Business Intelligence & Data Mining-5

Published on February 2017 | Categories: Documents | Downloads: 36 | Comments: 0 | Views: 356
of 26
Download PDF   Embed   Report

Comments

Content

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%

Cumulative % of No Cure Accounts

90%

Perfect
80%

Random
70%

Building

60%
50%

Validation

40%

Total

30%

Fico

20%

Behavior
score

10%
0%
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100
%
Cumulative % of Cure Accounts

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

1985

• Enhanced
Strategies

• Simple Strategies

• Behavioral Risk

• Decision platforms

• Response

• Profiling and
segmentation

• Revenue

• Champion /
Challenger Testing

• CRM Platforms

1990

• Attrition

1995

• Optimized
Strategies
• Optimization
Engines
• Acct-Level Profit
• Acct-Level Actions
• In-Market (Closedloop) Testing

2000

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

16,886
Offers

Sales

£543K
Revenue
s

Profits

Budget

Sponsor Documents

Or use your account on DocShare.tips

Hide

Forgot your password?

Or register your new account on DocShare.tips

Hide

Lost your password? Please enter your email address. You will receive a link to create a new password.

Back to log-in

Close