Modeling and Segmentation

1
Modeling and Segmentation

• Telecommunications Industry 2007
• GSU-MGS8040

2
Presentation Subtopics

• Telecom History
• Scope of Presentation
• Modeling
• Scoring Tracking
• Segmentation
• Whats Next?

3
Telecom History
4
Telecom History

• Pre-divestiture ATT
• Little innovation
• No competition
• No price pressure
• Divestiture 1974-1982
• USDoJ split ATT in return for entry into
  computers
• ATT split into 7 Regional Bell Operating
  Companies (RBOC)
• Ameritech Corporation
• Bell Atlantic Corporation
• BellSouth Corporation
• NYNEX Corporation
• Pacific Telesis Group
• Southwestern Bell Corporation
• U S West, Inc.

5
History (continued)

• Divestiture 1974-1982 (continued)
• Surge in long distance competition
• Sprint, MCI, ATT, BellSouth, Verizon, Quest
• LD prices drop
• Local monopolies remained
• local prices rise/static
• Telecommunications Act 1996
• State-by-state ? Uniform national law
• Meant to promote competition
• Incumbent Local Exchange Carriers (ILECs) made
  network elements available to Competitive LECs
  (CLECs) at cost plus regulated wholesale
• LECs gained ability to provide LD services
• Lead to consolidation of major media companies
  (80 gt 5)

6
Evolution of Telecom Companies
From Wikipedia
7
New Competitive Challenges

• New Technologies - Convergence
• Cellular Phone Messaging, E-mail, Ring Tones,
  TV/Video feeds
• Wireless Communication/Data
• VoIP
• Internet Access
• ISDN, DSL, T1
• Cable
• Cable/Wireless partnerships
• Television/Video (new)
• Bundle strategies

8
Presentation Scope
9
Presentation Scope

• Single ILEC providing B2B landline products and
  services
• 1.2M business customers, 2.4M lines
• 1 - 200 employees
• 1 - 50 lines
• 1 - 10 locations
• Top 5 industries Retail, Wholesale, Business
  Services, Manufacturing, Healthcare
• ILEC uses a three channel approach to the market
  including Inbound centers, Outbound sales and
  Sales Agents.

10
Modeling
11
Why Model

• Increase Profitability
• Ameliorate line losses
• CLEC competition
• Cellular
• Sales targeting outbound and Inbound, based on
  customer behavior/attributes
• New product development and advertising
  strategies
• Efficient use of marketing and sales resources
• Segmentation Strategies Identify groups of
  customers based on predictions of their possible
  business needs

12
Line Loss History
13
Line Loss History
14
Telecom Modeling

• Statistical propensity modeling is the backbone
  of telecom segmentation and offer strategy
• Every customer is scored by each model
  (probability and L, M, H score)
• Models have been built and continuously updated
  for all key products (Bundles, DSL, Lines, Line
  Add-ons, LD, T1, Direct Internet Access, complex
  data, complex voice, wireless, hosting, inert
  customers, customer vulnerability/churn, and
  growth index)
• Predominantly logistic regression models - 70
  variables initially, with 5-10 in the final model
• Sales improvement from the use of models varies
  from 20-50, over no targeting

15
Automated Data Sourcing/Flow
Sales Quotas and Targets
Billing
Modeling Reporting Datamart
List Generation
Product Usage

• Automated Acquisition
• Unit of Analysis
• Matching
• Cleaning
• Conflict Resolution
• Business Rules
• History
• Summarize
• Calculated Variables

Targeting
Service, Maintenance
Advertising Sales Campaigns
Tracking
Monthly Processing
Trouble Reports
New Product Strategy
Campaign Tracking
Reporting Scheduled, Ad hoc
Contracts
Data Views
Modeling Scoring
3rd Party - DB, InfoUSA
Scores, Segments
16
Modeling Scoring Flow
Store, Clean, Dummy variables, Categorize,
Standardize, Calculate new variables, Summarize
Modeling Reporting Datamart
Views
SAS Enterprise Miner
Insert
Refresh Models, New Models, Ad hoc Models
Score Customers Monthly
17
Data for Modeling

• Snapshot of customer data for the most current
  month
• Total of 350-400 variables
• Customer history (3-6 months) for some variables
• Aggregated with summary functions (mean, min,
  max, etc.)
• Data cleaning
• Null, 0, Missing, Blanks
• Impute
• Bad values (out of range, wrong type,
  subjectivity)
• Outliers
• Transformations
• Offsets
• Calculated variables
• Other pre-processing decision trees, factor
  analysis, etc.
• SAS Enterprise Miner

18
SAS Modeling Interface
19
Dataset Drill-Down
Variable labels intentionally covered
20
Logistic Drill-Down
21
Neural Net Drill-Down
22
Model Flow - Sample
23
Logistic Results Drill-down (Confusion Matrix)
24
Logistic Results Drill-down (T-scores)
25
Cumulative Response (Lift)
26
Scoring
27
Automated Scoring

• Score 1.2M customers for each of 25 models x 2
  variants/model x 1-4 updates/refreshes per year gt
  120 models/year
• Customers scored with 2 values probability
  (0.0-1.0) score (L, M, H) for each
  model/variant
• SAS code (32,354 lines ) - modularized,
  optimized for ease of maintenance and to some
  degree, speed
• Declare global macro variables
• Date
• Product mean revenue
• Declare Libnames
• Establish OLEDB connection with remote database
  (SQL Server 2005)
• Connection/references to local subdirectories
• Code
• Raw Data
• Scores
• Prep for new data delete datasets from previous
  months processing
• Retrieve data
• Connect to views and read data from remote server
  into local datasets
• Clean data, create calculated variables
• Launch scoring modules
• Score customers for 50 models
• Store scores locally

28
Scoring Process (include files)
Model 1 Scoring Code File
Master File SAS Pseudo-Code
Data scores.model1
Pre-scoring Code
set raw_data.cust

Model 1 Scores
run
Model 2 Scores
SAS Processing Flow
Modeling Platform
Model 3 Scores
Model 2 Scoring Code File

Model N Scores
Data scores.model2
set raw_data.cust
Post-scoring Code

run
include code.Score_Model_1.sas
29
Probability/Propensity vs Score
Score Abbreviation Probability Range Population Size
High H 0.50 H 1.00 20
Medium M 0.25 M 0.75 30
Low L 0.00 L lt 0.50 50
30
Tracking Model Effectiveness

• Monthly tracking with updating as needed
• Effectiveness Index (EI) actual sales compared
  to average sales rate
• EI multiplier showing how effective the model
  is. E.g. Product B model shows that a customer
  scored high is 3 times more likely to buy the
  product than an average customer
• Model differentiation compare High vs Low EI
  values. E.g. For Products C-E, a customer scored
  high is more than 7 times more likely to buy
  that product than one scored low

31
Model Performance Improvement - Refresh
32
Segmentation
33
Why Segment

• Increase Profitability
• Targeting
• efficient use of marketing and sales resources by
  targeting inbound and outbound sales
• Messaging
• development of targeted marketing communications
  (i.e., Hispanic language direct mail, women owned
  businesses) ensures messages reaches customers
  effectively
• Future Needs
• Identification of groups of customers based on
  their business needs, not bound by traditional
  telecom products

34
Segmentation Evolution
The segmentation process was continually evolved
- moving from one dimensional models to multi
dimensional schemes. Along the way, predictive
modeling was added to the process to ensure the
segmentation scheme was always actionable.
Product Targeted
Vulnerability
Value
Industry

• B2B
• Technology
• Retail/Service
• Small Stable

• Seg 1
• Seg 2
• Seg 3
• Seg 4
• Seg 5
• Seg 6

High
Customer Complexity
Vulnerability
Low
Customer Size
Location
One Dimensional
Multi Dimensional
1997 2001 2006
35
Product Based Segmentation
D
E
F
Complex
Products
Simple
A
B
C
Low
High
Size
36
Segment Profiles
Slide deliberately left blank.
37
Segmentation with Propensity Modeling

• Add propensity modeling to the static
  segmentation scheme
• Re-categorize customers into Segments
• Identify migrations from one segment to another
• Identify customer growth areas/products
• Promote stewardship for customer growth
• Anticipate new needs
• Develop new products

38
Needs Based Segmentation (Product Migration Paths)
D
E
F
Complex
Products
A
B
C
Simple
Low
High
Size
39
Additional Dimensions
D2
E2
F2
Complex
D1
E1
F1
Products
n
A1
B1
C1
Third Dimension
Simple
Locations
1
Low
High
Size
40
What Next?
41
Whats Next?

• Accommodate increased customer base (due to
  merger) and increased geographic footprint
• More products, more new product development
• Bundles
• Television/Video
• Etc.
• Shifting competitive landscape
• Cable
• New partnerships
• Revisit segmentation complexity (product) and
  size axes
• Evolve segmentation strategies
• Growth Index ? Lifetime value
• Other

42
Growth Potential/Index

• Customers Current Products and Value
• Product A x Revenue for A
• Product B x Revenue for B
• Product F x Revenue for F

X Current Value

• Customers Potential Products and Value
• Product A x Revenue for A
• Product B x Revenue for B
• Product C x Revenue for C
• Product F x Revenue for F
• Product G x Revenue for G

Y Potential Value
Y X Growth Potential/Index
43
Customer Lifetime Value

• CLV - value of a customer over the entire history
  of customer's relationship company
• Acquisition cost
• Churn rate
• Discount rate
• Retention cost
• Time period
• Periodic Revenue
• Profit Margin
• Possibly include Satisfaction Loyalty ?

44
Acknowledgements

• Special thanks to Tim Barnes Sam Massey, ATT -
  2007

45
Contact Information

• David Pope, Ph.D.
• Intelligent Strategies and
• Information Solutions, Inc.
• www.intelligentstrategies.com
• 770.271.9159