Case Study Interactive: How To Work With Structured And Unstructured Data To Increase Customer Acquisition And Reduce Churn With Relevant Communication
How can analytics improve your attribution model accuracy to highlight and transform your most successful marketing channels?
How can you introduce predictive analytics to increase your customer segmentation competency?
How can insights from consumer data help you to predict customer lifetime value and focus on your top customers?
How can split testing consumer data help to improve your customer offering and boost retention rates?
1. Case Study Interactive: How To Work With Structured And Unstructured
Data To Increase Customer Acquisition And Reduce Churn With Relevant
Communication
2. Harvinder Atwal
MoneySuperMarket.com
Web
dunnhumby
• previous : Insight Director, Tesco Clubcard
Lloyds Banking Group
• previous : Senior Manager, Customer Strategy and Insight
• Head of Data Strategy and Advanced Analytics
@harvindersatwal
British Airways
• previous : Senior Operational Research Analyst
{“about” : “me”}
@gmail.com
3. 3
£1.8B
SAVINGS
2016 estimate total of UK savings
1993 22M 6M MSM 14M MSE £316M 980
We started life as
mortgages 2000
Adults choose to
share their data
with us
Average monthly
users
2016
Revenue
2016
Providers
4. How can analytics improve your attribution model
accuracy to highlight and transform your most
successful marketing channels?
How can you introduce predictive analytics to
increase your customer segmentation
competency?
How can insights from consumer data help you to
predict customer lifetime value and focus on your
top customers?
How can split testing consumer data help to
improve your customer offering and boost
retention rates?
What you wanted to know
5. Warning: A data-driven customer
focussed strategy will not paper
over cracks in operational
performance or product deficiency
13. Last
View
Linear
or Fair
Share
First
Click
Linear or
Weighted
Share
Assumes only the “last
viewed” advert, email
or click counts – no
earlier activities are
given share of the
credit.
Weightings can be
arbitrary and need to
be constantly
updated
Not all interactions are
equally
Valuable. Not all activity
can easily be counted
e.g. offline
Assumes only the
first activity counts
– no later activities
are given any
credit
17. Display
Ad
Video
Ad
Social Email Search Website
Physical
Store
TV/
Radio/
Press
Outdoor
$88 $132
Time
Measure the direct
effect
What is the impact
of Outdoor
advertising on sales?
21. Display
Ad
Video
Ad
Social Email Search Website
Physical
Store
TV/
Radio/
Press
Outdoor
$88 $132
Time
Repeatedly iterate and
model. You can then
apply weightings
What is the impact
of TV/Display,
Video, Social…
spend on sales?
28. Think actionable needs, preferences
and states
Borrowing
Saving
Risk averse
Price-sensitive
Brand conscious
Financially Cautious
Financially confident
Time poor
29. Exercise:
What are some of the
actionable Customer
needs, preferences and
states for your
organisations?
30. Time of day responsiveness
Day of week responsiveness
Device preference
Marketing channel preference
Offer responsiveness
Help preferences
Social proof/review responsiveness
31. Traditional propensity modelling and
recommendation engine techniques can help
you if you have past outcome data
Customer
Data
Model
Highest probability
+
32. You can also think of customer history as a
sequence and predict using Deep Learning
Email
Open
Page view
Product
click
Product
click
Time
Customer history
Sale?
Future period
RNN
Cell
RNN
Cell
RNN
Cell
RNN
Cell
Prediction
34. Show pictures of
cats
Show pictures of
dogs
Show pictures of
People (control)
Test treatments at random
Conversion = 5% Conversion = 3% Conversion = 3%
But we’re not interested in which treatment
works best on average
35. Find the best treatment for each customer
Total Customers
(100% of customers)
(3% conversion)
Live alone
(30% of customers)
(4% conversion)
Don’t live alone
(70% of customers)
(2.6% conversion)
Urban
(56% of customers)
(2.7% conversion)
Rural
(14% of customers)
(2.1% conversion)
Live in apartment
(9% of customers)
(4% conversion)
Live in house
(21% of customers)
(4% conversion)
Cat conversion = 18% Cat conversion = 1% Cat conversion = 5.4% Cat conversion = 1%
Dog conversion = 2% Dog conversion = 2% Dog conversion = 2.5% Dog conversion = 7%
Cat segment People segment Dog segmentCat segment
Total segmented conversion =
6.5% vs 4% for best treatment
on average (Cat pictures for all)
36. A finite number of
predictive micro-
segmentations can
be combined to
create highly
personalised
individual
experiences
37. Test &
Collect
Model Embed Roll Out
Feedback
Plan
Pilot test
Collect Data
Build Model
Identify segments
Adjust model to fit
organisation
Re-engineer business
processes to support
segmented execution
Train organisation
Incorporate segments into
daily execution
Provide differentiated
services, products and
content
39. Traditional techniques like RFM and Pareto-NBD omit
many factors influencing Customer Lifetime Value
Contribution
Time
Buys second product
Complaint
Loss Leader
High Servicing costs
Complaint
resolution
Subscription revenues
40. Training Features
Random Forest Regression can create more
accurate CLV predictions
Training period Model Test period
Training period Prediction period
Time
Product Purchases
VisitsSpend
Demographics
Acquisition channel
Complaints
Future period
Location
Historic period
Segmenting models may improve accuracy further
User BehaviourShipping preferences
Payment preferences
Costs
47. Do not spend time AB testing small
cosmetic details
Simple UI changes are
ineffective.
Colour (changing the colour of
elements on a website) +0.0% uplift
Buttons (modifying website buttons) -
0.2% uplift
Calls to action (changing the wording
on a website to be more suggestive) -
0.3% uplift
Best test categories are:
Scarcity (stock pointers) +2.9% uplift
Social proof (informing users of others’
behaviour) +2.3% uplift
Urgency (countdown timers) +1.5%
uplift
Abandonment recovery (messaging to
keep users on-site) +1.1% uplift
Product recommendations (suggesting
other products to purchase) +0.4% uplift
Qubit meta-analysis of 6,700
experiments (2017)
50. You’re testing promotion of a new product in
an email campaign
What is the target variable?
C) Revenue per customer
B) Sales of the product
A) Click-through on the email
51. You’re testing an outbound telesales campaign
What is the unit of measurement for the
target variable (sales)?
A) A call
C) A telesales agent
B) A customer
52. A null hypothesis H 0 ('no effect') is tested against an
alternative hypothesis H 1 ('some effect'). The results
pass a test of statistical significance (P-value <0.05) in
favour of H 1.
What’s been shown?
1. H 0 is false.
2. H 1 is true.
3. H 0 is probably false.
4. H 1 is probably true.
5. Both (1) and (2).
6. Both (3) and (4).
7. None of the above.
Enough about me and MoneySuperMarket. From the research these are the questions you wanted answered.
Last view
Using this method, it’s not the click that counts but the last advert viewed
It’s important to have a vision of what you’re trying to achieve
Predictive Analytics is moving us from a world of having to Explicitly state our needs to having them Implicitly fulfilled
There’s no single segmentation
Use predictive analytics to understand actionable customer needs and preferences, without having to be told, so you can customise their experience
During World War II, the statistician Abraham Wald took survivorship bias into his calculations when considering how to minimize bomber losses to enemy fire. Researchers from the Centre for Naval Analyses had conducted a study of the damage done to aircraft that had returned from missions, and had recommended that armour be added to the areas that showed the most damage. Wald noted that the study only considered the aircraft that had survived their missions—the bombers that had been shot down were not present for the damage assessment. The holes in the returning aircraft, then, represented areas where a bomber could take damage and still return home safely. Wald proposed that the Navy instead reinforce the areas where the returning aircraft were unscathed, since those were the areas that, if hit, would cause the plane to be lost
A/B Testing
Summary: In an A/B test, a change is applied to a treatment group, and its performance is compared against a control group to estimate the impact of the change.
STEP 1: SELECT A PERFORMANCE METRIC
It’s important to understand the metric used to evaluate the results of the test. Whether the goal is to increase sales, profit, conversion rate, etc., this should be specified at the upfront.
STEP 2: SELECT THE EXPERIMENT DESIGN
Matched pair - when the sample size is small and/or the data is difficult to collect, a matched pair experiment should be used.
Randomized design - when the sample size is large and the data is easy to collect, then a randomized experiment should be used. Randomized experiments are very common for web-based AB tests.
STEP 3: SELECT TREATMENT AND CONTROL UNITS
Each individual in the test is considered a unit. The unit can be a person, store, etc. In a test, units are split into two groups, the treatment group and control group. Treatment and control units are compared against each other
STEP 4: SELECT EXPERIMENTAL AND CONTROL VARIABLES
Experimental variable - The experimental, or treatment, variable(s) is the variable that is different between treatment and control units. For example, if you are testing a new price point, the experimental variable would be price.
Control Variables - The control variables are the variables that should remain constant between test and control groups. These variables ensure that the treatment and control groups are representative of each other and that the results will apply to the population. Control variables are used to match each treatment unit to one or more control units.
STEP 5: DETERMINE TEST DURATION AND SAMPLE SIZE
These two go hand in hand and contribute most directly to statistical significance. You can improve statistical significance by either increasing the sample size or test duration. Generally the duration of a test should be at least as long enough to capture a representative sample.
STEP 6: RUN THE TEST AND PREPARE THE DATA
Now it’s time to run the test and collect the data. Preparing the data includes filtering for the dates of the test, ensuring there are no duplicate records, removing records with incomplete data, and removing outliers.
STEP 7: ANALYZE RESULTS
Lift - Compare the average performance between the two groups. It can also be useful to understand the distribution of the performance of the units.
Statistical Significance - Performing a t-test provides a p-value. P-values below 0.05, indicate statistically significant results. Paired t-test are used for matched pair experiments and unpaired t-test for randomized experiments.
Impact Estimation - In order to provide an expected impact of broad implementation of the treatment, apply the lift calculation to the entire population.