1. 1
Recent Trends in Credit Scoring Technology
Credit Scoring and Credit Control VII
September 5-7, 2001
Vijay Desai, Principal Scientist
HNC Software
2. 2
Agenda
• History of HNC and ProfitMax
• Overview of HNC Technologies
• Transaction Based Profiles in Credit Scoring
• Credit Risk, Attrition Risk, Revenue, and
Profit Score Performance
• Going Beyond the Credit Risk Score
• Questions and Answers
4. 4
Who is HNC Software?
• Founded 1986
• U.S. offices
– San Diego HQ
– Chicago
– Irvine
– Los Alamos
– Philadelphia
– New York
• International offices
– London
– Singapore
– Tokyo
• 1,000 + employees
• Initial Public Offering: June
1995
(NASDAQ symbol: HNCS)
• 2000 Revenues $254.9
million
San Diego Headquarters
7. 7
History of ProfitMax
• Developed in 1994
• Used by three of the top five credit card
issuers in USA
• More than 100 million accounts scored using
ProfitMax
• Provides real-time decision capability
• Provides a multi-dimensional view
8. 8
ProfitMax Uses HNC Technology
• HNC’s enhanced neural network technology
provides core predictive model capability
• Transaction data provide additional
information typically lost in summarization
• Models use account profiles
– Updated with every transaction
– Reflect entire account relationship
• Models and software are designed for
real-time decisions using the most recent
available information
9. 9
ProfitMax Profitability
Components
Revenue model Expected revenue if account does not attrite and
does not fall into collection
Credit risk model Expected loss due to failure to pay
Combines probability of loss with amount
Attrition risk model
Expected loss of revenue due to a sharp and lasting
reduction in balance and activity
Profit Calculator ties results together
Cost computation
Expected operation & funding costs, using expense
parameters & predicted transact, revolve, &
delinquency behaviors
• Graphical user interface
• Integrated testing and versioning capabilities
• Reports facility
12. 12
HNC Solutions in Network
Building...
• Architecture Settled by ANN Tournament System
• Automatic Training with Progressive Testing --
Prevents Over/Under Training
• Fast, HNC Hardware for Fine-Grain Learning
• Input Variable Selection Techniques:
– Proprietary Variable Clustering Method,
– Statistical Analysis of Partial Derivatives,
– Patented, Specific-Case Explanation Facility.
13. 13
Enhancement with Individual
Profiles
• Allows Network to Recognize Change
• Allows for Detection of Trends and Deviations From
Trend
• Allows Full Use of Event and Interval Information
• Avoids Wasteful Aggregation Into Fixed Periods
• Allows Network to Assess Events in Proper Context
• Changes Focus of Training Process From Unrelated
Events to Evolving Patterns.
15. 15
Use of Profiles in a Neural
Engine
Transaction Transaction Profile
Feature Detectors
Updated Profile
Neural
Network
Profile
Store
Output Score
16. 16
Mining Merchant Names for
Credit Risk Using Context Vector
• Observation:
– People who shop at Cartier are more likely to
shop at Tiffany, and very unlikely to shop at
GoodWill.
– People with ComCheck transaction have much
higher credit risk than people who shop at Ethan
Allen.
• Challenge:
– How to extract such information from credit card
transactions and incorporate it into credit risk
models to enhance the prediction?
17. 17
Context VectorTM
• Context Vector:
– High-dimensional representation
– Trained using co-occurrence
statistics ==> vectors are close if
they co-occur more often than
what is expected.
• Merchant Vector:
– Context Vector representation of
the merchant names.
– Neighboring merchant-vectors
==> tend to be shopped at by the
same people.
Context Vector™ Space of
Merchant Names
Context Vector™ Space of
Merchant Names
tiffanytiffany
AnnTaylorAnnTaylor
ComChekComChek
GoodWill
Harrah Casino
SalvationArmy EthanAllen
Cartier
ThriftyDrug
KaiserPharm
18. 18
Context Vector -- Cont.
• Question:
– What about people who go to casino
only once in a while, and usually
shop at high-end stores?
– Verses people who pay a lot of
medical bills and take greyhound?
• Account Vector:
– Aggregation of the merchant
vectors the account shops at.
– Represent the transactional
behavior over a certain period of
time.
– It migrates to different direction
when the behavior changes.
– Can be associated with different
credit risk.
Context Vector™ Space of
Accounts
Context Vector™ Space of
Accounts
Mktng-skill
Reading
Subscriber
Mktng-skill
Reading
Subscriber
Gambler
Thrifty
Shopper
High-end
Shopper
Medicine Spender
Family Person
19. 19
Score Fusion
• Issuers Have Access to Multiple Scores Today
• In the Credit Risk Management Area,
– Issuers Have Multiple Custom Scores Based Upon:
• Transaction Data
• Master File
• Credit Bureau
– Credit Bureau Score (Credit Bureau data Based)
– Custom Credit Risk Scores (Master File or Master File &
Transaction Based)
– Custom Bankruptcy Scores (Master File or Master File &
Transaction Based)
20. 20
How to Best Use Multiple Data
Sources?
• Combine all sources of data to build a “Super Duper”
score (1)
• Use existing scores as inputs to build new scores (2)
• Combine existing scores and make a “Super Duper” Score
(3)
• Literature suggests that approach 1 is as good or better
than 2 and approach 2 is as good or better than 3
• Organizational constraints make approach 3 the most
feasible solution today
21. 21
Organizational Constraints
• Reasons for not combining all sources of data
– Credit-bureau data used only on as-needed basis
– Processing system constraints
– Reluctance to shelve scores built with significant
effort
• Reasons for not using existing scores as
inputs
– Loss of flexibility
– Credit-bureau scores used only on as-needed basis
– Processing system constraints
22. 22
Score Fusion - HNC’s Approach
• The HNC Algorithm:
– is not dependent on a functional form
– makes sure that there are enough accounts in each
score combination
– is better at identifying non linear patterns in risk
– gives the best estimate of risk at each score
combination
– is a multidimensional approach
– is very easy to implement
25. 25
…Recognize Shifts in Individual
Cardholder Behavior
#CashAdvances
Revolving Balance Amount
High
Low High
John Doe’s Profile Two Weeks Ago
John Doe’s Profile Last Week
John Doe’s Profile This Week
27. 27
Current Profile
Cardholder B’s Profile
Cardholder A’s Profile
Cardholder C’s Profile
Cardholder D’s Profile
#CashAdvances
Revolving Balance AmountLow
High
...Understand Events in Context
High
28. 28
Transaction Data Contain a
Wealth of Information
Looking only at summarized data, these two cardholders
appear to have similar risk
Summarized
Data
Cardholder
#1
Cash Advance
Merchandise
Cardholder
#2
29. 29
Timing of Transactions is Vital
Information
Note the difference in apparent risk when transaction data are examined
Transaction DataSummarized
Data
Cardholder
#1
Cash Advance
Merchandise
Cardholder
#2
30. 30
More Timing of Transactions
Cash AdvanceTransaction DataSummarized
Data
Merchandise
PaymentMonth #1
No payment? Is this a
risky account or just a
late payer?
Month #2
31. 31
Knowledge of Individual is Key
Transaction
Data
Summarized
Data
Month #1
Cash Advance
Merchandise
Payment
Month #2
Month #3
Using the transaction
data can tell the
difference. Past
behavior is stored and
utilized to make
decisions.
Month #4
32. 32
Comparison of Dollar Savings
$0.00
$1.00
$2.00
$3.00
$4.00
$5.00
$6.00
0.50% 1.00% 1.50% 2.00% 2.50% 3.00% 3.50% 4.00%
Percentage of Accts W orked per Month
AnnualDollarBenefitperAccount
Cycle-Cut Real-Time Daily Batch
Benefit by Scoring and Decision
Points
37. 37
Transaction Data Lead to Better
Risk Management and Marketing
John Doe Behavior Trend Assessment
This Month Month Ago Conclusion
1. Summary
Data Scoring Balance $1000 No change in riskBalance $1000
2. Transaction-
based Scoring
7995-Gambli
6011-Cash Advance
Change to
higher risk from
1 month ago
ng Charges
6011-Cash Advance
5411-Grocery Store
7361-Employment
Service
6011-Cash Advance
Summary
unemployment,
live on card charges
Summary
work,
family,
stable
$200
$300
$150
$150
$100
$100
5651-Clothing
5641-Toys
5812-Family Restaurant
5942-Bookstore
7999-Travel
5732-Electronics
$200
$100
$50
$50
$400
$200
42. 42
Score Fusion: Improvement in
High Risk Accounts
0 .00 %
4 .00 %
8 .00 %
12 .00 %
16 .00 %
20 .00 %
24 .00 %
28 .00 %
32 .00 %
36 .00 %
40 .00 %
2 0% 3 0 % 4 0% 5 0 %
P e rc e n t B a d s D e te c te d
PercentImprovementinFalsePositiveRate
H N C S co re
C o m b in e d S c o re
43. 43
Score Fusion: Improvement in
Low Risk Accounts
0 .0 0 %
4 .0 0 %
8 .0 0 %
1 2 .0 0 %
1 6 .0 0 %
2 0 .0 0 %
2 4 .0 0 %
2 8 .0 0 %
3 2 .0 0 %
3 6 .0 0 %
4 0 .0 0 %
4 4 .0 0 %
4 8 .0 0 %
90 % 80 % 7 0% 6 0% 5 0%
P e rc e n t G o o d s D e te c te d
PercentImprovementinFalseNegativeRate
H N C S co re
C o m b in e d S co re
45. 45
Transaction Attrition Risk Model
With Profiles
Cardholder “A” Cardholder “B” Conclusion
1. Summary
Data
Scoring
Balance $6,000
Value High
Balance $6,000
Value High
No difference
2. Transaction
-based
scoring
5310-Discount Stores $200
5698-Subscription $100
7841-Videotape Rental $20
5411-Supermarket $50
5964-Catalog Merchant $400
6011-ATM Cash $200
Different
spending
patterns reveal
propensity to
attrite.
Summary
High-Ticket Charges
Non-repeating
charges,
T&E Spender
Summary
“Card of Choice”
4722-Vacation$1200
5812-Restaurant $150
3357-Car Rental $150
5944-Jewelry $400
“More likely to attrite” “Less likely to attrite”
46. 46
Balance at Risk is a Better
Predictor than Attrition Score by
Itself
0
10
20
30
40
50
60
70
80
90
100
1
7
13
19
25
31
37
43
49
55
61
67
73
79
85
91
97
Ranking According to Two Schemes
PercentofActualLosstoAttrition
Curbal*AttrProb Score
47. 47
Attrition Score Rank Orders
Attrition Balance Very Well
0
1 0
2 0
3 0
4 0
5 0
6 0
7 0
1 2 3 4 5
A ttritio n S c o re Q u in tile s
Percent
% Acco u n ts
% B alan ce
49. 49
Transaction Data In Revenue
Models
• Transaction data important in prediction of future
revenue
• Spending in certain SIC categories a better indicator
of continued future revenue than others; e.g. Home
related spending:
• 5411 - Grocery Stores, Supermarkets
• 4900 - Utilities - Electric, Gas, Sanitary & Water
• 5200 - Home Supply Warehouse
• 5211 - Lumber Services
• 0742 - Veterinary Services
• 7311 - Furniture & Tool Rental
• 5231 - Paint & Wall Paper Stores
52. 52
Going Beyond the Credit Risk Score
•Credit risk score not the best measure of
future profitability
•Multi-dimensional view of customer vital
53. 53
Managing Profitability:
Using Credit Risk Predictions
Portfolio organized by Credit Risk score
0 100 200 300 400 500 600
Avg. Profit
-50 0 50 100 150 200 250 300 350
CreditRiskScore
% Bad
Low
High
54. 54
Managing Profitability:
Using Credit Risk Predictions
Portfolio organized by HNC ProfitMax prediction
-50 0 50 100 150 200 250 300 350 50 70 90 110 130 150
Low
High
ProfitabilityPrediction
Avg. Profit % Bad
55. 55
A Better Measure of Future
profitability
-50 0 50 100 150 200 250 300 350-50 0 50 100 150 200 250 300 350
Organized by Credit Risk Score Organized by HNC Profit Score
Low
High
56. 56
Why a Multi-dimensional
View of Customers?
CreditRisk
Reven
Attrition Risk
low
low
high
low
high
high
• Small groups of
cardholders have a big
impact on bottom line
• Treatments are
profitably applied when
taken on select
cardholder groups
ue
58. 58
Risk-Revenue Tradeoff:
Balance Risk and Pricing
•Low Attrition
•High Credit Risk
•High Revenue
CreditRisk
Reven
Attrition Risk
low
low
high
low
high
high
ue
% of Accts: 1.00%
Bad Rate: 35.40%
Attrition Rate: 2.60%
True Revenue: $387.65
True Profit: ($198.99)
59. 59
“Save Me” - Retention Program
CreditRisk
Reven
Attrition Risk
low
low
high
low
high
high
% of Accts: 1.00%
Bad Rate: 0.04%
Attrition Rate: 20.00%
True Revenue: $354.20
True Profit: $222.55
ue
•High Attrition
•Low Credit Risk
•High Revenue
60. 60
Light Revolvers -
“Build Balance Program”
CreditRisk
Revenue
low
low
high
lowhigh
high
• Low Credit Risk
• Low Attrition
• Medium Revenue
• Low Credit Risk
• Low Attrition
• Medium Revenue
% of Accts: 3.00%
Bad Rate: 0.10%
Attrition Rate: 2.40%
True Revenue: $37.40
True Profit: $19.00
Attrition Risk
61. 61
Most Common Uses
• Profit Score
– Retention Queue
– Segmentation Variable
• Credit Risk Score:
– Minimal Acceptable level of risk
– “Knock-out” criteria for Marketing Programs
• Revenue Score:
– Needs based approach to credit line increases
– Revenue reason codes used for targeting
• Attrition Score
– Proactive Retention offers
– Profit@Risk metric
62. 62
Additional ProfitMax Benefits
Will Also Come From...
• Revenue Score
– Tracking and Reporting - trending over time
– Testing Validation - incremental change over control
position
• Profitability Score
– Prioritizing calls in VRU (best customers wait less)
– Making it more difficult for unprofitable accounts to talk to a
live representative
– Reactive retention queue has on-line access to profit number;
can tailor save offer
– Best customers routed to best offer
63. 63
ProfitMax and Basic
Segmentation
Risk Revenue Attrition Generic Strategy
Low Low Low Develop
Low High Low Grow
Low Low High Develop
Low High High Defend
High Low Low Exit
High High Low Maintain
High Low High Exit
High High High Selectively Defend
64. 64
ProfitMax and Advanced
Segmentation
Generic High Risk Affinity
Risk Revenue Attrition Strategy Strategy Strategy
Low Low Low Develop
Low High Low Grow
Low Low High Develop
Low High High Defend
High Low Low Exit Develop Maintain
High High Low Maintain Grow
High Low High Exit Develop Maintain
High High High Sel. Defend Defend Defend
