Intersection 672 666 672 672
Calibrated Generic Bureau Score 675 670 673 673
Stepwise 729 749 754 759
Parametric Resampling 727 748 752 758
Intersection 735 739 741 743
Clean History and Thin File Stepwise 733 759 765 770
Semi Parametric Resampling 727 748 764 769
Intersection 735 741 741 751
Calibrated Generic Bureau Score 722 737 740 747
Stepwise 750 751 756 761
Parametric Resampling 751 754 758 763
Intersection 750 751 754 756
Clean History and Thick File Stepwise 759 759 770 772
Semi Parametric Resampling 758 764 769 771
Intersection 760 760 766 766
Calibrated Generic Bureau Score 747 751 754 756
Stepwise 734 737 740 745
Parametric Resampling 735 740 742 747
Intersection 734 737 738 740
Stepwise 738 738 740 746
All (Pooled) Semi Parametric Resampling 737 746 746 748
Intersection 735 735 735 741

Nonparametric All variables 725 730 729 733

(CHAID) Intersection 724 729 726 730
Calibrated Generic Bureau Score 734 739 740 743

50
Table 16:
Out-of-Time Validation: Model Separation and Accuracy Measures (Pooled Across Segments)

Statistic and Sample

Scoring Model
Kolmogorov-Smirnov Hosmer-Lemeshow
Variable 1999 Dev 2000 Val 2001 Val 2002 Val
Model Form 1999 Dev 2000 Val 2001 Val 2002 Val
Selection value p-value value p-value value p-value value p-value
Stepwise 64.0 65.4 66.1 65.8 74.2 (<.0001) 1294.8 (<.0001) 1773.3 (<.0001) 2144.7 (<.0001)
Parametric Resampling 63.7 65.2 65.9 65.5 70.0 (<.0001) 1709.3 (<.0001) 2073.8 (<.0001) 2307.4 (<.0001)
Intersection 62.7 65.0 65.7 65.2 69.3 (<.0001) 1185.3 (<.0001) 1408.5 (<.0001) 1050.4 (<.0001)
Stepwise 63.9 65.4 66.1 65.8 NA NA 1233.3 (<.0001) 1679.5 (<.0001) 2051.5 (<.0001)
Semi Parametric Resampling 63.7 65.3 66.1 65.7 NA NA 1402.1 (<.0001) 1713.2 (<.0001) 1915.8 (<.0001)
Intersection 62.6 65.1 65.9 65.4 NA NA 932.2 (<.0001) 1063.4 (<.0001) 649.6 (<.0001)

NonParametric All Variables 58.3 59.3 60.0 59.9 2.1 .9778 3962.4 (<.0001) 3231.5 (<.0001) 2792.4 (<.0001)
(CHAID) Intersection 59.2 60.6 60.9 60.9 44.8 (<.0001) 4163.6 (<.0001) 3758.9 (<.0001) 3053.0 (<.0001)
Calibrated Generic Bureau Score 62.4 64.9 65.6 65.1 194.1 (<.0001) 2506.3 (<.0001) 4267.6 (<.0001) 2970.0 (<.0001)

51
Table 17:
Out-of-Time Validation: Separation (K-S) and Accuracy (H-L) Measures for Parametric and Semi Parametric Models, by
Segment

Statistic and Sample

Scoring Model
Segment Kolmogorov-Smirnov Hosmer-Lemeshow
Model Variable 1999 2000 2001 2002
1999 Dev 2000 Val 2001 Val 2002 Val
Form Selection Dev Val Val Val
value p-value value p-value value p-value value p-value
Stepwise 40.3 40.2 40.7 40.7 16.5 .0358 349.7 (<.0001) 181.2 (<.0001) 500.9 (<.0001)
Parametric Resampling 39.3 39.7 39.8 40.5 12.5 .1303 419.3 (<.0001) 227.6 (<.0001) 552.6 (<.0001)
Dirty History and Presently Intersection 37.5 38.5 38.9 39.5 6.1 .6360 313.7 (<.0001) 169.0 (<.0001) 386.9 (<.0001)
Mildly Delinquent Stepwise 40.3 39.5 39.6 40.8 489.8 (<.0001) 1991.3 (<.0001) 2079.4 (<.0001) 2348.7 (<.0001)
Semi
Resampling 39.1 39.7 39.9 40.9 526.8 (<.0001) 2049.7 (<.0001) 2219.2 (<.0001) 2453.1 (<.0001)
Parametric
Intersection 39.1 39.7 39.7 40.7 592.7 (<.0001) 2277.3 (<.0001) 2499.5 (<.0001) 2464.0 (<.0001)
Stepwise 42.9 43.2 42.6 42.0 30.8 .0002 113.9 (<.0001) 306.0 (<.0001) 310.2 (<.0001)
Parametric Resampling 42.4 42.8 42.3 41.7 26.0 .0011 121.1 (<.0001) 264.1 (<.0001) 290.3 (<.0001)
Dirty History and Presently Intersection 41.4 41.8 41.3 41.0 31.1 .0001 127.9 (<.0001) 336.4 (<.0001) 275.4 (<.0001)
Current Stepwise 43.0 43.0 42.5 42.0 34.7 (<.0001) 137.1 (<.0001) 348.3 (<.0001) 510.5 (<.0001)
Semi
Resampling 42.7 42.9 42.5 41.9 46.1 (<.0001) 125.6 (<.0001) 276.3 (<.0001) 480.6 (<.0001)
Parametric
Intersection 42.2 42.5 41.9 41.3 52.7 (<.0001) 98.3 (<.0001) 230.3 (<.0001) 420.2 (<.0001)
Stepwise 58.2 67.7 70.6 68.9 8.3 .4047 142.7 (<.0001) 111.3 (<.0001) 51.7 (<.0001)
Parametric Resampling 57.4 67.2 70.0 68.9 11.5 .1749 157.9 (<.0001) 118.9 (<.0001) 73.2 (<.0001)
Intersection 54.3 64.7 66.0 68.8 48.7 (<.0001) 188.2 (<.0001) 118.7 (<.0001) 130.6 (<.0001)
Clean History and Thin
File Stepwise 57.6 67.5 70.6 69.6 7.5 .4838 132.0 (<.0001) 107.3 (<.0001) 36.9 (<.0001)
Semi
Parametric Resampling 57.2 67.7 69.8 68.9 6.1 .6360 137.0 (<.0001) 107.5 (<.0001) 45.9 (<.0001)
Intersection 55.5 65.1 67.5 68.9 21.4 .0062 168.2 (<.0001) 111.6 (<.0001) 104.2 (<.0001)
Stepwise 60.2 61.2 63.2 63.2 84.9 (<.0001) 1342.1 (<.0001) 1850.5 (<.0001) 2635.1 (<.0001)
Parametric Resampling 60.0 61.0 63.2 63.1 74.6 (<.0001) 1874.0 (<.0001) 2181.4 (<.0001) 2957.1 (<.0001)
Clean History and Thick Intersection 58.7 60.6 63.1 62.8 67.6 (<.0001) 1199.3 (<.0001) 1283.7 (<.0001) 1338.6 (<.0001)
File Stepwise 60.1 61.2 63.4 63.3 9.7 .2867 1085.7 (<.0001) 1668.9 (<.0001) 2640.3 (<.0001)
Semi
Parametric Resampling 60.0 61.1 63.3 63.3 7.7 .4633 1477.5 (<.0001) 1827.3 (<.0001) 2703.6 (<.0001)
Intersection 59.1 60.9 63.3 63.0 6.8 .5584 830.0 (<.0001) 933.3 (<.0001) 965.4 (<.0001)

52
Table 18:
Out-of-Time Validation: Separation (K-S) Measures for Different Definitions of Default

Bad Event Type and Sample

Scoring Model
90+ Days Past Due or Worse 60+ Days Past Due or Worse
Variable Bad Event Hold- Hold-
Model Form Dev 2000 2001 2002 Dev 2000 2001 2002
Selection Horizon Out Out
24 64.0 64.0 65.4 66.1 65.8 61.5 61.6 63.6 64.6 63.9
18 65.8 65.8 66.9 67.6 67.6 63.1 63.4 64.8 65.9 65.6
Stepwise
12 67.8 67.6 68.9 69.2 69.9 65.4 65.4 66.4 67.7 67.8
6 71.8 72.2 73.0 73.3 73.8 68.6 68.7 70.5 71.0 71.1
24 63.7 63.8 65.2 65.9 65.5 61.4 61.6 63.5 64.4 63.7
18 65.7 65.5 66.7 67.4 67.4 63.0 63.3 64.6 65.8 65.4
Parametric Resampling
12 67.7 67.5 68.8 69.1 69.6 65.3 65.3 66.3 67.5 67.6
6 71.5 72.0 72.9 73.1 73.5 68.4 68.8 70.3 71.0 71.0
24 62.7 62.9 65.0 65.7 65.2 60.5 60.9 63.2 64.1 63.6
18 64.6 64.8 66.6 67.1 67.1 62.2 62.6 64.5 65.4 65.2
Intersection
12 66.7 66.8 68.4 68.8 69.3 64.5 64.8 66.1 67.3 67.4
6 71.0 71.7 72.5 72.7 73.1 67.8 68.2 70.0 70.6 70.9
24 64.0 63.9 65.4 66.1 65.7 61.6 61.7 63.6 64.6 63.9
18 65.8 65.8 66.9 67.6 67.6 63.0 63.3 64.7 66.0 65.6
Stepwise
12 67.8 67.6 68.9 69.3 69.8 65.3 65.3 66.5 67.7 67.8
6 71.7 72.2 73.1 73.3 73.7 68.6 68.7 70.5 71.0 71.2
24 63.8 63.7 65.2 65.9 65.5 61.4 61.6 63.4 64.4 63.7
Semi 18 65.7 65.5 66.7 67.4 67.4 63.0 63.2 64.6 65.8 65.4
Resampling
Parametric 12 67.6 67.5 68.8 69.1 69.6 65.2 65.3 66.3 67.5 67.6
6 71.6 72.1 72.9 73.0 73.5 68.4 68.8 70.4 70.9 71.0
24 62.6 62.8 64.9 65.7 65.3 60.5 60.9 63.2 64.1 63.6
18 64.6 64.8 66.5 67.1 67.1 62.1 62.5 64.4 65.4 65.3
Intersection
12 66.6 66.7 68.4 68.8 69.3 64.4 64.8 66.0 67.3 67.4
6 71.0 71.6 72.6 72.6 73.2 67.8 68.2 70.0 70.6 70.9
24 58.3 57.9 59.3 60.0 59.9 56.6 56.4 58.2 59.0 58.9
18 59.9 59.4 60.5 61.1 61.3 57.9 57.7 59.2 60.0 60.2
All variables
12 61.6 61.0 62.2 62.2 63.0 59.6 59.4 60.5 61.3 61.7
Non 6 65.0 64.9 65.4 65.2 66.1 62.3 62.2 63.6 63.6 64.3
Parametric 24 59.2 58.9 60.6 61.0 60.9 57.2 56.8 59.2 59.7 59.6
18 60.9 60.4 61.8 62.3 62.5 58.4 58.1 60.1 60.7 60.9
Intersection
12 62.7 62.0 63.5 63.7 64.5 60.1 59.6 61.3 62.2 62.7
6 65.8 65.7 66.3 66.6 67.1 62.8 62.0 64.3 64.4 64.9
24 62.4 62.6 64.9 65.6 65.1 60.3 60.4 63.1 64.0 63.2
Calibrated Generic Bureau 18 64.6 64.4 66.4 67.1 66.9 61.9 61.9 64.2 65.2 64.8
Score 12 66.5 66.4 68.2 68.8 69.2 63.9 63.8 65.7 66.9 66.9
6 70.2 70.7 71.8 72.2 72.5 67.2 66.7 69.1 69.9 69.8

53
Figure 1:
OCC/RAD CCDB Sample Design: 1999 & 2000

1999 2000

Random sample Matched

949,790 scoreable
from 1999 master scoreable
file
(950,000 scoreable
+ 50,000 Panel (1,066,290)
unscoreable)
50,000 unscoreable

Updating Sample:
66,500 scoreable

Cross-sectional
random sample from
2000 master file
(1,000,000)

883,500 scoreable

54
Figure 2:
1999 Development and 1999 Hold-Out Sample Construction, and Bad Rates
(Bad = 90+Days Past Due, or Worse, over the Following 24 Months)

Dirty Credit History & Presently

Mildly Delinquent
Dirty Credit History (Past
13,302 Individuals
Severe Delinquency or
(p=49.27%)
Major Derogatory Note)

81,116 Individuals Dirty Credit History & Presently

The (p=20.29%) Current
Development 67,814 Individuals
Samples (p=14.60%)

338,578
individuals Clean Credit History & Thin File
(p=7.19%) 15,132 Individuals
Future (p=4.76%)
bankcard Clean Credit History
performance
is observable 257,462 Individuals
at month 24 (p=3.06%)
Clean Credit History & Thick File
Not 242,330 Individuals
At least 677,262
Presently (p=2.96%)
with one individuals
open Severely
Delinquent Mildly Delinquent
bankcard
or Worse Dirty In-Time Hold-Out Segment
with
With on History 13,207 Individuals (p=49.04%)
balance
With CCDB Valid Bankcards The In-Time 80,942 Individuals Current
update
Attribute CCDB Hold-out (p=20.39%)
date In-Time Hold-Out Segment
Records Tradeline 714,698 Samples
between 67,735 Individuals (p=14.79%)
In 1999 Accounts individuals
1/99 and Thin File
In 1999 338,684
6/99 In-Time Hold-Out Segment
1 individuals Clean
Million 995,251 (p=7.22%) History 15,150 Individuals (p=4.86%)
733,820
Individuals individuals 257,742 Individuals Thick File
individuals
(p=3.09%) In-Time Hold-Out Segment
242,592 Individuals (p=2.97%)

Individuals for whom future performance is not observable

37,436 individuals

Individuals with at least one bankcard account presently severely delinquent or worse
19,122 individuals

Individuals without any open bankcards with a balance update date in 1/99 though 6/99
261,431 individuals

Individuals with attributes but without matiching valid tradeline data in 1999
4,749 individuals

55
Figure 3:

Gains Chart (Development Samples)

1

0.9

0.8

0.7

0.6
CDF of Bads

Parametric (Stepwise)
0.5 Parametric (Resampling)
Parametric (Intersection)
0.4
SemiParametric (Stepwise)
0.3 SemiParametric (Resampling)
SemiParametric (Intersection)
0.2 CHAID (All variables)
CHAID (Intersection)
0.1 Calibrated GBS
No Separation
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
CDF of All Samples (Sorted from Low to High score)

56
Figure 4:

Gains Chart (Hold-out Samples)

1

0.9

0.8

0.7

0.6
CDF of Bads

Parametric (Stepwise)
0.5 Parametric (Resampling)
Parametric (Intersection)
0.4
SemiParametric (Stepwise)
0.3 SemiParametric (Resampling)
SemiParametric (Intersection)
0.2 CHAID (All variables)
CHAID (Intersection)
0.1 Calibrated GBS
No Separation
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
CDF of All Samples (Sorted from Low to High score)

57
Figure 5:

Empirical Log Odds by RAD Scores (Development Samples)

20
Para(Stepwise)
Para(Resampling)
Para(Intersection)
Negative Log Odds = Log((1-PD)/PD)

15 SemiPara(Stepwise)
SemiPara(Resampling)
SemiPara(Intersection)
CHAID(All variables)
10
CHAID(Intersection)
Calibrated GBS
Calibration Target
5

550 600 650 700 750 800 850

RAD Score

58
Figure 6:

Empirical Log Odds by RAD Scores (Hold-out Samples)

20
Para(Stepwise)
Para(Resampling)
Para(Intersection)
Negative Log Odds = Log((1-PD)/PD)

15 SemiPara(Stepwise)
SemiPara(Resampling)
SemiPara(Intersection)
CHAID(All variables)
10
CHAID(Intersection)
Calibrated GBS
Calibration Target
5

550 600 650 700 750 800 850

RAD Score

59
Figure 7: RAD Scores, Full Sample, Development and Validation

All Samples
Dev
Hold
Parametric 00
(Resampling)
01
02
Dev
Hold
Semiparametric 00
(Stepwise)
01
02
Dev
Hold
Nonparametric
00
(CHAID)
01
02
Dev
Hold
Calibrated
00
GBS
01
02
500 550 600 650 700 750 800 850 900
Values

60
Figure 8: RAD Scores, Dirty/Delinquent Sample, Development and Validation

Dirty/Delinquent
Dev
Hold
Parametric 00
(Resampling)
01
02
Dev
Hold
Semiparametric 00
(Stepwise) 01
02
Dev
Hold
Nonparametric 00
(CHAID)
01
02
Dev
Hold
Calibrated
00
GBS
01
02
500 550 600 650 700 750 800 850 900
Values

61
Figure 9: RAD Scores, Dirty/Current Sample, Development and Validation

Dirty/Current
Dev
Hold
Parametric 00
(Resampling)
01
02
Dev
Hold
Semiparametric 00
(Stepwise)
01
02
Dev
Hold
Nonparametric
00
(CHAID)
01
02
Dev
Hold
Calibrated 00
GBS 01
02
500 550 600 650 700 750 800 850 900
Values

62
Figure 10: RAD Scores, Clean/Thin Sample, Development and Validation

Clean/Thin
Dev
Hold
Parametric 00
(Resampling)
01
02
Dev
Hold
Semiparametric 00
(Stepwise)
01
02
Dev
Hold
Nonparametric
00
(CHAID)
01
02
Dev
Hold
Calibrated 00
GBS 01
02
500 550 600 650 700 750 800 850 900
Values

63
Figure 11: RAD Scores, Clean/Thick Sample, Development and Validation

Clean/Thick
Dev
Hold
Parametric 00
(Resampling) 01
02
Dev
Hold
Semiparametric 00
(Stepwise) 01
02
Dev
Hold
Nonparametric 00
(CHAID)
01
02
Dev
Hold
Calibrated 00
GBS 01
02
500 550 600 650 700 750 800 850 900
Values

64
Figure 12:

Empirical Log Odds by RAD Scores (2002 Validation Samples)

20
Para(Stepwise)
Para(Resampling)
Para(Intersection)
Negative Log Odds = Log((1-PD)/PD)

15 SemiPara(Stepwise)
SemiPara(Resampling)
SemiPara(Intersection)
CHAID(All variables)
10
CHAID(Intersection)
Calibrated GBS
Calibration Target
5

550 600 650 700 750 800 850

RAD Score

65
Figure 13:

Gains Chart (2002 Validation Samples)

1

0.9

0.8

0.7

0.6
CDF of Bads

Parametric (Stepwise)
0.5 Parametric (Resampling)
Parametric (Intersection)
0.4
SemiParametric (Stepwise)
0.3 SemiParametric (Resampling)
SemiParametric (Intersection)
0.2 CHAID (All variables)
CHAID (Intersection)
0.1 Calibrated GBS
No Separation
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
CDF of All Samples (Sorted from Low to High score)

66
Figure 14:

Gains Chart: Parametric (Resampling) Model

1

0.9

0.8

0.7

0.6
CDF of Bads

0.5

0.4

Development
0.3
Hold-out
0.2 2000
2001
0.1 2002
No Separation
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
CDF of All Samples (Sorted from Low to High score)

67
Figure 15:

Gains Chart: SemiParametric (Stepwise) Model

1

0.9

0.8

0.7

0.6
CDF of Bads

0.5

0.4

Development
0.3
Hold-out
0.2 2000
2001
0.1 2002
No Separation
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
CDF of All Samples (Sorted from Low to High score)

68
Figure 16:

Gains Chart: NonParametric (CHAID) Model

1

0.9

0.8

0.7

0.6
CDF of Bads

0.5

0.4

Development
0.3
Hold-out
0.2 2000
2001
0.1 2002
No Separation
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
CDF of All Samples (Sorted from Low to High score)

69
Figure 17:

Gains Chart: Calibrated Generic Bureau Score

1

0.9

0.8

0.7

0.6
CDF of Bads

0.5

0.4

Development
0.3
Hold-out
0.2 2000
2001
0.1 2002
No Separation
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
CDF of All Samples (Sorted from Low to High score)

70