UV3846
Rev. April 23, 2009
SELECT COLLECTIONS, INC.
Marissa Wells, manager of collections for Select Collections, Inc., called in Marcos
Kilduff, a summer intern, to describe an assignment:
I want to see if we can build a model to predict how much money we will collect
from delinquent accounts. Such a model would prove very useful in deciding
which accounts to purchase and how much to pay.
You learned regression during your first year, didn’t you? Well, see what you can
do for us. There are 3,570 accounts in a data set that I will call the training set.
We purchased these accounts and then collected the amount shown in the last
column, labeled totalpay. Totalpay is the variable of interest. I want you to come
up with a way to forecast totalpay.
We’ve also assembled a bunch of potential predictor variables for you to use. All
of these variables are “useable” in the sense that we will know them for new
accounts about which we are making decisions. Your model can use any of these
variables in any combination and in any form. The data dictionary [see Exhibit 1]
explains each of these variables.
To make this assignment interesting, I want you to use the model you come up
with to predict totalpay for the 3,570 accounts in a separate worksheet of data that
I will call the test set. The accounts in this test data set are no different from the
accounts in the training set except that I know the totalpay values and you don’t.
So, by tomorrow morning, I want an Excel file from you containing a single
column of numbers. In cell A1, please put the adjusted R-squared for your model.
In cells A2 through A3571, put your predicted values for the accounts in the test
data set in ID order. Tomorrow, I’ll use the actual totalpay values for the accounts
in the test set to calculate your model’s mean squared error—the average of your
model’s 3,570 squared errors.
This case was written by Thomas A. Pomroy, Phillip E. Pfeifer, and William Scherer of the University of Virginia.
It was written as a basis for class discussion rather than to illustrate effective or ineffective handling of an
administrative situation. Copyright 2002 by the University of Virginia Darden School Foundation, Charlottesville,
VA. All rights reserved. To order copies, send an e-mail to [email protected]. No part of this
publication may be reproduced, stored in a retrieval system, used in a spreadsheet, or transmitted in any form or by
any means—electronic, mechanical, photocopying, recording, or otherwise—without the permission of the Darden
School Foundation. Rev. 4/09.
This document is authorized for use only by Maria Marcela Bonilla Gutierrez in Data Mining y Analítica Predictiva (MAIT-2025) at INCAE Business School, 2024.
-2- UV3846
And, in the interest of full disclosure, I’m giving this same assignment to all our
interns. It’ll be kind of fun, don’t you think?
Oh, I almost forgot. We’re also giving you some descriptive statistics and plots
[see Exhibit 2] to give you a head start.
Company Background
Select Collections, Inc., was a start-up subsidiary of a major credit card company. The
company purchased distressed consumer debt at discounted rates from such major credit card
companies as Chase, Wells Fargo, and Bank of America and then used data-driven decision-
making and dynamic value assessment to optimize the collection processes associated with the
purchased accounts.
For each purchased account, the first decision was whether to resell or attempt to collect.
For accounts the company decided to attempt to collect, a host of tactics was available for use in
any sequence and with any frequency.
Select Collections’ strategic intent was to become the best in the world at tailoring and
optimizing the collection process. Like other collection companies, Select Collections used the
telephone and the legal system as its two major collection tools. The company had recently
opened a large, state-of-the-art call center in Topeka, Kansas, and its legal department was active
in 46 states. Select Collections strictly followed all state and federal regulations regarding the
collection process.
This document is authorized for use only by Maria Marcela Bonilla Gutierrez in Data Mining y Analítica Predictiva (MAIT-2025) at INCAE Business School, 2024.
-3- UV3846
Exhibit 1
SELECT COLLECTIONS, INC.
Data Dictionary
Field Definition
acctid Account ID, a unique index used to identify accounts
state State in which the account-holder lives (TEXT)
zip ZIP code in which the accountholder lives
rollout Card issuer from whom the account was purchased
cobal The balance of the account at the point of “charge-off” when the account
was purchased
collscr The interpretation of this field is not fully known; however, it is produced
by an in-house legacy system and exported to the data warehouse for every
account. This is a text variable
cs The left-most two digits of collscr. This variable is a number.
accessscr The account’s accessibility score, an a priori estimation of how likely it is
that the accountholder is reached via phone.
lnacscr The natural log of the accessscr
bureauscr The output of an in-house prediction model based on the likelihood of
receiving payments from the accountholder using credit-bureau attributes as
predictors
eaglemod The output of the Eagle System’s proprietary model, an estimate of the risk
of nonpayment of the account
numcalls The number of telephone calls made (to date) to the accountholder
numrpcs The number of “right-party connects” or phone calls in which the collection
agent speaks with the accountholder
totalpay The total amount of payments received from the accountholder (in dollars)
Source: Created by case writer.
This document is authorized for use only by Maria Marcela Bonilla Gutierrez in Data Mining y Analítica Predictiva (MAIT-2025) at INCAE Business School, 2024.
-4- UV3846
Exhibit 2
SELECT COLLECTIONS, INC.
Descriptive Statistics and Plots of Training Data Set
Variable N Mean Median TrMean StDev Minimum Maximum
cobal 3570 3231.7 2525 3030 2268.2 263 17534
cs 3570 4.4031 4 4.2808 2.8821 1 10
accesssc 3570 0.49524 0.4928 0.49485 0.2895 0 0.9999
lnacscr 3570 -1.0163 -0.71 -0.9052 1.0088 -11.15 0
bureausc 3570 121.27 123 121.91 15.81 66 149
eaglemod 3570 50.042 50 50.045 28.531 1 99
numcalls 3570 80.87 62 72.05 75.49 1 731
numrpcs 3570 3.3922 2 2.9312 3.3448 1 34
totalpay 3570 807.9 463.5 673.8 1016.8 1 13867
Rollout Count
Associates 548
Bank_Of_Am 631
Chase 336
Chase_Bony 23
Chase_Rev 197
Discover 1178
Wells 614
Wells_FIB 43
TrMean refers to trimmed mean. The trimmed mean is the average calculated after removing the 5% highest and 5%
lowest values.
Source: Created by case writer.
This document is authorized for use only by Maria Marcela Bonilla Gutierrez in Data Mining y Analítica Predictiva (MAIT-2025) at INCAE Business School, 2024.
-5- UV3846
Exhibit 2 (continued)
Source: Created by case writer.
This document is authorized for use only by Maria Marcela Bonilla Gutierrez in Data Mining y Analítica Predictiva (MAIT-2025) at INCAE Business School, 2024.
-6- UV3846
Exhibit 2 (continued)
Source: Created by case writer.
This document is authorized for use only by Maria Marcela Bonilla Gutierrez in Data Mining y Analítica Predictiva (MAIT-2025) at INCAE Business School, 2024.
