Paper 11060-2016
Debt Collection through SAS® Analytics Lens
Karush Jaggi, AFS Acceptance; Thomas Waldschmidt, SquareTwo Financial;
Dr. Goutam Chakraborty, Oklahoma State University
ABSTRACT
Debt Collection! The two words can trigger multiple images in one’s mind – mostly harsh. However, let’s try
and think positively for a moment. In 2013, over $55 billion of past due debt was recovered by agencies in
the United States. What if all of these debts were left as is and the fate of credit issuers in the hands of
good will payments made by defaulters? Well, not the most sustainable model to say the least. In this
situation, debt collection comes in as a tool that is employed at multiple levels of recovery to keep the credit
flowing. Ranging from in-house to third party to individual collection efforts, this industry is huge and plays
an important role in keeping the engine of commerce running.
In the recent past, with financial markets recovering and banks selling less of charged off accounts and at
higher prices, debt collection has increasingly become a game of efficient operations backed by solid
analytics. This paper takes you in to the back alleys of all the data that is in there and gives an overview of
some ways modeling can be used to impact the collection strategy and outcome. SAS® tools such as
Enterprise Miner™ and Enterprise Guide™ are extensively utilized for both data manipulation and
modeling. Decision trees are given more focus to understand what factors make the most impact.
Along the way, this paper also gives an idea of how analytics teams today are slowly trying to get the ‘buy-
in’ from other stake holders in any company which surprisingly is one of the most challenging aspects of our
job.
INTRODUCTION
While collection agencies would love to recover all the money owed from its customers in one go, it is a
long road before any money is actually collected. Each customer account generally goes through a certain
lifecycle that can be summarized in the below diagram.
Preliminary Establishing right Collection efforts
skip tracing party contact through non-legal
channels
Detailed skip
tracing
Collection efforts File suit &
Close
through legal secure a
account
channels judgement
Figure 1. Debt collection lifecycle
In the figure 1, we see that once a customer is acquired, they are put through a journey that looks broadly
the same for most of them barring some exceptions like customers above the age of 65 years who cannot
be sued.
Beginning with preliminary skip tracing, every customer is put through a process of establishing certain
basic details like existing phone number, address and credit report. The aim here is to get a general history
of the customer that serves as a good starting point for making initial decisions about contacting them.
Once their identity is mapped, the process of establishing contact with them is begun. It may come as a
surprise, but even after all the skip tracing efforts, getting the right person on the phone or reaching them
via snail mail is one of the biggest challenges affecting ROI. The next step after establishing a right party
contact (RPC) is to begin non -legal collection efforts (unless they dispute the debt, which leads to another
story). This step includes offering payment plans, settlement options, upfront payments etc. However, it is
observed that a considerable chunk of customers are still unable to make the required payments after which
their account proceeds to a more stringent legal channel. This channel involves a detailed skip tracing effort
aimed at exploring details like properties owned by the customer or their employment details. This channel
usually leads to a judgement giving the collections company options such a putting a lien on property or
garnering wages. An important thing to note here is that the legal channel is an expensive option for the
collections company considering it involves an array of court costs and attorney fees.
What is interesting to note here is that at all stages of this process, considerable amount of data collected
and stored about each and every customer that makes way for statistical analysis and predictive modelling
in a way that optimizes operational efficiency. It is not only in-house data that is available, but also a whole
spectrum of third party data that can be bought off the shelf to aid such advanced analytics.
Getting in to the analytics side of this, a desired analytics process is one that will equip the collections
agency with insights into basic questions like “What are the chances this customer will pay?”, “When is the
right time to collect from this customer?”, “How much should be collected from this customer?” etc. With this
in mind, the broad aim of this paper is really to understand what factors play an important role in maximizing
collections and how we can capitalize on them. A few of the industry standard models that are usually used
in tandem today are:
1. To predict if a customer is going to pay or not in its entire lifetime (Probability of Payment or PoP)
2. If yes, to predict the approximate month of first payment
3. To predict the $ amount that will be recovered from the customer
The results from all the above models will go into deciding the course of treatments an account will be
subject to in its recovery lifecycle. For example, an account predicted to be a payer as well as predicted to
make the first payment in month 6 from date of buying is a good indication to make collection efforts
through traditional calling etc. However, an account predicted to be a payer but predicted to make the first
payment in month 24 from date of buying is a sure shot indication to take legal action against them
considering we know that all payers after month 18 are known to pay after being sued.
HOW TO GET PROBABILITY OF PAYMENT?
Most industry experts will concur that predicting the probability of payment is one of the most critical set of
information to have while buying & working charged off accounts. This probability can be calculated for
payments within a certain time frame (like 6 months) or over the life of an account. We will see ahead how
to adjust the target variable to suit different needs. Naturally, predicting this probability is a relatively
complex task that is affected by a number of external factors that are not captured by the system. However,
a wide variety of internally available data is a good starting point for us.
To begin with, let us look at a few meaningful attributes about a customer that are available in-house.
ATTRIBUTE TYPE USABLE TRANSFORMATION
ACCOUNT_OPEN_DATE Date DAYS_SINCE_ACCOUNT_OPENED
CHARGEOFF_DATE Date DAYS_SINCE_CHARGEOFF
DELINQUENCY_DATE Date DAYS_SINCE_DELINQUENT
LAST_PREPURCHASE_PAYMENT_DATE Date DAYS_SINCE_LAST_PREPURCHASE_PAYMENT
FIRST_PAYMENT_MONTH Date FIRST_PAYMENT_MONTH_FROM_PURCHASE
CHARGEOFF_AMOUNT Numeric USE AS IS
ORIGINAL_PURCHASE_AMOUNT Numeric USE AS IS
ORIGINAL_LOAN_AMOUNT Numeric USE AS IS
ORIGINAL_CREDIT_LIMIT_AMOUNT Numeric USE AS IS
INTEREST_RATE Numeric USE AS IS
Table 1. In house customer attributes
As we can see, the date type variables are converted to a more usable format of number of days since the
date captured. This can be easily performed using the SAS function INTCK (interval, from, to) as shown
below.
DAYS_SINCE_DELINQUENT=
INTCK("DAYS",DATEPART(LAST_PREPURCHASE_PAYMENT_DATE),
DATEPART(PURCHASE_DATE)) + 1;
DATEPART function is used to extract date from a date time variable.
Moving forward, it is worth while discussing some of the third party data available today without much
hassle. To begin with, the US Census provides a broad range of ZIP code level demographic data that can
be crucial considering we know where our customers live. Some examples are median household income,
median home value, median per capita income and more. These are generally good indicators of the
neighbourhood that a customer lives in.
Other sources of third party data include credit bureaus like FICO®, Experian® and TransUnion® or public
record aggregators like LexisNexis® and TLo. Let’s look at some attributes that we can get from them.
ATTRIBUTE RANGE/TYPE SOURCE DEFINITION
PROPENSITY RECOVERY 350-850 TransUnion® Likelihood of collecting $50 or more within
12 months
SCORE 3.0
YIELD RECOVERY SCORE 3.0 000-999 TransUnion® Identifies accounts likely to pay more
money
RECOVERY BANKCARD SCORE 350-850 TransUnion® Likelihood of collecting $100 or more within
6 months
INCOME ESTIMATOR 3.0 000-999 TransUnion® Estimates an individual’s income based on
credit history
FICO CREDIT SCORE 300-850 FICO® Customer’s credit score
VerifiedSSN BINARY LexisNexis® Indicates if the customer’s input SSN is
verified
VerifiedPhone BINARY LexisNexis® Indicates if the input phone is verified
VerifiedAddress BINARY LexisNexis® Indicates if the input address is verified
InputAddrAgeNewestRecord 0-960 LexisNexis® Indicates if the input address in address
history
FelonyAge 0-84 LexisNexis® Months since most recently recorded
felony conviction
Table 2. Third party customer attributes
The above attributes only scratch the surface and are nowhere close to the exhaustive list that is available
for deployment. We can easily access detailed data revolving around education, income, professional
associations, address etc.
At this point, we are equipped with a lot of data that will be helpful in predicting probability of payment
(PoP). To understand PoP, let’s visualize a set of customers as show below.
Credit Demographic High PoP
Attributes Attributes
Med PoP
In-house
Attributes
Low PoP
Figure 2. Visualizing Probability of Payment
In figure 2, we see that all of the customer attributes are used as input to SAS tools for modeling. To begin
with, the first step here is to get insights into what attributes are actually useful and have some predictive
value.
While there are multiple ways to do this, the Kolmogorov Smirnov statistics & charts are widely used in risk
analytics. In its simplest form, this is basically a non-parametric test of equality of two probability
distributions. Basically checking a null hypothesis that the two distributions are identical. If the KS statistic is
large enough, the two are significantly different. Visually, this is measured as the distance between
distributions of a positive outcome and negative outcome for the analysis variable.
Let’s see how to do this.
We can use PROC NPAR1WAY to get KS statistics as shown below.
ODS GRAPHICS ON;
/*GENERATING KS STATISTICS & CHARTS*/
PROC NPAR1WAY EDF DATA=LIB.DATASET;
CLASS PAYER_FLAG ;
VAR ADDRCHANGECOUNT01 CTR3P2E;
OUTPUT OUT = LIB.OUTPUT (KEEP = _VAR_ _KS_ _D_);
RUN;
As can be seen, I have used the EDF option in the PROC NPAR1WAY statement that gives us KS statistic.
A host of other options can be used to get other statistics.
The CLASS statement above can be used to specify the attribute we are trying to classify, payers versus
non-payers in this case (PAYER_FLAG). VAR statement is optional and can be excluded if all variables are
to be used.
ODS GRAPHICS ON ensures we get the below empirical distribution graph (figure 3).
Figure 3. Empirical Distribution using KS Statistic
As can be seen above in figure 3, variable ADDRCHANGECOUNT01 is not very helpful in separating
payers from non-payers. An important consideration with the KS Statistic is that since it is a univariate
analysis, we do not get any other details such as collinearity of variables and hence we use multivariate
analysis as described ahead.
This information about the predictive capability of an attribute is very crucial for third party data since most
of them come at a certain price. When dealing with millions of customer records, this makes a considerable
difference in the operational costs of collections.
Once we have identified the important variables, we can move to our next step which involves using a
decision tree to see how the selected variables stack up and look at them from a business point of view.
Let us begin with this diagram from Enterprise Miner
Figure 4. SAS® Enterprise Miner Diagram
In the above figure 4, we can see our data set EM_SET deployed with a simple regression, neural network,
decision tree and tree regression.
Going into the decision tree, we set it up for rank ordering customers based on their propensity to pay and
hence the target is the variable PAYER_FLAG. The detailed properties are as below in figure 5:
Figure 4. SAS® Enterprise Miner Diagram Properties
An important point to note here is the assessment measure chosen for the stopping point of the tree is
average square error. We chose to use this average square error because the end goal of this model is to
provide a rank order of accounts for incoming debt purchases. This rank order will be used in establishing
operational strategies and prioritizing work effort. Also, please note the other options like max branch, max
depth etc. are good to work with for optimizing complexity of the decision tree.
Decision tree have a variety of outputs that we can use. From estimating variable importance to scoring a
new data set, they can be used for a variety of purposes. One of the more used application of decision trees
is with regression as a variable selection technique.
Looking at the decision tree map show below, we get a very compact overview of the separation between
the cases and non-cases. The node width is representative of the number of cases contained in the node
after the split. Also, by default, the color of the node is reflective of the number of cases in the node. In
figure 5 below, we can see that most of our cases (black nodes) are in the bottom right of the map which
gives a sense of the payment pattern that we are trying to unravel from the data.
Figure 5. SAS® Decision Tree Map
Coming to the variable importance report generated by the decision tree, we have the following variables
that are ranked by their predictive value. The importance measure below is basically a relative score given
to each variable when compared to the most important variable, TU_RECOVERY_SCORE in this case. An
interesting application of this score is to identify irregularities or inconsistencies with the node splits that the
tree comes up with. This is done by comparing the importance of a variable with the vimportance (validation
importance) for the same variable and if they are markedly different, we know there is an issue. The ratio
column on the extreme right is a ratio of vimportance/importance that can also be used. A ratio of 1 is
desired.
Figure 6. Decision Tree Variable Importance
Now that we have used the decision tree and have some results from it, we will deploy this decision tree in
tandem with a regression model (also called a tree regression). Consider the highlighted branch of the
diagram below.
Figure 7. SAS® Enterprise Miner Diagram – Tree
Regression
Since regression does not deal well with missing values, the impute node is used for performing missing
value imputation. This node offers multiple imputation options for class variables and interval variables. One
of the interesting methods of imputation offered is a tree imputation technique which is a versatile method of
replacing missing values using an inbuilt decision tree. In essence, it tries to predict what the missing value
is most likely to be based on all other available attributes.
As this project is a part of an internship, non-disclosure agreements keep me from divulging more
interesting details of the various models that were developed.
TREATMENTS BASED ON PROBABILITY OF PAYMENT
Now that we have results from a fairly detailed statistical analysis, the next step is to take a business view
of the results and take some decisions based on the same. Taking a step back to revisit the final aim of this
analysis, we are trying to maximize our collections by way of increasing operational efficiency and deploying
the right strategy for each account. So with this in mind, let us see how to use our new found knowledge for
establishing treatments that are best suited for each account.
The below example takes in to account results from a few other models that were mentioned above in the
paper and work in tandem with the PoP model we just went through.
Consider an account that gets a high probability of payment and an approximate month 4 of first payment.
We can then use the below chart to assign a suitable plan of action for the account based on past
knowledge. So if an account is highly likely to pay in month 4 from acquisition, we can confidently assign it
to a call center through a non-legal channel which would keep our over-head expenses very low and return
on investment high.
However, in case we have an account at hand that has a high probability of payment but the estimated
month of first payment is well after 24 months, we will need to deploy a different treatment. In this case, we
can subject the account to professional skip tracing and gather information about mortgage payments made
in the recent past or a recent increase in credit rating. Using these indicators, a legal rout can be adopted
based on the fact that such positive financial activity indicates that the customer is financially sound and can
be profitably sued. However, the efficiency of skip tracing at this point becomes crucial considering the legal
route is very expensive to go through.
Figure 8. Treatments
CONCLUSION
With banks selling less delinquent accounts and at higher prices, the margins for debt collection companies
are slim to work with. The debt collection industry is maturing in many aspects and it is vital for analysts to
leverage ways that give the company an edge over others. Analytics is bound to be one of the most
important tools that will aid better decision making. As described above, data is abundantly available and
more can be used (for example customer data from mobile carriers) to support any analysis that can be
conceived.
REFERENCES
Collections Information: Facts and statistics about third-party debt collection.
http://www.acainternational.org/products-collections-information-5431.aspx
US Census
http://www.census.gov/
Decision Tree Validation: A Comprehensive Approach
http://www2.sas.com/proceedings/sugi30/256-30.pdf
ACKNOWLEDGMENTS
I would like to thank my professor Dr. Goutam Chakraborty for being a great guide and pushing me to bring
out my best. I also thank my colleagues at SquareTwo Financial – Eric Hayes, Annette Girmus Orford,
Harold Dickerson and Thomas Waldschmidt for taking me under their wings and guiding me as their intern.
RECOMMENDED READING
Enhancing delinquent debt collection using statistical models of debt historical information and
account events
US 7191150 B1
http://www.google.com/patents/US7191150
CONTACT INFORMATION
Your comments and questions are valued and encouraged. Contact the authors at:
Karush Jaggi
Risk Analyst, AFS Accpetance
Ft. Lauderdale, FL – 33309
405-762-1946
karush.jaggi@okstate.edu
Thomas Waldschmidt
SquareTwo Financial
Denver, CO - 74075
720-217-6046
thomasw2101@icloud.com
SAS and all other SAS Institute Inc. product or service names are registered trademarks or trademarks of
SAS Institute Inc. in the USA and other countries. ® indicates USA registration.
Other brand and product names are trademarks of their respective companies.
