You are on page 1of 21

Applying Diverse Analytics to

Improve Automotive Collections


Gene Grabowski, Jr.
Operations Analytics Supervisor
Ford Motor Credit Company

Copyright © 2007, SAS Institute Inc. All rights reserved.

Introduction
ƒ Webster’s Online Dictionary definition of “Diversity”
• Noticeable heterogeneity
• The condition or result of being variable

ƒ Diverse Analytics
• The ability to employ distinct disciplines to develop
knowledge or address complex problems

ƒ Goal
• Encourage the use and integration of diverse analytics

2
Copyright © 2007, SAS Institute Inc. All rights reserved.

Copyright © 2007, SAS Institute Inc. All rights reserved. 1


Problem Resolution

ƒ “When people do not have the ability to


resolve a problem, they not only become
tolerant to the problem, but actually start to
deny its existence”

ƒ “Those who have the ability to resolve


problems, start looking for larger problems to
solve”

Of Paradise and Power by Robert Kagan

3
Copyright © 2007, SAS Institute Inc. All rights reserved.

Session Outline

ƒ Ford Credit Overview

ƒ Role of Collection Operations

ƒ Role of Collection Analytics

ƒ Detail how diverse analytical methods are


currently being integrated
• SAS facilitates this integration
4
Copyright © 2007, SAS Institute Inc. All rights reserved.

Copyright © 2007, SAS Institute Inc. All rights reserved. 2


Role of Ford Credit

ƒ Ford Motor Credit Company (FMCC) is an indirect,


wholly-owned subsidiary of the Ford Motor Company
ƒ Since 1959, FMCC has financed over 50 million
customers
• Primary offerings are car loans and leases
• Also offers auto insurance, interest bearing accounts and fixed rate notes
• Services Retail and Commercial customers
• Represented in North America, Europe, Latin America and Asia Pacific

ƒ Serves as parent organization to the following:


Primus Financial Services Mazda Finance
Volvo Car Finance Jaguar Financial Services
Land Rover Finance
5
Copyright © 2007, SAS Institute Inc. All rights reserved.

Role of Ford Credit

ƒ Goals of Ford Credit:


• Serve as revenue source for Ford Motor Company
• Help enhance overall vehicle sales
− Provide financing to targeted customer segments
− Work in tandem with Ford Marketing & Customer Services
• Ensure customer fulfills contractual obligations
ƒ Analytics integrated with SAS can help by:
• Growing the number of outstanding contracts
• Pricing contracts appropriately
• Limiting delinquencies and defaults

6
Copyright © 2007, SAS Institute Inc. All rights reserved.

Copyright © 2007, SAS Institute Inc. All rights reserved. 3


Goal of Ford Credit Collections

Limit
Delinquencies &
Defaults

Collection Collection
Operations Analytics

7
Copyright © 2007, SAS Institute Inc. All rights reserved.

Ford Credit Collection Operations

ƒ Comprised of 6 U.S. and 1 Canadian business centers


- Serve as call centers to notify delinquent customers
- Help expedite delinquent payments
- Handle contract originations work
- Regionally distributed to meet national needs

ƒ Business customer of analytical research

8
Copyright © 2007, SAS Institute Inc. All rights reserved.

Copyright © 2007, SAS Institute Inc. All rights reserved. 4


Ford Credit Collection Operations

ƒ Collections workload is divided into 2 primary


categories
• Net Campaign:
- Lower risk customers
- Customers in initial stages of delinquency
• Base Campaign:
- Higher risk customers
- Customers in continued stage of delinquency
- Worked by more experienced and effective collectors
ƒ Campaign distinction was not analytically driven
• Created by Collection Operations
9
Copyright © 2007, SAS Institute Inc. All rights reserved.

Ford Credit Collection Analytics – Key Tasks

ƒ Analytical Data Management


ƒ Risk Assessment of Default
ƒ Optimal Assignment Timing
• Initiation of collections work
ƒ Seasonality
ƒ Time Series Forecasting
ƒ Optimal Staffing

10
Copyright © 2007, SAS Institute Inc. All rights reserved.

Copyright © 2007, SAS Institute Inc. All rights reserved. 5


Ford Credit Collection Analytics - Goals

Look Familiar??

11
Copyright © 2007, SAS Institute Inc. All rights reserved.

Analytical Data Management

ƒ Collection Operations use:


• Teradata RDBMS for storage
• Business Objects for reporting
ƒ Collection Analytics requires the versatility and
statistical capabilities of SAS
ƒ Foundation of analytics: Sound programming and
data manipulation skills
• Ability to implement very specific business rules
• Ability to gather and combine data from multiple sources

12
Copyright © 2007, SAS Institute Inc. All rights reserved.

Copyright © 2007, SAS Institute Inc. All rights reserved. 6


Analytical Data Management – Best Practices

ƒ Use SAS Macros, SAS/GRAPH, reporting PROCs


and SAS/ODS etc. to manage and report data
ƒ Learn basics of different operating systems
• e.g. UNIX, LINUX, Mainframe
ƒ Investigate Structured Query Language (SQL)
• Relational Databases & SAS/ACCESS
ƒ Understand Data Warehousing concepts
• Single source of truth
ƒ Goal: Increase analytical independence and
decrease reliance on IT Department
13
Copyright © 2007, SAS Institute Inc. All rights reserved.

Risk Assessment of Default

ƒ Probability of Default (POD) Model


• Assesses risk of default over multiple timeframes
• Logistic Regression Model (PROC LOGISTIC)
ln(p/(1-p)) =b0 + b1x1 + b2x2 + … + bnxn
• Account level binary dependent variable
1= Default - “Bad”
0= No default - “Good”
• Uses variables from three broad categories:
- Application Type, Payment History, Credit Bureau

14
Copyright © 2007, SAS Institute Inc. All rights reserved.

Copyright © 2007, SAS Institute Inc. All rights reserved. 7


Risk Assessment of Default

The Development Sample:

Performance Window

All active accounts at time T Default indicator at time T+N as


stored by Ford Credit

ƒ POD Model “Splits”


• Lease / Retail Contract (Auto Loan)
− Non-Delinquent
− 1 Payment Delinquent
− 2 + Payments Delinquent
15
Copyright © 2007, SAS Institute Inc. All rights reserved.

Risk Assessment of Default

ƒ POD Initial Steps


• Extract / Cleanse / Merge Data
• Explore Data (PROC UNIVARIATE, SAS/GRAPH)
• Variable Reduction & Transformation
− Bin Continuous & Categorical Variables
− Create dummy variables
PROC TRANSREG DATA = INPUT DESIGN;
MODEL CLASS ( VAR1 VAR2 …. VARN / ZERO = NONE);
OUTPUT OUT = DUMMY_OUTPUT ;
RUN;
− Automatically creates a Dummy Variable for each level
of the input variable (VAR1, VAR2, … VARN)

16
Copyright © 2007, SAS Institute Inc. All rights reserved.

Copyright © 2007, SAS Institute Inc. All rights reserved. 8


Risk Assessment of Default

ƒ Type of code eliminated with PROC TRANSREG


• Example Variable: STATE=(‘TEXAS’,’OHIO’)

• IF STATE=‘TEXAS’ THEN DUMMY_TEXAS=1; ELSE


DUMMY_TEXAS=0;
• IF STATE=‘OHIO’ THEN DUMMY_OHIO=1; ELSE
DUMMY_OHIO=0;

• DUMMY_TEXAS=(STATE=‘TEXAS’);
• DUMMY_OHIO=(STATE=‘OHIO’);

17
Copyright © 2007, SAS Institute Inc. All rights reserved.

Risk Assessment of Default

ƒ POD Model Specification


• SELECTION = BACKWARD selection process
• Rank order significant variables by chi-square level
ƒ Compute Concordance
• All combinations of Good and Bad accounts
• Count pairs where POD of Goods < POD of Bads
− “Concordant” Good/Bad pair
• Divide Concordant Good/Bad pairs by all Good/Bad pairs
ƒ Test Kolmogorov Smirnov (KS) Statistic

18
Copyright © 2007, SAS Institute Inc. All rights reserved.

Copyright © 2007, SAS Institute Inc. All rights reserved. 9


Risk Assessment of Default

ƒ Behavior scores help differentiate delinquent


customers who are likely to pay versus those who
will not
"Non-payers" MEDIUM
"Payers"
G
B G G G LOW
B G G G
B Apply behavior scores HIGH G
B G G B G G G
G G B G G G
B B
B B B G G
G G B B B B B G G G G
G
B BB B B G G B G G G G
high risk Beh Scr low risk

G represents a customer who will pay


B represents a customer who won’t pay
19
Copyright © 2007, SAS Institute Inc. All rights reserved.

Risk Assessment of Default

ƒ Entire active portfolio is scored every month


ƒ Scores transformed to POD
ƒ POD validated against actual default rates
ƒ All results compared to Validation sample
• 25% random selection from Development Sample

⎛ POD ⎞
ln⎜ ⎟ = β 0 + β1 ( Score)
⎝ 100 − POD ⎠
⎡ 1 ⎤
POD = 100 × ⎢ − ( β 0 + β 1 ( Score )) ⎥
⎣1 + e ⎦ 20
Copyright © 2007, SAS Institute Inc. All rights reserved.

Copyright © 2007, SAS Institute Inc. All rights reserved. 10


Optimal Assignment Timing

ƒ A subsequent analysis is performed to


determine when a delinquent account should
be worked by a collector
• When = Number of Days Past Due (DPD)

ƒ An “Optimal Assignment Timing” is the


number of DPD at which:
Operational Costs = Losses Saved

21
Copyright © 2007, SAS Institute Inc. All rights reserved.

Optimal Assignment Timing


ƒ Operational Cost inputs:
• Annual salary of collector
• Optimal number of Full Time Employees (FTEs) to work 10,000
accounts

ƒ Saved Loss inputs:


• Observed default rates by risk level
• Losses per repossession (Severity)
• Number of opportunities to work an account within a year

ƒ Both set of inputs are in terms of annual dollars


22
Copyright © 2007, SAS Institute Inc. All rights reserved.

Copyright © 2007, SAS Institute Inc. All rights reserved. 11


Optimal Assignment Timing

ƒ Inputs are combined to determine a “break-even


probability of default”
• A level of risk an account must attain before it is worth
assigning to a collector
ƒ Break-even levels are calculated for Net and Base
Campaign
ƒ Analysis stills lacks time as risk dimension

23
Copyright © 2007, SAS Institute Inc. All rights reserved.

Optimal Assignment Timing

ƒ Group accounts by POD for a given month


• 20 equally sized “bins” – initially!

PROC RANK DATA=INPUT DATA OUT=RANK_DATA GROUPS=20;


VAR POD;
RANKS POD_BINS;
RUN;

Break-Even
ƒ Estimate 12 month Probability of Level

Default conditioned on the


number of “days to clear”
for each risk bin
24
Copyright © 2007, SAS Institute Inc. All rights reserved.

Copyright © 2007, SAS Institute Inc. All rights reserved. 12


Optimal Assignment Timing - Overlay Curves

Break-Even
Level

Since bin 1’s default rate is always less than the


Break-Even default rate, these accounts should not be
issued for work.

25
Copyright © 2007, SAS Institute Inc. All rights reserved.

Seasonality – Prelude to Forecasting

ƒ Behavior that occurs on the basis of time


• Annual, semiannual, quarterly, monthly or any other
unit of calendar time
ƒ Conceals important characteristics
• Requires an adjustment process to enable period-
over-period comparisons
ƒ 2 Types of Seasonality
• Deterministic: Patterns repeats year over year
• Stochastic: No systematic pattern year over year

26
Copyright © 2007, SAS Institute Inc. All rights reserved.

Copyright © 2007, SAS Institute Inc. All rights reserved. 13


Seasonality- Prelude to Forecasting
ƒ A “Time Series” Yt consists of 4 factors:
Yt = Ct * St * It * TDt (Multiplicative – factors are related)
Yt = Ct + St + It + TDt (Additive – factors are independent)

• Trend Cycle (Ct): Captures long-term trends and business cycle variations
• Seasonal (St): Monthly pattern of variation that repeats year over year
• Irregular (It): Unpredictable shocks that are transitory in nature
• Trading Day (TDt): Measures the frequency and intensity of business days
within the calendar month

ƒ Isolate the Trend Cycle (Ct) to make forecasting easier


ƒ Factors have a base of 100 (or 1). 100=No Seasonality
ƒ Ford Credit factors are Multiplicative
27
Copyright © 2007, SAS Institute Inc. All rights reserved.

Seasonality - PROC X11

ƒ Based on the U.S. Bureau of the Census X-11


seasonal adjustment methodology
ƒ Decomposes monthly or quarterly series into
seasonal, trend, trading day and irregular
factors
ƒ Provides multiplicative and additive form of
the decomposition
ƒ Enhanced version of this methodology,
PROC X12, is also available

28
Copyright © 2007, SAS Institute Inc. All rights reserved.

Copyright © 2007, SAS Institute Inc. All rights reserved. 14


Seasonality- PROC X11
Percentage of
%MACRO X11(CENTER); Work by Day
of Week
PROC X11 DATA=COMBINE NOPRINT YRAHEADOUT;
VAR &CENTER._CAMPAIGN ;
ID SASDATE; /* DATE VARIABLE */
MONTHLY DATE=SASDATE CHARTS=NONE TDREGR = ADJUST;
PDWEIGHTS MON=&MON TUE=&TUE WED=&WED THU=&THU FRI=&FRI SAT=&SAT;

OUTPUT OUT=X11_&CENTER /* OUTPUT DATASET */


A1=&CENTER._ORIGINAL /* ORIGINAL TIME SERIES */
C17=&CENTER._IRR /* IRREGULAR FACTOR */
C18=&CENTER._TD /* TRADING DAY FACTOR */
D10=&CENTER._SA; /* SEASONAL FACTOR */
RUN;

%MEND X11;
%X11
29
Copyright © 2007, SAS Institute Inc. All rights reserved.

Time Series Forecasting

ƒ Requirements
• 6-15 month workload outlook for 7 business centers
− 3 forecasting cycles within year
• Performed for both Net and Base Campaign
ƒ Candidate SAS Procedures
• PROC FORECAST – Simple but effective
• PROC AUTOREG – Regression with autoregressive errors
• PROC ARIMA – Utilizes Box-Jenkins methodology
− Others: PROC SPECTRA, PROC VARMAX, PROC STATESPACE
ƒ Autoregressive Integrated Moving Average (ARIMA) provides
most robust approach in the short run
30
Copyright © 2007, SAS Institute Inc. All rights reserved.

Copyright © 2007, SAS Institute Inc. All rights reserved. 15


Time Series Forecasting – ARIMA

ƒ PROC ARIMA
• Allows for independent variables
• Allows for intervention (dummy) variables
• Integrates well with PROC X11
ƒ Autoregressive Process (AR)
• Yt = A1 Yt-1 + A2Yt-2 + ... ApYt-p + et
• Relates the current value to its own previous “p” values
ƒ Moving Average Process (MA)
• Yt = W1 et-1 + W2et-2 + …. Wqet-q
• Relates the current value to the previous “q” errors
31
Copyright © 2007, SAS Institute Inc. All rights reserved.

Time Series Forecasting – Stages of ARIMA

ƒ Identification Stage

ƒ Estimation Stage

ƒ Diagnostic Stage

ƒ Forecasting Stage

32
Copyright © 2007, SAS Institute Inc. All rights reserved.

Copyright © 2007, SAS Institute Inc. All rights reserved. 16


Time Series Forecasting – Stages of ARIMA

ƒ Identification Stage
• Time Series Yt must be “Stationary”
− Mean, Variance and Lag Covariance constant over time
− No “Unit Root”
• Corrected through differencing and log transformation
− Tested with SAS Macros: %DFTEST, %LOGLIK
− PROC ARIMA STATIONARITY option

Example of Non-
Stationary Time Series

33
Copyright © 2007, SAS Institute Inc. All rights reserved.

Time Series Forecasting – Stages of ARIMA

ƒ Estimation Stage
• Approximate the coefficient of Autoregressive and
Moving Average terms (if any!)
− Utilize autocorrelation & partial autocorrelation methods
• Determine the significance of date dummy variables
− Dates derived from X11 Irregular Factors (It)
• Determine the significance of independent variables
− Examine cross correlation of independent and dependent
variables to fit Transfer Function

34
Copyright © 2007, SAS Institute Inc. All rights reserved.

Copyright © 2007, SAS Institute Inc. All rights reserved. 17


Time Series Forecasting – Stages of ARIMA

ƒ Diagnostic Stage
• Goodness of Fit
− Minimize Akaike’s Information Criterion (AIC) & Swartz’s
Bayesian Criterion (SBC)
• Autocorrelation of Model Residuals
− Non significant “Q Statistic”

ƒ Forecasting Stage
• Are forecasts realistic?
• Within expectations?
• Check forecast errors Æ Et = Yt − Ft

35
Copyright © 2007, SAS Institute Inc. All rights reserved.

Times Series Forecasting – Error Tests

ƒ Mean Absolute Error

ƒ Mean Absolute Percentage Error

ƒ Mean Square Error

ƒ Root Mean Squared Error

36
Copyright © 2007, SAS Institute Inc. All rights reserved.

Copyright © 2007, SAS Institute Inc. All rights reserved. 18


Optimal Collections Staffing

ƒ Defined as the optimal number of Full Time


Employees (FTEs) required to work 10,000
accounts
• Estimate is used in conjunction with forecast
ƒ Calculated in the context of minimizing losses
subject to constraints
• Estimated for each business center
• Estimated for both Net & Base Campaign
• Serves as input to Optimal Assignment Timing

37
Copyright © 2007, SAS Institute Inc. All rights reserved.

Optimal Collections Staffing

Losses
Optimal Collections Staffing
L1

Total FTE Cost

L2

0 FTE1 FTE’

FTEs per 10,000 Accounts to be Worked


38
Copyright © 2007, SAS Institute Inc. All rights reserved.

Copyright © 2007, SAS Institute Inc. All rights reserved. 19


Optimal Collections Staffing

ƒ Optimization performed with PROC AUTOREG


• Augments OLS with autoregressive term of error
− Used when estimates are continuously above (below) trend
• Take first derivative of quadratic function and set to zero
− f(x) = ax2+bx +c
− f’(x) = 2ax +b =0 Æ x= -b /2a where x=Optimal FTEs

et = −eλt − 1 − eλt − 2 − ... − eλt − m + v

39
Copyright © 2007, SAS Institute Inc. All rights reserved.

Conclusions

ƒ Diverse analytical methods can be effectively


integrated to meet complex business needs
ƒ Integration enables a greater number of needs
to be addressed
ƒ Don’t hesitate to expand into different analytical
subject areas
ƒ Effective data management and programming is
the foundation for all analysis
ƒ SAS enables diverse analytics

40
Copyright © 2007, SAS Institute Inc. All rights reserved.

Copyright © 2007, SAS Institute Inc. All rights reserved. 20


Thank you!

Gene Grabowski, Jr.


Operations Analytics Supervisor
Ford Motor Credit Company
ggrabow1@ford.com
734-347-2430

Special thanks to:


Alfredo Cateriano
Ishita Gupta

41
Copyright © 2007, SAS Institute Inc. All rights reserved.

42
Copyright © 2007, SAS Institute Inc. All rights reserved.

Copyright © 2007, SAS Institute Inc. All rights reserved. 21

You might also like