iab sm measurement talk slides 19 oct slideshare version
TRANSCRIPT
predictive modelling
and
benchmarking
with
social media metrics
a
road map for development
Stephen Haggard, Z/Yen Group Limited
Presentation to IAB Social Media Council, Research and
Measurement Group 19 October 2010
over half of marketing budgets
spend <1% on social media Emailvision 2010
social media data framework and weighting
we do not have:
known relationships
standard regressions
organising principles
comparable datasets
we do have:
outcome criteria
eg CPR>$x
volume of data
social media data narrative will be fuzzy
“solve wicked problems”
social media
in
professional
communities
avatars
as
interfaces
to
data worlds
risk analysisoutcome prediction
Support Vector Machine (SVM)
Predicting outcome variables from datasets with complex
multiple and dynamic correlations and patchy/noisy data
analysis by hyperplane fit
applications
PropheZy
major gift prediction
(Big Lottery Fund)
straight-through-
processing in dealing
rooms
union branch
performanceDynamic Anomaly and Pattern Response
(DAPR)
media applications / buzzdeck by AWAL
/rural network for MTN Zambia
DAPR machine - PropheZy
best execution trading: compliance analysis
live demonstration available – contact me
next - data
Historical data tests
gauge reliability
build confidence
Build model of correlation
training sets
Harvest ongoing campaign datasets across time
non-numerical is OK eg sentiment index, content tags,
external content OK eg weather, interest rate, other
media, Google Analytics
next– format ?
proprietary specialist tool
industry-standard tool
data sharing club
next - what is it ? benchmark for performance
prediction and planning tool
analysis and evaluation tool for insight
Scenarios in social media - prediction
Scenario: predictive use.
Working with the target values for domains given in IAB metrics proposal, the Agency is planning a SM campaign with KPIs: {i(A) Appreciation: CPE < €0.20} and {i(A) Action CPL but {i(A) Awareness n 3}
The DAPR machine looks in the data for instances that yield this outcome and assigns a p=20% probability of success for the input scenario. But the DAPR machine has also established (by benchmarks across the whole community of data) that variables d, g and j are reliable predictors of the target CPE for this product type. So the Agency enters new target values for fields d, g, j and this yields p=50%, allowing the planners to target their resources more smartly.
Scenarios in social media - evaluation
Scenario: evaluation.
The client is midway through an SM campaign that
has attracted 275,000 visitors so far (on target) but
has yielded only 900 plays of its video, way below
target. How bad is this, and should the Agency be
fired?
DAPR will spot the extent to which this outcome is
anomalous and below/above trend for this point in a
campaign of this or any kind. Data points can be
tracked in 3 dimensions (time, number, type).
Scenarios in social media - analysis
Scenario: analysis.
SM campaigns on Platform X are achieving
bookmarking and referrals at 12 % more per visitor
than those on Platform Y. Why ?
DAPR will identify the strongest patterns around
bookmarking and referrals both across the entire
data pool, and then on Platforms X and Y; the
clusters of difference that correlate best with
bookmarking and referral scores will be identified for
both Platforms X and Y. Histogram analysis shows
best predictors of bookmarking referral.