1

Recent Trends in Credit Scoring Technology

Credit Scoring and Credit Control VIISeptember 5-7, 2001

Vijay Desai, Principal ScientistHNC Software

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

3

History of HNC and ProfitMax

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

5

Representative Financial Customers

6

History of Analytic Innovation

Fraud Management•Falcon, eFalcon

Application Decision Management•Capstone•Capstone Online•4Score

Marketing Optimization•Fee Enhancement•Pricing Optimization•Cross sell Optimization

Risk Management•ProfitMax•ProfitMaxBankruptcy•Strategy Manager

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

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

ProfitMax Profitability ComponentsRevenue model Expected revenue if account does not attrite and

does not fall into collection

Credit risk model Expected loss due to failure to payCombines probability of loss with amount

Attrition risk modelExpected loss of revenue due to a sharp and lasting reduction in balance and activity

Profit Calculator ties results together

Cost computationExpected operation & funding costs, using expense parameters & predicted transact, revolve, & delinquency behaviors

• Graphical user interface• Integrated testing and versioning capabilities • Reports facility

10

Profitability Overview

Risk Adjusted RevenueRisk Adjusted Revenue

(Cost of Funds + Operation costs)

(Cost of Funds + Operation costs)

less

equalsProfit ForecastProfit Forecast

(Forecasted Revolving Revenue + Fees +

Forecasted Interchange Revenue)

(Forecasted Revolving Revenue + Fees +

Forecasted Interchange Revenue)

(Credit Risk Adjustment + Attrition Risk Adjustment)(Credit Risk Adjustment + Attrition Risk Adjustment)

less

equalsRisk Adjusted RevenueRisk Adjusted Revenue

11

Overview of HNC Technologies

•Neural Networks•Context Vectors•Score Fusion

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

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.

14

Enhancement with Individual Profiles

Profile

Input

Input

Input

Input

Input

Input

Input

Input

Input

Input

FD

FD

FD

FD

FD

FD

PROFILE

Output

15

Use of Profiles in a Neural Engine

Transaction Transaction Profile

Feature Detectors

Updated Profile

NeuralNetwork

ProfileStore

Output Score

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

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

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-endShopper

Medicine Spender

Family Person

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

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

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

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

23

Transaction Based Profiles in Credit Scoring

24

High

High

John Doe’s Profile

Portfolio’s Aggregate Profile

# C

ash

Adv

ance

s

Revolving Balance AmountLow

Transaction-based Neural Network Scoring…Profile Historical Cardholder Activity

25

…Recognize Shifts in Individual Cardholder Behavior

# C

ash

Adv

ance

s

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

26

Understand Events in Context

Cardholder B’s ProfileCardholder B’s Profile

Cardholder A’s ProfileCardholder A’s Profile

Cardholder D’s ProfileCardholder D’s Profile

# C

ash

Adv

ance

s#

Cas

h A

dvan

ces

Revolving Balance AmountRevolving Balance AmountLowLow

HighHighCardholder C’s ProfileCardholder C’s Profile

HighHigh

27

Current Profile

Cardholder B’s Profile

Cardholder A’s Profile

Cardholder C’s Profile

Cardholder D’s Profile

# C

ash

Adv

ance

s

Revolving Balance AmountLow

High

...Understand Events in Context

High

28

Transaction Data Contain a Wealth of InformationLooking only at summarized data, these two cardholders appear to have similar risk

SummarizedData

Cardholder#1

Cash Advance

Merchandise

Cardholder#2

29

Timing of Transactions is Vital InformationNote the difference in apparent risk when transaction data are examined

Transaction DataSummarizedData

Cardholder#1

Cash Advance

Merchandise

Cardholder#2

30

More Timing of Transactions

Cash AdvanceTransaction DataSummarizedData

Merchandise

PaymentMonth #1

No payment? Is this a risky account or just a late payer?

Month #2

31

Knowledge of Individual is KeyTransaction

Data

SummarizedData

Month #1

Cash Advance

Merchandise

PaymentMonth #2

Month #3

Using the transaction data can tell the difference. Past behavior is stored and utilized to make decisions.

Month #4

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

Ann

ual D

olla

r Ben

efit

per A

ccou

nt

Cycle-Cut Real-Time Daily Batch

Benefit by Scoring and Decision Points

33

Transaction Based Neural Net Score is Stable Over Time

Month 1 Month 2 Month 3 Month 6 Month 9%units Avg. Scr. %units Avg. Scr. %units Avg. Scr. %units Avg. Scr. %units Avg. Scr.

1st ProfitMax 100 545 79.5 543 74.2 543 68.8 544 66.1 537Quartile Monthly 100 553 60.1 552 58.2 551 63.1 556 60.7 5512nd ProfitMax 100 614 65.5 614 60.0 616 50.2 618 46.4 622Quartile Monthly 100 610 47.5 609 46.5 609 45.0 618 43.8 6203rd ProfitMax 100 654 62.7 657 55.7 659 44.0 663 40.7 667Quartile Monthly 100 657 48.2 645 47.0 645 35.2 654 31.8 6544th ProfitMax 100 707 84.2 709 78.4 712 71.9 715 66.8 717Quartile Monthly 100 707 77.1 688 76.8 688 69.3 691 67.7 681

34

Benefit Over Monthly Score Increases with Time

Worst 20% of AccountsDollar Based Performance Spread over Monthly Score

10 Months 12 Months 14 Months 19 MonthsC/O 31* 65 97 191BKO 44 45 43 96Ever 95+ 30 88 114 185Ever 65+ -14 22 39 100

* 31 basis points higher than Monthly score

35

Benefit Over Monthly Score Increases with Time (contd.)

Best 30% of Accounts

Dollar Based Performance Spread over Monthly Score10 Months 12 Months 14 Months 19 Months

C/O 6 11 15 32BKO 8 9 10 20Ever 95+ 13 27 35 51Ever 65+ 18 30 37 55

* 31 basis points lower than Monthly score

36

Credit Risk Model Performance

37

Transaction Data Lead to Better Risk Management and Marketing

John Doe Behavior Trend Assessment

This Month Month Ago Conclusion1. Summary

Data Scoring Balance $1000 No change in riskBalance $1000

2. Transaction-based Scoring

7995-Gambli6011-Cash Advance Change to

higher risk from 1 month ago

ng Charges

6011-Cash Advance 5411-Grocery Store 7361-Employment Service 6011-Cash Advance

Summaryunemployment,

live on card charges

Summarywork,family,stable

$200$300$150$150$100

$100

5651-Clothing5641-Toys5812-Family Restaurant5942-Bookstore7999-Travel5732-Electronics

$200$100$50$50$400$200

38

Risk Model Comparison: Overall Results

ProfitMax Monthly ScoreFALSE Positive Rate @ 5% Detection 0.3 2FALSE Positive Rate @ 10% Detection 1 4FALSE Positive Rate @ 20% Detection 5 11FALSE Positive Rate @ 30% Detection 15 27FALSE Positive Rate @ 40% Detection 31 54FALSE Positive Rate @ 50% Detection 60 100KS Statistic* 106 100

*K-S Statistic and False Positive Rates are scaled to Monthly Score = 100

39

Transaction Data Improves Predictive Power

0

10

20

30

40

50

60

70

80

0 2 4 6 8 10

% Goods

% B

ads

ProfitMax Monthly

40

Transaction Data Particularly Useful for Young Accounts

0

10

20

30

40

50

60

70

0 2 4 6 8 10

% Goods

% B

ads

ProfitMax Monthly

41

Score Fusion: Improvement in K-S

SCORE K-S

Monthly Score 100

HNC Score 106

Combined Score 111

42

Score Fusion: Improvement in High 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 %

2 0 % 3 0 % 4 0 % 5 0 %

P e rc e n t B a d s D e te c te d

Perc

ent I

mpr

ovem

ent i

n Fa

lse

Posi

tive

Rat

e

H N C S c o reC o m b in e d S c o re

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 %

9 0 % 8 0 % 7 0 % 6 0 % 5 0 %

P e rc e n t G o o d s D e te c te d

Perc

ent

Impr

ovem

ent

in

Fals

e N

egat

ive

Rat

e

H N C S c o re

C o m b in e d S c o re

44

Attrition Risk Model Performance

45

Transaction Attrition Risk Model With Profiles

Cardholder “A” Cardholder “B” Conclusion

1. Summary Data Scoring

Balance $6,000Value High

Balance $6,000Value High

No difference

2. Transaction -based scoring

5310-Discount Stores $2005698-Subscription $1007841-Videotape Rental $205411-Supermarket $505964-Catalog Merchant $4006011-ATM Cash $200

Different spending patterns reveal propensity to attrite.

SummaryHigh-Ticket Charges

Non-repeating charges,

T&E Spender

Summary“Card of Choice”

4722-Vacation$12005812-Restaurant $1503357-Car Rental $1505944-Jewelry $400

“More likely to attrite” “Less likely to attrite”

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

Perc

ent o

f Act

ual L

oss

to A

ttriti

on

Curbal*AttrProb Score

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 ttrit io n S c o re Q u in tile s

Perc

ent

% Acco u n ts

% B a lan ce

48

Revenue Models Performance

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

50

Revolving Balance

0

1000

2000

3000

4000

5000

6000

7000

8000

9000

1 2 3 4 5 6 7 8 9 10

51

Transaction Volume

0

100

200

300

400

500

600

700

800

900

1 2 3 4 5 6 7 8 9 10

52

Going Beyond the Credit Risk Score

•Credit risk score not the best measure of future profitability•Multi-dimensional view of customer vital

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

Cre

dit R

isk

Scor

e

% BadLow

High

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

Prof

itabi

lity

Pre

dict

ion

Avg. Profit % Bad

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

Why a Multi-dimensional View of Customers?

Cre

dit R

isk

Reve

n

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

57

Dream Customer

low

low

Cre

dit R

isk

Cre

dit R

isk

Reve

nRe

ven

Attrition RiskAttrition Risk

low

low

high

low

high high

Bad Rate: 0.00%Attrition Rate: 2.00%True Revenue: $342.12True Profit: $212.04

ueue

•Low Attrition•Low Credit Risk•High Revenue

58

Risk-Revenue Tradeoff:Balance Risk and Pricing

•Low Attrition•High Credit Risk•High Revenue

Cre

dit R

isk

Reve

n

Attrition Risk

low

low

high

low

high high

ue% of Accts: 1.00%Bad Rate: 35.40%Attrition Rate: 2.60%True Revenue: $387.65True Profit: ($198.99)

59

“Save Me” - Retention Program

Cre

dit R

isk

Reve

n

Attrition Risk

low

low

high

low

high high

% of Accts: 1.00%Bad Rate: 0.04%Attrition Rate: 20.00%True Revenue: $354.20True Profit: $222.55

ue

•High Attrition•Low Credit Risk•High Revenue

60

Light Revolvers -“Build Balance Program”

Cre

dit

Ris

k

Rev

enue

low

low

hig

h

low

high

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: $37.40True Profit: $19.00

Attrition Risk

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

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

ProfitMax and Basic SegmentationRisk Revenue Attrition Generic Strategy

Low Low Low DevelopLow High Low GrowLow Low High DevelopLow High High DefendHigh Low Low ExitHigh High Low MaintainHigh Low High ExitHigh High High Selectively Defend

64

ProfitMax and Advanced Segmentation

Generic High Risk AffinityRisk Revenue Attrition Strategy Strategy Strategy

Low Low Low Develop Low High Low GrowLow Low High DevelopLow High High DefendHigh Low Low Exit Develop MaintainHigh High Low Maintain GrowHigh Low High Exit Develop MaintainHigh High High Sel. Defend Defend Defend

65

Questions &Answers