Upload
vijay-desai
View
213
Download
0
Embed Size (px)
Citation preview
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