nedma14: answering marketing’s top 3 questions using predictive analytics - william b. disch, ph.d
DESCRIPTION
This informative presentation will teach you how predictive modeling will answer difficult marketing questions, allowing you to focus your resources where you will achieve the highest ROMI. This presentation covers three business cases that answer the three questions: How to effectively improve response rates? How to reduce churn? How to identify customers who are most likely to become best customers? This presentation was given by William B. Disch, SVP of Analytics at Virtual DBS, at NEDMA's Annual Conference on May 14, 2014.TRANSCRIPT
How can we boost response, reduce churn andupsell current customers?
William B. Disch, Ph.D.Senior Vice President of Analytics, Virtual DBS
May 14th, 2014
Answering Marketing’s Top 3 Questions Using Predictive Analytics
Today’s Presenter
William Disch, Ph.D.Senior Vice President of Analytics
• Heads the analytics division at Virtual DBS
• Primary focus is on client collaboration and employing ROI-oriented multivariate predictive modeling and algorithm creation, product sequencing, segmentation, and custom analytics specific to industry verticals
• Key is collaboration in operationally defining context of measurable objectives
• Presenter at major database, analytics and academic conferences nationwide including the DMA, DM Days, the AMA, and others
How can I effectively improve my campaign response rates?
A Marketer’s Top 3 Questions
My customer churn rate is too high. How can I reduce it?
My company sells many products. How can I identify customers who will buy more than one?
Issue/Question Modeling Solution
Lower than optimal conversion rate for new leads.
How do I increase new acquisitions while at the same time keeping costs down?
Customer Acquisition Model
Churn rate higher than acceptable.
Can I identify current customers at the highest risk to churn before they leave?
Churn (Attrition) Model
Missing upsell opportunities.
How can I identify customers most likely to buy their next product?
Upsell Model
Three Cases
Case 1
Customer Acquisition Model:
Specialty Foods Retailer
Customer Acquisition Model Specialty Foods Retailer Business Case
Specialty Foods Retailer Current State600,000 prospect mail pieces sent annually
Current Results
2% gross response rate (6,000 responders)
25% conversion rate (1,500 customers purchased)
Specialty Foods Retailer Desired StateNo change in mail volume
Desired Results
2.6% gross response rate (9,000 responders, 30% improvement)
25% conversion rate (2,250 customers purchased)
Customer Acquisition Model –The Process Simplified
Campaign Responders
Campaign Non-Responders
Virtual DBS Appended
Demographics
Purchasers(subset of Campaign
Responders)
Predictive Analytics
Processing
Predictive Algorithm Reveals Top Response/Acquisition Drivers and
their Predictive Weight
Virtual DBS Compiled B2B and B2C Data Includes Hundreds of Demographics and Related Elements
Demographic
Income Wealth Age Ethnicity Occupation Household Type Marital Status Length of Residence Home Ownership Home Value Mortgage Info Home Size (Sq Ft) Lender Codes Age of Home Dwelling Type Small Office/Home Office Presence of Children Ages of Children
Interests
Fitness Outdoors Athletic Cultural Charitable Events Community Involvement Gardening Financial Travel Donor Do It Yourselves Etc.
Buying Behavior
Product Types Travel Upscale Retail Finance Etc.
Life Stage Clusters
Mutually ExclusiveClusters Life Stages:
- Springs: 18 - 24- Summers: 25 - 44- Autumns: 45 - 64- Winters: 65+
Income Range:- Low: 40k- Mid: 40k – 75k- High: 75k+
Family Type:- Single- Couples- Families
Community Type:- Rural- Suburban- Urban
B2B Firmagraphics
SIC Division SIC/NAICS Codes No. Employees Annual Sales Ownership Type Location Type Years In BusinessBusiness Verticals Technology Use SOHO Etc.
0 0.05 0.1 0.15 0.2 0.25 0.3
Gender
Household Type
Dwelling Type
Marital Status
Household Income
Gifts
Sports/Leisures
Mail Order: Food Products
Health
Buyer Orders: Home Care
Garden
Number of Children
Assessed Median Home Value
Political Donor
Hobby: Knitting/Needlework
Likes to Read
Hobby: Cooking
Reading: Cooking/Culinary
Health/Institutional Donor
Predictive Attributes Driving Response/Acquisition
A specialty foods retailer wants to increase response rates of new customers buying holiday food products.
We operationally defined the event group of those who had purchased during the past season at a dollar value of X or higher.
The drivers in the algorithm show that the best prospects tend to be females in single family households with children, with moderate to high income, and who have a propensity to use discretionary income for a variety of personal and social needs and behaviors.
GenderFemale - 63%
Household TypeAdult Male & Female Present w Kids - 49%
Dwelling TypeSingle Family – 74%
Marital StatusMarried - 58%
Household Income$150k + - 35%$125-$150k - 10%$100-$125k - 15%
Assessed Median Home Value$750k + - 5%$700-$750k - 1%$500-$550k - 2%
Customer Acquisition Algorithm Example
An algorithm is a mathematical equation that incorporates predictive drivers and their weights.
Constant (unique to each algorithm)
+ Gender (x .43)
+ Household Type (x .38)
+ Dwelling Type (x .32)
+ Marital Status (x .31)
+ Household Income (x .27)
+ Gift Behavior (x .25)
+ Sport/Leisure Interest (x .23)
+ Mail Order Food Product (x .23)
+ Health Interest (x .22)
+ Home Care Buyer (x .21)
+ Garden Interest (x .20)
+ Number of Children (x .18)
+ … remaining predictors… (x .XY)
= Propensity to Purchase
How Do We Know the Model Works?
VALIDATION PROCESS:How we assess the power and efficacy of a model:
Acquisition model strength is tested on a validation dataset:
1. Randomly see event group/target records into the prospect universe set of records
2. Run the algorithm
3. Event group/targets should score near the top of the scored file
4. Conduct multiple iterations
Model Scoring Validation Gains Table
Probability Random Validation Acquisition Score Rank
5% 0.50 7.34Ranks
1, 2 and 3, 410% 0.50 1.50
15% 0.50 1.37
20% 0.50 1.14
25% 0.50 1.15Ranks 5,6
30% 0.50 0.93
The validation shows that the top 5% of scored prospects are 7.3 times more likely to become a customer than a random prospect
Before Acquisition Model Scoring We Are Here
FirstName LastName Address1 City State Zip Phone Email
Acquisition Score
Aimee Falcigno 9 Rocamora Road Rocky Hill CT 06067 8605718495 [email protected] ?
Ann Reynolds 94 Brookmoor Rd West Hartford CT 06109 8603130922 [email protected] ?
Amy Lubas 823 Edgemere Lane Sarasota FL 34242 9419939899 [email protected] ?
Andy Skarzynski 34 Ballard Drive West Hartford CT 06119 8602333240 [email protected] ?
Ann Schmidt 21 Onlook Rd Wethersfield CT 06109 8605294388 [email protected] ?
Anthony Paquette 88 Oak Leaf Drive Colchester CT 06415 8605378667 [email protected] ?
William Derech 132 Randy Lane Wethersfield CT 06109 8605298060 [email protected] ?
Suzanne Costanzo 9 Deerfield Trace Burlington CT 06013 8603067433 [email protected] ?
Gina Cunningham 28 Cockle Hill Road Salem CT 06420 8608893943 [email protected] ?
Caperton Beirne 7100 Quail Hill Road Charlotte NC 28210 7047808753 [email protected] ?
Carol Heminger 515 Edgebrook Lane West Palm Beach FL 33411 5617750098 [email protected] ?
Christine Peterson 60 Iroquois Road West Hartford CT 06117 8602332239 [email protected] ?
Craig Meeker 44 Garden Gate Farmington CT 06032 8602840123 [email protected] ?
Cynthia Poglitsch 98 Cobblestone Way Windsor CT 06095 8606885207 [email protected] ?
Carrie Stephens 3135 Crittenton Place Baltimore MD 21211 5618012538 [email protected] ?
David Iaia 37 Wilson Avenue Belmont MA 02478 6174894839 [email protected] ?
David Wurts 6404 S Ross Road Morrison CO 80465 3036974486 [email protected] ?
David Vining 65 Falcon Lane Glastonbury CT 06033 8604636450 [email protected] ?
We have no way of predicting a prospect’s response/purchase behavior.
After Acquisition Model Scoring We Are Here
FirstName LastName Address1 City State Zip Phone Email
Acquisition Score
Aimee Falcigno 9 Rocamora Road Rocky Hill CT 06067 8605718495 [email protected] 1
Ann Reynolds 94 Brookmoor Rd West Hartford CT 06109 8603130922 [email protected] 6
Amy Lubas 823 Edgemere Lane Sarasota FL 34242 9419939899 [email protected] 9
Andy Skarzynski 34 Ballard Drive West Hartford CT 06119 8602333240 [email protected] 1
Ann Schmidt 21 Onlook Rd Wethersfield CT 06109 8605294388 [email protected] 1
Anthony Paquette 88 Oak Leaf Drive Colchester CT 06415 8605378667 [email protected] 2
William Derech 132 Randy Lane Wethersfield CT 06109 8605298060 [email protected] 10
Suzanne Costanzo 9 Deerfield Trace Burlington CT 06013 8603067433 [email protected] 1
Gina Cunningham 28 Cockle Hill Road Salem CT 06420 8608893943 [email protected] 4
Caperton Beirne 7100 Quail Hill Road Charlotte NC 28210 7047808753 [email protected] 7
Carol Heminger 515 Edgebrook Lane West Palm Beach FL 33411 5617750098 [email protected] 8
Christine Peterson 60 Iroquois Road West Hartford CT 06117 8602332239 [email protected] 2
Craig Meeker 44 Garden Gate Farmington CT 06032 8602840123 [email protected] 9
Cynthia Poglitsch 98 Cobblestone Way Windsor CT 06095 8606885207 [email protected] 10
Carrie Stephens 3135 Crittenton Place Baltimore MD 21211 5618012538 [email protected] 2
David Iaia 37 Wilson Avenue Belmont MA 02478 6174894839 [email protected] 1
David Wurts 6404 S Ross Road Morrison CO 80465 3036974486 [email protected] 10
David Vining 65 Falcon Lane Glastonbury CT 06033 8604636450 [email protected] 4
We know exactly which prospects are likely to respond and buy.
Case 2
Churn Model:
Telecommunications Business Case
Churn Model Telecommunications Business Case
Telecom Company Current State (2013 results)Subscriber count on January 1 : 1,000,000Net Subscriber count on December 31 : 1,010,000
Current ResultsCustomers churned : 200,000Net subscriber gain : 10,000Churn rate : 20%
Telecom Company Desired State (2014 plan)Subscriber count on January 1 : 1,010,000Net Subscriber count on December 31 : 1,119,500
Desired ResultsCustomers churned : 100,500Net subscriber gain : 109,500Churn rate : 10%
How Difficult Is It to Sell Something?The Economics of Marketing
It is 3x to 7x HARDER to sell to new customer than an existing one.
Existing Customer New Customer
Existing Product 1X 3X
New Product 2X 7X
Retention makes a hard job many times easier
Churn makes an already difficult job many times harder
If we can acquire new customers at the lowest possible cost, extra resources can be applied to retention efforts
Churn Model The Process Simplified
Current Customers
Lapsed CustomersVirtual DBS Appended
Demographics
Payment history, price/promo,
products purchased, CS
calls, etc.
Predictive Analytics
Processing
Predictive Algorithm Reveals Top Churn Drivers and their
Predictive Weights
Predictive Attributes Driving Churn
The top predictive churn drivers show us why customers left:
7. Aggressive Competitive Offer
6. Promotion Period Expiring
5. SOHO*
4. Technical Issues
3. GeoVector*
2. Price
1. Service
0% 5% 10% 15% 20% 25% 30%
Care Call: Service - Last 30 Days
Care Call: Price - Last 30 Days
Care Call: Tech Probs - Last 30 Days
Duration to Promo Roll-Off
Care Call: Service - Last 90 Days
Competitor Aggressive Promotion
Product Grade (single to bundles)
Care Call: Service - Last 60 days
Active or Inactive Promo Flag
Last Package (single, double, triple)
Promo Duration
Significant Churn Predictors (Customer Variables)
* Virtual DBS appends
Telecom Churn Algorithm Example
A churn algorithm is a mathematical equation that incorporates predictive drivers and their weights.
Constant + Aggressive Competitive Promotion (x .16) + Time to Promo Expiration (x .21) + SOHO (small office, home office), (x .23) + Number of Tech Support Calls (x .27) + GeoVector (Age, Income, Geo, Family Type), (x .32) + Price (x .36) + Number of Service Issues (x .43)
= Propensity to Churn
Before Churn Model Scoring We Are Here
We have no way of predicting future churn behavior.
FirstName LastName Address1 City State Zip Phone Email Churn Score
Aimee Falcigno 9 Rocamora Road Rocky Hill CT 06067 8605718495 [email protected] ?
Ann Reynolds 94 Brookmoor Rd West Hartford CT 06109 8603130922 [email protected] ?
Amy Lubas 823 Edgemere Lane Sarasota FL 34242 9419939899 [email protected] ?
Andy Skarzynski 34 Ballard Drive West Hartford CT 06119 8602333240 [email protected] ?
Ann Schmidt 21 Onlook Rd Wethersfield CT 06109 8605294388 [email protected] ?
Anthony Paquette 88 Oak Leaf Drive Colchester CT 06415 8605378667 [email protected] ?
William Derech 132 Randy Lane Wethersfield CT 06109 8605298060 [email protected] ?
Suzanne Costanzo 9 Deerfield Trace Burlington CT 06013 8603067433 [email protected] ?
Gina Cunningham 28 Cockle Hill Road Salem CT 06420 8608893943 [email protected] ?
Caperton Beirne 7100 Quail Hill Road Charlotte NC 28210 7047808753 [email protected] ?
Carol Heminger 515 Edgebrook Lane West Palm BeachFL 33411 5617750098 [email protected] ?
Christine Peterson 60 Iroquois Road West Hartford CT 06117 8602332239 [email protected] ?
Craig Meeker 44 Garden Gate Farmington CT 06032 8602840123 [email protected] ?
Cynthia Poglitsch 98 Cobblestone Way Windsor CT 06095 8606885207 [email protected] ?
Carrie Stephens 3135 Crittenton Place Baltimore MD 21211 5618012538 [email protected] ?
David Iaia 37 Wilson Avenue Belmont MA 02478 6174894839 [email protected] ?
David Wurts 6404 S Ross Road Morrison CO 80465 3036974486 [email protected] ?
David Vining 65 Falcon Lane Glastonbury CT 06033 8604636450 [email protected] ?
After Churn Model Scoring We Are Here
We know exactly which customers are most likely to churn.
FirstName LastName Address1 City State Zip Phone Email Churn Score
Aimee Falcigno 9 Rocamora Road Rocky Hill CT 06067 8605718495 [email protected] 1
Ann Reynolds 94 Brookmoor Rd West Hartford CT 06109 8603130922 [email protected] 6
Amy Lubas 823 Edgemere Lane Sarasota FL 34242 9419939899 [email protected] 9
Andy Skarzynski 34 Ballard Drive West Hartford CT 06119 8602333240 [email protected] 3
Ann Schmidt 21 Onlook Rd Wethersfield CT 06109 8605294388 [email protected] 1
Anthony Paquette 88 Oak Leaf Drive Colchester CT 06415 8605378667 [email protected] 2
William Derech 132 Randy Lane Wethersfield CT 06109 8605298060 [email protected] 10
Suzanne Costanzo 9 Deerfield Trace Burlington CT 06013 8603067433 [email protected] 1
Gina Cunningham 28 Cockle Hill Road Salem CT 06420 8608893943 [email protected] 4
Caperton Beirne 7100 Quail Hill Road Charlotte NC 28210 7047808753 [email protected] 7
Carol Heminger 515 Edgebrook Lane West Palm BeachFL 33411 5617750098 [email protected] 8
Christine Peterson 60 Iroquois Road West Hartford CT 06117 8602332239 [email protected] 2
Craig Meeker 44 Garden Gate Farmington CT 06032 8602840123 [email protected] 9
Cynthia Poglitsch 98 Cobblestone Way Windsor CT 06095 8606885207 [email protected] 10
Carrie Stephens 3135 Crittenton Place Baltimore MD 21211 5618012538 [email protected] 8
David Iaia 37 Wilson Avenue Belmont MA 02478 6174894839 [email protected] 1
David Wurts 6404 S Ross Road Morrison CO 80465 3036974486 [email protected] 10
David Vining 65 Falcon Lane Glastonbury CT 06033 8604636450 [email protected] 4
Case 3
Upsell Model:
Utility Company
Upsell Model Utility Company Business Case
Utility Company Current State
Customers can buy an add-on insurance product to protect their furnace
Quarterly direct mail campaign with insurance offer sent to all 100,000 customers
Current Results
Gross response rate 3% (3,000 responses)
Conversion rate 10% (300 sales)
Utility Company Desired State
Quarterly direct mail campaign sent to 20,000 current customers most likely to buy insurance product
Desired Results
Mail 80,000 fewer records
Achieve 18% gross response rate (3,600 responses)
Conversion rate 10% (360 sales)
Upsell Model –The Process Simplified
Customers without Insurance
Product
Customers with Insurance Product
Virtual DBS Appended
Demographics
Predictive Analytics
Processing
Predictive Algorithm Reveals Top Upsell Purchase Drivers and their
Predictive Weights
Predictive Attributes Driving Upsell Purchase
GeoVector3323: 45-64, $75k+, Suburban, Families-~18%3313: 45-64, $75k+, Urban, Families-~8%2323: 25-44, $75k+, Suburban, Families-~8%
Household Income$50,000-$74,999-~19%$150,000+-~19%$75,000-$99,999-~18%
Dwelling TypeSingle Family-~100%
Homeowner StatusOwner- ~97%Renter- ~1%
Pro-Environmental StatusYes- ~3%
Mail ResponderMultiple- ~78%Single- ~1%
Length of Residence15+ Years- ~39%11-14 Years- ~17%8-10 Years- ~14%
Socio-Demographic Clusters3-Corporate Clout- ~5.68%9-Platinum Oldies- ~5.58%5 Sitting Pretty- ~5.45%
InterestsGarden- ~14%Investments- ~38%Travel- ~56
DonorReligious Donor- ~20%Health Institutional Donor- ~22%
Before Upsell Model Scoring We Are Here
FirstName LastName Address1 City State Zip Phone Email Upsell Score
Aimee Falcigno 9 Rocamora Road Rocky Hill CT 06067 8605718495 [email protected] ?
Ann Reynolds 94 Brookmoor Rd West Hartford CT 06109 8603130922 [email protected] ?
Amy Lubas 823 Edgemere Lane Sarasota FL 34242 9419939899 [email protected] ?
Andy Skarzynski 34 Ballard Drive West Hartford CT 06119 8602333240 [email protected] ?
Ann Schmidt 21 Onlook Rd Wethersfield CT 06109 8605294388 [email protected] ?
Anthony Paquette 88 Oak Leaf Drive Colchester CT 06415 8605378667 [email protected] ?
William Derech 132 Randy Lane Wethersfield CT 06109 8605298060 [email protected] ?
Suzanne Costanzo 9 Deerfield Trace Burlington CT 06013 8603067433 [email protected] ?
Gina Cunningham 28 Cockle Hill Road Salem CT 06420 8608893943 [email protected] ?
Caperton Beirne 7100 Quail Hill Road Charlotte NC 28210 7047808753 [email protected] ?
Carol Heminger 515 Edgebrook Lane West Palm Beach FL 33411 5617750098 [email protected] ?
Christine Peterson 60 Iroquois Road West Hartford CT 06117 8602332239 [email protected] ?
Craig Meeker 44 Garden Gate Farmington CT 06032 8602840123 [email protected] ?
Cynthia Poglitsch 98 Cobblestone Way Windsor CT 06095 8606885207 [email protected] ?
Carrie Stephens 3135 Crittenton Place Baltimore MD 21211 5618012538 [email protected] ?
David Iaia 37 Wilson Avenue Belmont MA 02478 6174894839 [email protected] ?
David Wurts 6404 S Ross Road Morrison CO 80465 3036974486 [email protected] ?
David Vining 65 Falcon Lane Glastonbury CT 06033 8604636450 [email protected] ?
We have no way of knowing which customers will make a next purchase.
After Upsell Model Scoring We Are Here
FirstName LastName Address1 City State Zip Phone Email Upsell Score
Aimee Falcigno 9 Rocamora Road Rocky Hill CT 06067 8605718495 [email protected] 7
Ann Reynolds 94 Brookmoor Rd West Hartford CT 06109 8603130922 [email protected] 1
Amy Lubas 823 Edgemere Lane Sarasota FL 34242 9419939899 [email protected] 9
Andy Skarzynski 34 Ballard Drive West Hartford CT 06119 8602333240 [email protected] 6
Ann Schmidt 21 Onlook Rd Wethersfield CT 06109 8605294388 [email protected] 4
Anthony Paquette 88 Oak Leaf Drive Colchester CT 06415 8605378667 [email protected] 2
William Derech 132 Randy Lane Wethersfield CT 06109 8605298060 [email protected] 9
Suzanne Costanzo 9 Deerfield Trace Burlington CT 06013 8603067433 [email protected] 1
Gina Cunningham 28 Cockle Hill Road Salem CT 06420 8608893943 [email protected] 2
Caperton Beirne 7100 Quail Hill Road Charlotte NC 28210 7047808753 [email protected] 8
Carol Heminger 515 Edgebrook Lane West Palm Beach FL 33411 5617750098 [email protected] 1
Christine Peterson 60 Iroquois Road West Hartford CT 06117 8602332239 [email protected] 2
Craig Meeker 44 Garden Gate Farmington CT 06032 8602840123 [email protected] 10
Cynthia Poglitsch 98 Cobblestone Way Windsor CT 06095 8606885207 [email protected] 3
Carrie Stephens 3135 Crittenton Place Baltimore MD 21211 5618012538 [email protected] 2
David Iaia 37 Wilson Avenue Belmont MA 02478 6174894839 [email protected] 1
David Wurts 6404 S Ross Road Morrison CO 80465 3036974486 [email protected] 3
David Vining 65 Falcon Lane Glastonbury CT 06033 8604636450 [email protected] 1
We know exactly which customers are most likely to make a next purchase.
Summary
• Predictive modeling effectively answers difficult marketing questions
• Predictive modeling allows you to maximize your ROI by concentrating your resources on those customers or prospects most likely to buy or churn
• Scored data resulting from a predictive model is immediately actionable
• Predictive algorithms are portable and can be used to score a variety of internal and external lists
• Client collaboration and the operationally defined metrics specific to the current state business state are key – model performance is highly correlated with the quality of the metrics used to build the model
About Virtual DBS
What we do
Virtual DBS offers technology, data, and analytics allowing corporate decision makers to gain strategic business insights they can use to make profitable business decisions.
Best in class tools for CDI, predictive analytics, and campaign management to organize, extract, and monetize customer and prospect databases and generate positive ROI.
Highly effective and affordable products and services for B2C and B2B marketers.
Founded and managed by industry veterans with a focus on mathematical precision, customer service, and client collaboration.
For questions or further information, please contact
John Dodd, EVP
Direct 401.667.7595
www.virtualdbs.com
Q and A
Appendix
After successful deployment of hundreds of modeling initiatives for a
multitude of clients in widely varying marketing scenarios, we have often
seen response performance improvements of 20% to 40% over
established baseline.
• For example, where a particular package-list-offer combination has historically
generated a 2% response rate, we often see our clients enjoying response
rates ranging from 2.4% to 2.8% (i.e. 20% to 40% above established baseline)
by utilizing Virtual DBS predictive modeling in their customer development
targeted marketing campaigns.
• We have seen highly profitable modeling initiatives in which lesser gains were
achieved (often as low as a couple of percentage points over control) – but
have also seen campaigns come in with much higher response lift (e.g. 2x over
baseline).
AppendixA Note on Response Performance
AppendixPredictive Modeling Overview
• Modeling uses past behaviors (respond, buy, churn) to optimize those behaviors going forward
• We combine appended demographics with customer-specific fields (transaction values, dates, product details, etc.)
• Two Primary Outcomes:
1. Behavioral Profile
2. Scoring Algorithm
AppendixModeling Answers Key Strategic Questions
What do my best customers look like?
How do I find more prospects who look like them?
What is my market penetration?
Where are my new clients going to come from?
How do I stay relevant to my various customer groups?
Which of my customers are most likely to leave?
Which customers are going to spend the most?
What is the lifetime value of a customer?
What is the cost to acquire a new customer?
How do I help low-performing customers to become high-performing customers?
AppendixTypes of Models
Churn
Acquisition
Customer Optimization (cross and upsell)
Cluster/Segmentation
Best Payer
Best Customer
Price Elasticity
Product Sequencing
Others
0.00
2.00
4.00
6.00
8.00
10.00
Algorithm Performance
Random Validation
There are two primary steps for validating a predictive algorithm once the algorithm has been created.
First, the event group sample is randomly seeded into the universe sample, using multiple iterations of random samples, then the file is scored using the algorithm.
If the algorithm is successful, the event group sample should score in the “Best” deciles, and up and to the left in the above bar chart. The results mean that the randomly seeded event group sample is being successfully predicted by the algorithm.
In this case, model performance indicates that deciles 1-2 have the greatest lift. Detection of seeds suggests suppressing the top ~20% of the top scoring records yields a probability of capturing ~68% of current customers.
Model Scoring Validation Gains Table
Probability Tier
Random Validation Prospect Selection
5% 0.50 8.04
Deciles 1 thru 2
10% 0.50 2.52
15% 0.50 1.69
20% 0.50 1.29
25% 0.50 1.08Decile 3
30% 0.50 0.89
35% 0.50 0.76
Deciles 4 thru 10
40% 0.50 0.59
45% 0.50 0.53
50% 0.50 0.50
55% 0.50 0.50
60% 0.50 0.50
65% 0.50 0.50
70% 0.50 0.50
75% 0.50 0.50
80% 0.50 0.50
85% 0.50 0.50
90% 0.50 0.50
95% 0.50 0.50
100% 0.50 0.50
AppendixHow Do We Know the Model Works?
0.00
0.20
0.40
0.60
0.80
1.00
Lift and Gains Performance
Random Validation
Model Scoring Validation Gains Table
Probability Tier
Random Validation Prospect Selection
5% 0.50 8.04
Deciles 1 thru 2
10% 0.50 2.52
15% 0.50 1.69
20% 0.50 1.29
25% 0.50 1.08Decile 3
30% 0.50 0.89
35% 0.50 0.76
Deciles 4 thru 10
40% 0.50 0.59
45% 0.50 0.53
50% 0.50 0.50
55% 0.50 0.50
60% 0.50 0.50
65% 0.50 0.50
70% 0.50 0.50
75% 0.50 0.50
80% 0.50 0.50
85% 0.50 0.50
90% 0.50 0.50
95% 0.50 0.50
100% 0.50 0.50
AppendixHow Do We Know the Model Works? (cont.)
Second, the probability of increased responding for the modeled event group is plotted again a random sample.
Using the gains table to the right, the results also show that scored prospects in the top 5% of the prospect file are 8x more likely to look like a current Best Responder, and those from the second tier are 2.5x more likely to look like a current Best Responder.
Overall, prospects in the top 10% are approximately 5.5x more likely to look likely to look like current Best Prospects, compared to only a 50/50 probability by using change alone.
Again, in this case, model performance indicates that deciles 1-2 have the greatest lift. Detection of seeds suggests suppressing the top ~20% of the top scoring records yields a probability of capturing ~68% of current customers.
For questions or further information, please contact
John Dodd, EVP
Direct 401.667.7595
www.virtualdbs.com