Population Identification
• The entire portfolio population is usually too
heterogeneous for a single model
• The population is segmented either based on
some broad criteria (e.g. based on age groups:
18-25, 25-45, 50-65, >65) or empirically, to
segments with widely different good / bad odds
• Scorecards need to be built for each segment
• Availability of data (both credit bureau data and
internal data) needs to be considered
Observation / Performance Timeline
Observation
Snapshot
Performance
Snapshot
PAST
Observation
Period
PRESENT
FUTURE
Performance
Period
Good-bad Definition
• The accounts are classed as ‘good’ / ‘bad’ /
‘indeterminate’ based on performance during the
performance period. For example:
• Bad = bankrupt or 3 or more payments missed
within 9 months
• Indeterminate = ever 2 payments missed or
always inactive or very low balance during 9
month observation period
• Good = always up-to-date payments or ever 1
payment missed
Good-bad evolution
Good-Bad Evolution
Cum. Bad Rate
120.00%
100.00%
80.00%
69% of accounts that go
bad in 18 months can be
identified after 9 months
60.00%
40.00%
20.00%
0.00%
1
2 3
4
5
6
7 8
9 10 11 12 13 14 15 16 17 18 19
Months since Observation
Characteristic Selection
•
Not all predictive characteristics are used in the model.
– An inter-correlation effect may exist between variables.
– For example, age may be correlated with time at current employment
and therefore only one is necessary in the model.
•
Some credit bureau characteristics and some internal ones are
selected based on their ‘marginal contribution’ to the outcome, and
‘monotonicity’ of the odds:
12.00
70
60
10.00
50
40
6.00
30
Scores
Good-Bad Odds
8.00
G/B Odds
Scores
4.00
20
2.00
10
0.00
0
1-50
50-70
70-80
80-85
85-90
90-95
95-97
97-100
100-101
Other
Bands
10/30/2013
Ratio of current balance to maximum lifetime balance (all credit cards)
35
Model Build
•
•
Once the characteristics have been selected a
statistical model can be developed.
Multivariate statistical methods include
– Linear Regression
– Logistic Regression
– Heuristics like Decision Tree / Neural Network
10/30/2013
36
Model Build
n
The model is built on dichotomous data. In this case a 1 for “Good”
customers and a 0 for “Bad” customers.
1
0.8
0.6
0.4
0.2
0
0
200
400
600
800
1000
Logistic Regression
n
The logistic regression fits the probability better than Linear
regression.
1
0.8
0.6
0.4
Good/Bad Probability
0.2
Logistic
Linear (Good/Bad Probability)
0
0
200
400
600
800
1000
Models
•
Logistic Regression has the following form:
p
k
= ∑ j =0 β j x j
ln
1− p
p=
(
exp ∑ j= 0 β j x j
k
(
)
1 + exp ∑ j = 0 β j x j
k
)
Reject Inference and Validation
•
Reject Inference
– Reject Inference is necessary for application scorecards
because there is no performance information for the
rejected applications
• Applications that are rejected should be included in the final
model
– Behavioural scorecards deal only in existing customers,
therefore do not require reject inference.
•
Validation
– A randomly selected control group (hold out sample) or
proxy portfolio is used to test the model.
Scoring
System A
% APPLICANTS
Comparison of Scoring Systems
Cutoff
Bads
Goods
10%
Scoring
System B
% APPLICANTS
5%
Cutoff
Bads
Goods
20%
5%
SCORE
Measures of Discrimination
Scorecard performance can be judged on the level of discrimination
• Two measure that can be used are:
Gini (or ROC) – the area between Lorenz Curve and random line
PH - % of Goods below 50% of bads
Lorenz Curve
1
0.9
Cumulative Bads
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.2
0.4
0.6
10%
Cumulative Goods
0.8
1
Gini = 62%
Advantages of Scoring
– Defines degree of credit risk for
each applicant
– Ranks risk relative to other applicants
– Allows decisions based on degree of risk
– Enables tracking of performance
over time
– Permits known and measurable adjustments
– Permits decision automation
but BI in Consumer Lending goes much beyond ...
Maximizing Lifetime Customer Value
Focus on Value of Customer to Organization - Lifetime Customer Value
First Premier NSF Model Seg1 : NSF last 30 days
Score comparisons
100%
Cohort or
“Segments”
• Groups of
accounts with a
similar profile
Predictive
Models
• Collapses
information into a
Score to rank-order
populations based
on account
characteristics
Strategies
• Matrix of Segments
and Scores to
increase granularity
of decisions
Optimization
• Combine predictions –
risk, revenue - into a
single metric or objective
• Assign an optimal action
for each account while
satisfying business
constraints
Evolution of Methods
Optimization
Behavioral Scoring
Risk Scoring
Criteria Based Rules
• Little / No data
• Criteria-based
rules
• Decision impacts
not understood
Increasing Customer View (Data and Models) and Competitive Pressures
Credit Limit & Pricing Optimization using
Action-Effect Models
Inputs
Target
Behaviour
Score
Actions
Credit Limit
Increase /
Decrease
Action-Effect Models
P(Attrition)
Expected
Revenue
P (Default)
Utilization
PROFIT
Change in
Balance
Propensity
to Revolve
Credit Price
Increase /
Decrease
LGD
Expected
Loss
Action-Effects Models
Action-Effects models are:
• A prediction of customer reaction to an action taken
• What your models look like depend on your action set as well as what
information is known about the customer
• Action-Effects models are sensitive to actions (traditional predictive
models aim to be robust over possible actions)
Example: How does a Credit Limit Action affect customers’
behaviour?
Example: Action-Effect Model of Attrition
Given Credit-Limit Change
Attrition Rate . Conseq Months Revolver L12 Months
0.25%
>2 Months
<2 Months
0.20%
Attrition Rate
Extrapolation
No Historical Inf :
Extrapolation
0.15%
0.10%
Action effected the
behaviour
0.05%
0.00%
£500
£1,000
£1,500
£2,000
CL Action
£2,500
£3,000
£3,500
Improved portfolio performance
Low
Cutoffs
Optimized
Scores
High
Cutoffs
Optimized
Scores
E[Volume]
Low
Cutoffs
Optimized
Scores
E[Profit]
E[Loss]
Single Score
Single Score
High
Cutoffs
Low
Cutoffs
E[Loss]
E[Volume]
Single Score
Efficient Frontier
High
Cutoffs
E[Profit]
Value of Strategy Optimization
Example users
Strategy Optimization is delivering
outstanding results to over 25
world-class organisations, driving
true competitive advantage
Some Caselets Follow
New Business : Initial Credit Limit
Assignment
Challenge
• Determine the most profitable initial credit card limits
• Must not increase overall exposure or bad debt when
compared to existing rule based strategy
Leading
Retail Bank
Solution
• Optimization used in real-time during credit card application process
• Identifies which customers receive which initial limit*
Adaptive control rules
Customer level optimization
Benefit
Revenues
£111.2M
Revenues
£118.2M
+ 6%
Credit
Losses
£48.4M
Credit
Losses
£46.3M
- 4%
Other
Costs
£52.2M
Other
Costs
£51.2M
- 2%
Profit
£10.6M
Profit
£20.7M
+ 96%
*Results normalised to a customer base of 1 million
Account Management : Limit Increase
Challenge
• Determine the best increase in credit limit to maximise
customer profitability across a 10 million customer base
• 5 different percentage-based increases to consider
• Must not increase overall exposure or bad debt when
compared to existing strategy
Leading
Credit Card Issuer
• Implemented through tree based logic
Solution
• Optimization used during end-of-month statement process
• Identifies which groups of customers receive an increase in credit limit and by how much
• Implemented through a refined strategy tree
• Generates incremental profit of £3 per account per annum when optimizing tree-based
strategies (estimated at £7 per account per annum when optimizing individual
customer strategies)
Limit Increase : Large Retailer
Challeng
e
• 5 million retail store card customers
• Determine best limit increase for each eligible
customer to maximise overall retail and credit
profitability
• Must not increase overall exposure or bad debt when
compared to existing strategy
• Existing champion strategy fine-tuned over many years
Solution
• Optimal limit applied during end-ofmonth billing process
• Considers off line retail transaction
data to predict seasonal card usage
Benefit
• Generates incremental profit of 7% over
the existing Champion strategy for same
bad debt and exposure levels
• Now considering optimal timing of
increases to coincide with seasonal
patterns in retail category spend
New Business : Loan Pricing
Challenge
• Improve new personal loan customer profitability at
point of sale
• 7 different APR rates to consider
Solution
Leading UK
Personal Finance
Lender
• Optimization used to determine the optimal price (APR) to offer new customers
• Optimal decision based on:
• Deal profitability, Propensity to take up the offer and Credit risk losses
• Optimization applied dynamically at the individual customer level to maximise
decision performance
• 13% increase in profit contribution for the same lending amount bad debt
value
Contact Planning : Top 10 UK
retail bank
Challenge
• Shifting focus: from campaign-centric marketing to customer
data-driven forecasting of outbound contacts
• Product maturities, renewals, end of term and other cross sell
opportunities;
• Exploit a wealth of customer information to drive timely,
appropriate and profitable customer contacts
Solution
Benefit
• Optimization used to schedule million’s
of customer contacts over a financial
year
• 20% increase in number of customer
contact opportunities in the financial
year;
• Adheres to strict customer contact
frequency policies
• Improved predictability of sales
volumes, budgets and revenues over
time;
• Identifies best opportunities to meet
financial year budget, sales and revenue
expectations across product marketing
units
• Finds contact opportunities missed by
previous approaches
• Highlights gaps between customer
needs and the supporting propositions
that could be delivered;
• Scenario planning process reduced
from days to less than an hour;
Cross Sell : Top 5 UK retail bank
Challenge
Benefit
Business as usual
• 10 million customers
£6.66M
£6.08M
1.969M
• Cross selling within personal
finance customer base:
• Loans, Cards and Cheque
Accounts
Offers
Sales
- 3%
+ 19%
• Must not increase overall budget
spend or outbound channel usage
• Maximise Net present Value
Solution
£572K
14,202
Revenue
s
Profits
Budget
+ 32% + 35% - 5%
Optimization
£8.80M
£8.25M
1.914M
• Optimization used to select offers for
customers in monthly telemarketing
and direct mail campaigns
• Fully utilises available models