5. 5
Every outlet is
similar…but different!
Different sizes, different
customer base, different staffing
levels, different formats etc…
Unfair to compare
outlets directly
An outlet with lower footfall and
excess staff will outperform an
outlet that is under pressure!
Need to control for
differences
We use statistical modelling to
level the playing field so that a
direct comparison is possible…
8. 8
Data
Range of variables to test means data
availability is not consistent – weekly,
monthly, quarterly etc.
Continuous data needs to be rolled up to
the relevant frequency used for modelling
Data on a number of factors will be point
in time and relatively static over time
Some data may shift seasonally
Typically use a single data point per factor
per outlet – latest quarter / 6 months /
year
9. 9
Modelling Approach
Pre-analysis, Data Visualisation &
Correlations help to form initial picture of
variable relationships and shapes
Typically, no data transformation
Linear multiple regression modelling
– Don’t tend to see exponential
relationships
– Ease of interpretation
– Other model forms are available!
Updates determined by client business /
reporting cycle
15. 60%
40%
Uncovering existing knowledge
Identified drivers of
redemption through
regression modelling (type of
customer, type of
supermarket, time of year, …)
& response based sampling
Used as covariates in
CBC analysis in step 4
16. 16
EFFORT LEVELS
€
GENEROSITY
LEVELS
X
X X X
X XX
X
X
X
X
X
X
X
X
X
X X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
AND
ACCORDING
TO YOUR
PREVIOUS
BEHAVIOUR...
XCBC DESIGN
WITH TWO
ATTRIBUTES
+ one control group of consumers
with no points offered
17. 17
CBC DATA =
REAL
BEHAVIOUR
DATA
One CBC choice task per consumer:
a specific effort/generosity combination that can be
redeemed or not
Shopping data showed whether redeemed or not
(=answer to single CBC choice task) and level of
(extra) spend versus control group
31. 31
Conjoint design details
CBC design with 2 attributes:
number of points
required spend on next basket (also called effort)
Not all combinations were
possible – as only realistic
scenarios were tested
Spread of CBC design over
sample: dependent on average
basket value of past 3 Months
= points/effort
32. 32
Conjoint analysis details
HB estimation
Full design split in 3 designs for
realistic results (i.e. to avoid
borrowing data from respondents
with too different shopping behaviour
– much lower or much larger
shopping basket values)
Contraints on both attributes
Given the little data per respondent (one CBC task) a lot of data
borrowing required
Looked at several PV and DF settings
Final choice: PV = 1, DF = half the sample size
3 sub designs
Each estimated
separately
through HB
Editor's Notes
When I talk to companies about this most people tell me they expected me to be older
Performance Drivers – the model obviously quantifies the impact and relative importance of each of the key drivers
Benchmarks – performance assessment vs local operating conditions
Targeting – more differentiation in target setting, accounting for relevant benchmark, current performance and ability to change / influence local operating conditions
This can get serious! We have worked with some clients that link bonus payments to these targets.
Data – given the wide range of variables to test, data availability is not consistent – some weekly, monthly, quarterly etc.
Some data will be continuous – sales, transactions etc. – and needs to be rolled up to the relevant frequency of data that will be used for modelling
Some data will be point in time and relatively static over time – e.g. outlet grade
Some data may shift seasonally – demographics, customer mix
How did we get on with this????
Big DATA and MR – introduce it here?
Conjoint
Talk here about big data, i.e. where the big data element comes from