Description of HERC project at eFinance
Project: Revamp credit scorecards for Hertz Equipment Rental Corp.
Hertz Equipment Rental Corp. (HERC) is an industry leader in
construction equipment rental (pumps, lifts, backhoes, etc.).
In 2003, HERC engaged eFinance to improve and automate
their credit scorecards. Their business objectives were:
● Fast decisions to customers
● Establish a rigorous credit score for each customer
● Reduce writeoffs
● Reduce costs to provide credit assessments
Core need: External and internal scorecards
HERC needed scorecards for two situations. When a customer was new, they needed external
scorecards. These scorecards would be based on Dun & Bradstreet credit information. When
credit decisions were needed for existing customers (e.g. increased limits, collections handling),
internal scorecards would leverage internal transaction and payment data.
Key elements of the project
HERC provided access to a large set (hundreds of thousands) of data: customer accounts,
rental information, payment information, writeoffs. We worked with D&B to pull historical credit
file information for customers at the time their accounts were opened.
Linear regression was used. While logistic regression was better suited (i.e. measuring a binary
outcome: writeoff or not), at the time the need for expedited work required reducing the learning
curve for new tools. I knew linear regression well, and it would provide analysis comparable to
logistic. Similarly, Microsoft Access was used to run data analysis, as I did not know SQL.
Key activities during the project included:
● Determine what outcomes were to be measured
● Hypothesizing which factors would produce better predictive results
● Creating new data for use in the analysis
● Transforming data to make it more useful
● Running correlations on data from six months prior to assess predictive ability
● Analyzing regression results and iterating the models to improve them
● Adjusting data sets based on outliers and demographic characteristics
● Running models generated from one sample on a control group to assess their integrity
● Working closely with HERC executives to incorporate their insights and to ensure buyoff
on the final models
Hutch Carpenter eFinance HERC Project 1
Examples: Creating, transforming and adjusting data
In hypothesizing factors that would provide predictive insights, I created several new variables
that became part of the scorecards.
Average Open Days (AOD): The average time a customer’s invoices have been open, weighted
by dollar amount. This input was not tracked by HERC, but made sense as a marker for
companies that would be struggling. HERC had only tracked Average Pay Days (APD; how long
it took customers to pay on invoices). AOD provided better correlation to writeoffs than APD.
Average Pay Days trend: HERC tracked APD, but did not know about the trend in APD. I
created a new variable that captured the trend in APD (increasing, decreasing, stable) over
SIC writeoff percent: HERC tracked the SIC code for customers. But it wasn’t using this
information for scorecard assessments. I analyzed the writeoffs by SIC; this became their
industry writeoff experience. The demographicbased data became part of their scorecards.
Credit Score ClassCompany size groups: D&B’s Credit Score Class (CSC) showed good
predictive potential, but the correlations were diminished by large companies. Large firms would
have poor CSC, but were not written off generally. I created a new variable that combined a
company’s CSC with its size. This new variable was powerful in predicting writeoffs.
Data transformations: In addition, basic data transformations were done to make the inputs
more useful. For example, I used the natural log function to smooth the distribution of Number of
Employees and Time in Business.
Data adjustments: One customer was skewing the regression models as a writeoff despite
good credit stats. This customer was disputing its obligations for equipment stolen from a
construction site. This was a business, not credit, issue. Separately, large sized customers were
analyzed separately, on their own internal scorecard.
Anecdotal example of scorecards’ benefit: Kenny Manta
During the course of the project, a large national customer (Kenny Manta) defaulted on its
invoiced amounts to HERC. Kenny Manta had not previously been written off. So the customer
became a test of the power of the new internal scorecard.
I took the data that was available for Kenny Manta six months prior to the default. Running the
credit scoring algorithm on Kenny Manta resulted in a Risk Rating of 8. On a 1 (best) 10
(worst) scale, an ‘8’ is among highest risk customers. If HERC had the internal scorecard at the
time, they would have had advanced warning prior to the writeoff.
The Kenny Manta analysis solidified HERC’s confidence in the statistical approach to credit
Hutch Carpenter eFinance HERC Project 2
Analysis tied directly to outcomes
A key aspect of the data analysis was to ensure the statistical models matched the outcomes
that HERC was seeking. These objectives defined whether the scorecard algorithm was
successful or not. An example of this approach is shown below. It shows the results of running a
scorecard on a sample of customers.
1 2 3 4 5 6 7 8 9 10
# Writeoffs 2 7 16 3 1 4 4 9 11 2
# Accounts 1,047 1,196 457 43 20 35 11 55 59 3
Writeoff % 0.2% 0.6% 3.5% 7.0% 5.0% 11.4% 36.4% 16.4% 18.6% 66.7%
Revenue (000) $68,331.5 $146,841.8 $54,545.6 $5,613.4 $2,023.3 $5,400.8 $1,285.4 $9,441.5 $7,812.8 $442.5
Net writeoffs $51.0 $156.3 $282.4 $22.2 $115.7 $53.9 $61.5 $333.5 $755.8 $115.9
0.1% 0.1% 0.5% 0.4% 5.7% 1.0% 4.8% 3.5% 9.7% 26.2%
Key scorecard characteristics for HERC:
● High percentage of good accounts scored Risk Rating 1, 2, 3
● High percentage of bad accounts scored 8, 9, 10
● High percentage of “good” revenue covered by Risk Ratings 1, 2, 3
● High percentage of dollar writeoffs covered by Risk Ratings 8, 9, 10
HERC implemented four statistically derived scorecards, two each for external (new) customers
and internal (existing) customers. Based on the analysis, HERC stopped using full D&B data
reports, switching to using only a subset of the data. They made this decision directly because
of the data analysis that I provided to them. The combination of operational efficiencies and
reduced data costs generated an estimated $500,000 savings per year.
These scorecards were implemented in 2003. As of this October 2015 writeup, HERC
continues to use them.
Hutch Carpenter eFinance HERC Project 3