Rapid ROI Online Seminar

1
Rapid ROI Online Seminar

• Peter Caron, SPSS Inc.
• Ksenija Krunic, Verizon Wireless
• Joe Somma, HSBC

To speak with an SPSS representative, please
call [phone]
January 29, 2003
2
The Challenges

• Customer Retention
• Customer Profitability
• Key use prediction to focus marketing efforts
  and retain customers

3
Where Prediction Fits
Benefit Identify and exploit new opportunities

• OPTIMIZE
• Customer Retention
• Product Affinities
• Promotions
• Demand Planning
• Quality Improvement
• Employee Utilization

• ENABLE
• Customer Interaction
• Inventory Control
• Supply Chain Management
• Quality Measurement
• Employee Self Service

ERP
Predictive Analytics
CRM
Recommendations
Automate Decision
SCM
Exploit Value
Scoring
Identify Opportunity
Web
Data Mining
Operational Systems
Benefit Reliable effectiveness measurements
OLAP
ETL
Business Intelligence
Data Quality

• UNDERSTAND
• Customer Satisfaction
• Product Revenue
• Cost of Goods Sold
• HR Turnover

Benefit Institutionalized, repeatable
processes
Query/ Report
Data Warehouse
4
Data and Prediction in the Customer Life Cycle
Loyalty
Transaction
Target by profile/score Predict
profit/valueSelect best approach
Retention
Conversion
Reactivation
Acquisition
Time
5
CRISP-DM the Methodology of Prediction
www.crisp-dm.org
6
The Keys to ROI

• Choose a quick-win project
• Use cross-functional teams
• Deploy a repeatable process
• Measure and communicate your ROI

7
Rapid ROI Online Seminar

• Ksenija Krunic, Verizon Wireless

8
About Verizon Wireless

• Largest wireless provider in US
• Customer base 30.3 million
• Covering 90 of US population
• 1220 Company Stores and Kiosks
• 40,000 employees
• 140 Switching Centers

9
Sizing the Issue

• Key Challenge to reduce churn with limited
  resources
• Assumptions (based on Industry averages by Yankee
  Group)
• Average cost of new customer acquisition 320
• Churn 2.0 per month
• So for a company the size of Verizon (30M)
• 600,000 customers disconnect per month
• Associated replacement cost in hundreds of
  millions per year

10
Possible Solutions

• Targeted direct mail to known churn risks
• High churn price plans
• Customers out of contract
• Not targeted enough linear relationship with
  churners
• Develop predictive model with scores and rules
• Analyze rule outputs to identify actionable
  customer sets
• Use findings to target customers with specific,
  relevant, and timely offers

11
Previous Attempts

• Had been mostly IT initiated
• Business did not understand possibilities
• Business conditions were easier fewer
  competitors
• Changing players with corporate merger

12
Success Team

• Data Warehouse Group
• Marketing
• Management team
• Consultants from Data Mining vendor SPSS

13
Building the Team

• IT brought idea to Marketing team and presented
  it as partnership
• Marketing recommended additions to attributes to
  use in building model
• Marketing learned the modeling process as well as
  capabilities and weaknesses of modeling

• IT learned the business processes and direct
  marketing strategies
• Strong vendor relationships for campaign
  management
• Individual and team commitment

14
The Modeling Process

• Build Model
• Included hundreds of basic attributes
• Derived and Ratio fields added to enrich the
  model
• Test Model
• Using unseen set of data
• Select the models with best accuracy
• Score the base
• Collect the data for the active customer base
• Run data through the model
• Outputs score and churn segment profile for
  every customer

15
The Modeling Process cont.

• 4. Output rule sets
• Describe different churn customer segments that
  are vital for marketing campaign design
• 5. Validate Model
• real life accuracy

16
Model Results/Validation
17
IT Team Outputs

• Monthly scoring by billing system
• Populated in Data Warehouse for use by Marketing
• Rules triggered for each customer

• Number of rules generated
• 85 Active customer rules
• 43 Voluntary Churn customer rules
• Campaign Outputs
• Multiple file formats

18
Qualitative Learnings

• Predictors
• Not one or two silver bullets
• Reinforced business knowledge

• Surprises
• Example Dropped calls usually not predictive

19
Marketing Campaigns using Predictive Modeling

• Began with one campaign
• 40-60K pieces per month
• Very personalized unique offer
• Approximately 15 take rate
• Currently four main campaign types
• Multiple segments
• Three billing systems
• 400,000 pieces/month
• Up to 35 take rate of high churn risk customers

20
Multiple Touchpoints

• Outbound Direct Mail and Telemarketing
• Creative Strategy
• One shell, many offers
• Customized one-to-one mailings

• Inbound Customer Care Application
• Customer flagged by offer
• Used By Customer Service, Retail Channels
• To catch customers that
• Outbound reps were unable to contact
• Call to disconnect

21
Benefits

• Cost Reduction
• Customers saved up to 80 more takes
• Direct Mail budget for same churner mailing
  reduced by 60

• Revenue Increase
• Average monthly revenue increase per bill
• Monthly usage increased

• Switched customers from analog to digital
• Contract Renewals increased
• Business Learning
• Some rule sets point out process flaws
• Some rule sets apply across Enterprise without
  separate modeling

22
Key Learnings

• Win back is a different strategy
• Highest score least chance of success
• Garbage in, Garbage out test helps with data
  quality
• Try new creative treatments
• Talk to the customers yourself

23
Closing the Loop - Data Understanding

• Add new data fields as available to enhance
  accuracy
• Infrastructure upgrades
• Up-sell, cross-sell
• Acquisition Models, Response Models

24
Closing the Loop - Business Understanding

• Identify multiple additional customer segments
• Do not need large numbers
• Discrete, actionable groups
• Must be able to execute

• Leverage CRM software
• Multiple offers
• Specific churn triggers
• Improve business processes based on rule sets

25
Summary of Success

• Accurate predictive models
• Quality of data
• Data mining environment
• Effective marketing campaign
• Specific, Relevant, and Timely
• Closing the loop
• Ability to keep models up-to-date
• High performance on large scale data (Teradata)
• Fast, repeatable process enabled by data mining
  tool (Clementine)
• Multi-disciplinary team

26
Rapid ROI Online Seminar

• Joe Somma, HSBC

27
HSBC

• New York Regional bank located in Buffalo, New
  York
• Owned by HSBC (London) -- Third largest bank
  globally
• HSBC USA has over 85 billion in assets and 480
  branches
• Has been actively implementing data mining and
  one-to-one marketing strategies for the last
  three years

28
Business Rationale at HSBC

• 2001-02 Strategic Plan called for enhanced
  cross-sell efforts designed to target deposit
  products
• Wanted to take advantage of money coming out of
  stock market, robust consumer credit market, and
  test various approaches for 2003 CRM initiative

29
Internal Variables

• Length of time as customer
• Length of time since last purchase
• Product last purchased
• Total balances
• Total deposit balances
• Total loan balances
• Number of transactions
• ATM
• Debit
• Number of total accounts

• Number of deposit accounts
• Number of loan accounts
• Total investment balances
• Number of investment products
• User of electronic services
• Geographic identification (e.g., census tract)
• Current profit
• Life time profit
• HSBC segment

30
External Data

• Geo-demographic segmentation data (Microvision
  and Pyscle)
• InfoBase Premier variables

31
How to Decide Which Data to Use

• Define the business problem
• Determine the value each customer provides
• Determine the current costs associated with
  generating that value
• Assess how your current marketing process could
  be improved to solve your business problem
• Identify how data could help you implement the
  improvement in customer value
• Estimate the incremental cost of the data
  technique your considering
• Estimate the incremental return from using
  non-transactional data
• Calculate ROIincremental return/cost

32
First Iteration

• First Variable to differentiate Stock Market CD
  users was Lifetime Value (LTV)
• Initially included as a continuous variable was
  re-coded into five categories and an unknown
• Top two tiers (LTV greater than 300/year)
  provided most significant results

33
Second Iteration

• The second variable to provide differentiation
  was the number of months since the customer
  purchased a product
• This was a continuous variable segmented into
  five categories
• Customers who had purchased products within the
  last year and two years ago displayed the
  greatest association with Stock Market CD
  customers

34
Third Iteration

• Total deposit balances were the third most
  important variable
• Total balances were grouped into ten groups
• Customers with total deposit balances in excess
  of 30,000 were significantly more likely to be
  associated with Stock Market CD purchasers

35
Stock Market CD

• Customers with high LTVs, who purchased a
  product in the last year and had greater than
  30,000 in total deposit balances
• Customers with high LTVs, purchased a product
  within last two years and had CD balances in
  excess of 40,000
• Customers with high LTVs, purchased a product
  within last two years and most recently purchased
  a mutual fund money market

36
Deposit Offers

• Moderate to Low LTV customers with total
  household loan balances (non-mortgage) in excess
  of 20,000
• Low LTV customers who had purchased a product in
  the last three months and had total deposit
  balances under 2,000

37
Loan Offers

• Moderate to Low LTV customers with total
  household loan balances (non-mortgage) in excess
  of 20,000
• Low LTV customers who had purchased a product in
  the last three months and had total deposit
  balances under 2,000
• High transaction customers (3 or more
  transactions per month)

38

• While the output sounds logical, how do we know
  that we have developed an acceptable model?
• Business user review
• Evaluation of gains charts

39
Misclassification Matrix
Actual Category
40
Gains Chart

• The first 7 of file includes 33 of respondents
• Response rate of this segment was greater than
  16
• The top decile accounted for 42 of respondents
• Deciling is an excellent technique for
  determining performance at different file depths

41
Financial Performance -- Deposit

• Used model to reduce mailing costs by 45
• Saved 58,000
• In comparison to prior offer, model attained 95
  of sales, while mailing to a smaller number
• Average opening balances exceeded prior effort by
  22

42
Deployment

• Mailing is only 1 way to deploy
• Critical for sales strategies as well
• Cross-sell information used at branches
• Scripts used to cross-sell

43
Cross-Sell Effect on Retention

• Cross-selling and retention campaigns are useful
  in their own right
• But they do affect each other
• Using predictive techniques, the number of
  customer contacts has decreased, the amount of
  revenue has increased and the volume of mail has
  stayed the same
• Less junk mail can affect loyalty
• A better experience with each interaction can
  affect loyalty

44
Where is HSBC Going From Here?

• Development of a warehouse to speed usage of data
  mining methodology
• Modify warehouse data model to take speed mining
• the paradox of warehouse patterns
• data mine as an analytical repository
• Matching a full family of tools to the mine

45
Rapid ROI Online Seminar

• Questions?
• Peter Caron, SPSS Inc.
• Ksenija Krunic, Verizon Wireless
• Joe Somma, HSBC

46
Thank you!

• Contact information
• Peter Caron
• pcaron_at_spss.com
• www.spss.com
• 800.259.1028
• Ksenija Krunic
• Ksenija.Krunic_at_VerizonWireless.com
• Joe Somma
• Joe.Somma_at_us.hsbc.com