The document discusses a project between Wharton Interactive Media Initiative (WIMI) and ESPN to build a predictive model to forecast media consumption across ESPN's digital properties during the World Cup. WIMI developed a multi-channel predictive model using user data from ESPN.com, video, and mobile to accurately forecast daily reach and cumulative usage. The model provided insights into how consumption varied by channel and day of week, and how users' behaviors on different channels were related. It also showed potential for estimating the relative media drawing power of individual soccer teams.
Academically-Practical and Practically-Academic Learnings in Interactive Media
1. Academically‐Practical and Practically‐Academic
Learnings in Interactive Media
Wharton Interactive Media Initiative
Professor Eric T. Bradlow
K.P. Chao Professor
Professor of Marketing, Statistics, and Education
Vice‐Dean and Director, Wharton Doctoral Programs
Co‐Director, Wharton Interactive Media Initiative
www.whartoninteractive.com
2. T H E R E I S N O G R E AT D I V I D E !
Practice
Academics
Academically‐Practical Practically‐Academic
3. H O W D O I K N O W W H AT A C A D E M I C S K N O W A N D H O W D O I
K N O W W H AT P R A C T I T I O N E R S C A R E A B O U T ?
• WIMI Corporate Partners
o Travel and Listen!
• Matchmaking Webinars
o Take Corporate Partner Business Problems and Present Them to
the Academic Community
• My Own Academic Research
• Academic Research Conferences and WIMI‐Funded Research
4. W IM I’ S “ L E A R N I N G N ET W O R K”
Global network of research partners
Wharton Lab
for Publishing WIMI
Research,
Innovation Student Placement
Interns,
Partners
W H A R T O N I N T E R A C T I V E M E D I A I N I T I AT I V E
6. What is the #1 Problem Today
for Internet Ad Publishers?
NOBODY KNOWS HOW MUCH TO PAY THEM!
ADVERTISING ATTRIBUTION
* NOT LAST CLICK
* NOT EQUALLY SPREAD
WHARTON INTERACTIVE MEDIA
INITIATIVE
7. O R G A N I C TA C K L E S A D AT T R I B U T I O N !
ORGANIC DEVELOPED AND MANAGED A COMPLETE DIGITAL MARKETING
STRATEGY FOR THE CLIENT, A NEW CAR MANUFACTURER
Display advertising on media sites
Sponsored search
Shopping sites
Advertiser sites
8. DIGITAL ADVERTISING “PATHS” FOR NEW CAR SHOPPERS
(HYPOTHETICAL)
DATA
View Ad View Ad
Edmunds.com CNN.com
Day 1
View Ad View Ad View Ad Click‐through @
CNN.com KBB.com CNN.com CNN.com
Day 6
Page view at advertiser “Conversion” at
Click‐through @ Google
site advertiser site
Day 20
User 1
User 2
User 3
9. D ATA
AVAILABLE FIELDS
Display advertising impressions
For each activities at the advertisers site
(including conversions)
• User • User
• Date & time • Date and time
• Advertiser organization (i.e., brand) • Type of activity
• Media buy name • “Conversion” or “Success” activities
• Site where ad was displayed (28 sites) • Search inventory
• User’s country, state & area code (based on IP) • Find a dealer
• Build & price
• Get a quote
• Other activities
Click‐throughs • User’s state & area code (based on IP)
• User • Whether the conversion occurred in the
• Date & time same session as a click‐through
• Advertiser organization (i.e., brand
• Media buy name
• Site where ad was displayed
• Ad id number
(no info on ad content)
• User’s country & state code KEY IS HAVING ALL THREE
(based on IP) LINKED TOGETHER
11. W H AT T O S H O W W H E N S O M E O N E S E A R C H E S ?
WHARTON INTERACTIVE MEDIA INITIATIVE
12. E X P E D I A TA K E S O N O P T I M A L S E A R C H R E S U LT S
Retail
Opaque
Package
Corporate
Media
13. DATA ON 10,000K+ HOTEL SEARCHES CONDUCTED OCT 1-15, 2009
Free Text
Associated with
Search
Region/Distinct Keyword
Assigned to that Text
Travel Dates
Number of Number of
Rooms Travelers
Time/Date of Search
14. FOR EACH SEARCH WE OBSERVE WHICH HOTELS WERE DISPLAYED
Number of
Hotels that Meet
Search Criteria
Hotels Displayed
Price Displayed
for Each Hotel
Was the Price a
Promo?
15. WE ALSO OBSERVE WHICH HOTELS WERE VIEWED AND PURCHASED
Which Hotels
Got Click‐
Throughs?
(if any)
Which Hotels got
“Book It” Click‐
Throughs?
Which Hotels
Were Purchased?
16. ERIC “THE WIZARD”: PREDICTING AND
M O N E T I Z I N G F U T U R E B E H AV I O R
W H A R T O N I N T E R A C T I V E M E D I A I N I T I AT I V E
17. D E S C R I P T I O N O F D ATA
• Contains 23,000 users of hulu.com who registered during February
2009.
o Take 10% random sample
• Tracking daily incidence of visiting to view videos for each of 120 days
starting March 1, 2009.
• Summary Statistics of 90‐day in‐sample period:
o Reach: 46% of people visit at least once
o Frequency: 4.3 visits on average, among those who visit
o Streakiness: 446 total streaks of visits lasting 3 or more consecutive days (across all
people)
• Last 30 days are the holdout (out‐of‐sample) period used for model
validation.
17
18. EXAMPLE 1: MAKING MONEY FOR HULU
ALIVE, THEN
DEAD
ALIVE, THEN
COLD
ALIVE OR
“DEAD”
ALIVE OR
COLD
WINNER ‐> DON’T PAY TO BRING BACK FROM THE “DEAD”
18
19. MAKING MONEY FOR MECOX LANE
• A retailer (with catalogs, stores and a website) would like a tool to identify which
consumers active and which ones have ended their relationship with the firm
• The retailer provided transaction history across three channels (web, store & catalog) for
a random sample of 30,000 customers
• Using this data, researchers at WIMI are developing a model that can be used to:
o Identify ‘inactive’ customers
o Forecast future sales
o Plan capacity
o Understand multi‐channel behavior
• Unlike many other forecasting
approaches does not require any
information about the consumer other
than her purchase history
o Easily applied in many settings
20. REMARKABLE PREDICTION ACCURACY
Cummulative Orders • The model is based on the simple
Actual cummulative Forecast Cummulative idea that people buy at a steady rate
8000 until they become inactive
7000
o But different people have
6000 different rates
5000 • By using the data to estimate the
4000
rates at which people buy and
become inactive, we build a model
3000
that can forecast orders into the
2000 future
1000
• These models have proven accurate
0 across many industries and contexts
* All results preliminary
21. INSIGHT INTO DIFFERENCES BETWEEN CHANNELS
Beta Density for Dropout Rate
• Even though we never observe when
a customer becomes inactive, the
Overall Catalog (Method=1)
1000.0
.com (method=8) Store (Method=M)
model gives us an estimate of the
Proportion of Customers (Probability Density)
900.0 drop‐out rates
800.0 .com customers are
700.0
less likely to drop out o Most .com customers have a
than others
600.0 very low drop‐out rate
500.0
Catalog customers are
o Most catalog customers have a
400.0 more likely to drop much higher drop‐out rate
out than others
300.0 o Store shoppers vary widely in in
200.0
their propensity to drop out
100.0
0.0
0.000 0.002 0.004 0.006 0.008 0.010 0.012 0.014
Daily Dropout Rate (!)
* All results preliminary
22. FO R ECAST ING M ULTI ‐ C H A NNEL M E D I A CO N S U M PT ION D U R I N G T H E
WO R LD C U P
Elea McDonnell Feit
Pengyuan Wang
Eric T. Bradlow
Peter S. Fader
Wharton Interactive Media Initiative
25. ESPN OBJECTIVE FOR WIMI
Build a state‐of‐the‐art predictive model to
understand and project "multichannel"
consumption habits across digital properties
(Internet, Mobile & Streaming Video)
26. M U LT I - C H A N N E L T O U R N A M E N T F O R E C A S T I N G
• The Wharton Interactive Media Initiative developed a state‐of‐the‐art predictive
modeling method to understand and project multi‐channel consumption habits across
media platforms (web, video and mobile).
• We tested this model using usage data for individual fans across three channels and were
able to make accurate forecasts, measure the relationships between channels, and
estimate the media attractiveness of individual teams.
Soccer Reach for ESPN Digital Properties During World Cup
.com U U U U
(number of registered fans visiting daily)
Video S S S S
Mobile
Soccer Reach
6/4
6/5
6/6
6/7
6/8
6/9
6/10
6/11
6/12
6/13
6/14
6/15
6/16
6/17
6/18
6/19
6/20
6/21
6/22
6/23
6/24
6/25
6/26
6/27
6/28
6/29
6/30
7/1
7/2
7/3
7/4
7/5
7/6
7/7
7/8
7/9
7/10
7/11
Day
27. FINDINGS
Daily Soccer Reach Forecast for ESPN.com
• Forecasting during World Cup
(Number of registered fans visiting daily
Predicted Actual
o By summing up predictions for
100
individual fans, we make accurate
per 1,000 fans)
forecasts of overall reach for each 80
Reach
channel. 60
• Multi‐channel behavior 40
20
o Fans are less likely to use ESPN.com
0
on weekends, but Mobile usage is
6/4
6/6
6/8
6/10
6/12
6/14
6/16
6/18
6/20
6/22
6/24
6/26
6/28
6/30
7/2
7/4
7/6
7/8
7/10
unaffected by weekends.
o Among those who use mobile and Cumulative Soccer Frequency for ESPN.com
.com, the more a fan uses Mobile, the during World Cup
2000
less he uses ESPN.com.
(Total visits for 1000 registered fans)
1800
• Team strengths
Cumulative Frequency
1600
1400
o The method we have developed can 1200
be used to estimate the media 1000
800
attractiveness of individual teams. 600
400
200
0
6/4 6/11 6/18 6/25 7/2 7/9
28. A C A D E M I C A L LY P R A C T I C A L I N T E R A C T I V E M E D I A 2009-2010
Cross‐Platform Data: Dec 2010
Social Networking: Jan 2009
Impact and Emergence of UGC:
Dec 2009
Mobile Marketing: 2011
29. Mine Your Own Business
Market Structure Surveillance through Text Mining
Feldman, Goldenberg, Netzer
Is “classic” Marketing
Research dead? Use analytics to explore the
relationship between brands
Perceptual Map of US Car Makes
Text mine consumer posts
compact sport old
Audi A6 67 345 56
Honda
1384 539 245
Civic
Toyota
451 128 211
Corolla
Customers are telling us things for “free”
30. Does Chatter Really Matter?
Dynamics of User-Generated Content and Stock Performance
Tirunillai and Tellis
Short‐term Long‐term
effect on effect on
stock stock
returns returns
Chatter 3.8 4.8
Consumer ‐2.1 ‐3.6
Opinion
Negative ‐2.9 ‐3.9
Chatter
Negative ‐3.7 ‐4.7
Expressions
What your customers are saying matters (if you own stock)
“You can take UGC to the Bank”
31. Crowdsourcing New Product Ideas
Bayus
“The goal is for you, the customer,
is to tell Dell what new products or
services you’d like to see Dell develop.”
Prior Experience Relationship to Future Daily: Feb 2007 – Feb 2009
Performance 7,100+ ideas
4,300+ ideators
# prior good ideas not significant 170 ideas implemented
# prior reviewed ideas not significant
# prior ideas not significant
# prior comments not significant
The value of the crowd is in the “crowd”
32. Modeling Connectivity in Online Networks
Ansari, Koenigsberg & Stahl
• Social network data helps to improve predictions of behavior above
and beyond just behavior
+ >
• More popular social networkers are also more active
• Online popularity is a more important correlate of online behavior than
offline
>
Knowing a customers social graph helps predict their purchases
33. Econometric Modeling of Social Interactions
Hartmann
Michael Michael’s
Promote to
goes golfing friends golf
Michael
more more
Fraction of customer’s
value that derives from
others in the group
Direct Value Indirect Value
65% 35%
Consumers bring additional value through their community
34. Opinion Leadership and Social Contagion
in New Product Diffusion
Iyengar, Van den Bulte, Valente
Target social influencers
Physician most often nominated by his
peers as influential is targeted and is
persuaded to increase his/her
prescription by 10 units
vs.
Across the board promotion
Each physician is given an
additional detailer visit
Influencers work, but slowly and “locally”
35. Pricing Digital Content
Iyengar, Abhishek, & Bradlow
Popularity begets
popularity; but
how do you get it?
But, free is free!
Freemium works!
40. S O L U T I O N : K E E P W H AT I S N E E D E D , F I T W H AT I S T H E R E
41. SUMMARY
• It is all about the data!
o In many cases, practitioners have it – academics want it.
o Scraping programs mean we can now all have it and in real‐time.
• Convergence of problems between academia and practice, in the
interactive media space, has never been higher.
o Advances still need to be made on scale of academic methods.
• Let’s look for the next great divide! It demonstrates an opportunity for
further study.
42. W H A RTO N I N T E R A C T I V E M E D I A I N I T I AT I V E
Eric T. Bradlow
ebradlow@wharton.upenn.edu
www.whartoninteractive.com
W H A R T O N I N T E R A C T I V E M E D I A I N I T I AT I V E