Optimising price and marketing mix.
Concept of learning. When an account/product has too little sales data, bayesian shrinkage allows us to borrow information from other accounts.
Deals with outliers, by shrinking estimates towards each other.
Allows one hierarchical model instead of multiple models.
More robust, stable estimates with significant regional and account variation in estimates that cannot be done in a classical linear model.
Provides price elasticity measure that shows the impact of price changes on volume
3.
How much do sales increase when we run a
temporary price reduction?
When I promote do I cannibalise my own sales?
When my competitor promotes do they take sales
from me? ......If so, how much?
What effect would a 4% price rise have on my sales
?
5.
Concept of learning. When an account/product has
too little sales data, bayesian shrinkage allows us to
borrow information from other accounts.
Deals with outliers, by shrinking estimates towards
each other.
Allows one hierarchical model instead of multiple
models.
More robust, stable estimates with significant
regional and account variation in estimates that
cannot be done in a classical linear model.
6.
Multilevel/hierarchical model
Centering is a helpful way of parameterising
models so that the results are easily interpreted.
Fixed effect and random effect.
SAS Proc Mixed to fit the hierarchical model.
7.
We use the Restricted Maximum Likelihood (REML)
as the estimator and the Newton-Raphson as the
search algorithm.
To gauge the fit of the model, we use Akaike’s
Information Criterion (AIC) and Schwarz’s Bayesian
Criterion (BIC).
The estimator generates reasonable MAPEs at the
total National level of about 1% to 6%.
8.
Measures the impact of price changes on volume
Price Elasticity = %Change in Sales
%Change in Price
For example, if an item had a price elasticity of
-1.5,it would lose 15% of its volume if it raised its
price by 10%, (β1= -1.5).
9.
For example, the lift due to catalogue is:
Example
Catalogue
Vol = Vol*exp(Beta)
Beta (β8 )
0.49
exp(Beta)
1.49
Lift
49%
10.
Marketing activities that are not found statistically significant by
the model are deemed to be ineffective.
Based on the size of the promotional price elasticity, clients are
recommended to continue/discontinue temporary price
reductions.
Based on the ‘Return of Investment’ for each marketing
activity, clients are advised which marketing activities best
increase their sales.
A simulation tool, using the coefficients from the model, is
presented to clients so that they are able to simulate and plan
their future strategy, so that they obtain the best profit possible.
11.
Price Elasticity
- Challenge: A declining sales in salad
dressing. A regression on historical data
suggested that reducing price would
significantly boost sales
- A category manager was told by market
research that the price elasticity for one of
the products was -2. The product, however,
still had strong sales in spite of several
increases. (Competitive Reaction and time
horizon).
12.
There is most times insufficient variation in base
price to accurately estimate its elasticity with scanner
data.
A major source of controversy is the appropriate level
of aggregation at which to study advertising effects
with regression. While some agents argue in favour of
store or account level analysis, third party consultant
argue in favour of market level analysis.
A careful comparative study of the estimation of
advertising effect sizes with store level versus market
level data would be of substantial interest to
practitioners who come into contact with this issue.
13.
What is the duration or length of the long
term effect of advertising. Some researchers
say 6months others show a duration of well
over 6months and others show that duration
might last several years.
14.
Studies are needed to determine whether or not
base price elasticity’s can be reliably estimated
from scanner data and, if so, what is the extent of
natural price variation or number of observations
needed.
These estimates could also be compared with and
tested against those obtained from survey-based
methods such as discrete choice analysis.
15.
Research is needed to resolve aggregation issues in
assessing advertising effects. Determining the best
approach to assess advertising effects with scanner
data could resolve some of the methods and data
aggregation debate in this area.
Manufacturers and retailers could benefit from
methods to determine the costs and benefits of
broader versus narrower product assortment
(e.g., Number of flavours and/or varieties)
16.
Establishment of a ‘Methods Standards’ for Scanner
Data Analysis.
Determine the generalizability of empirical results
(e.g., meta-analysis of price elasticity or
advertising elasticity) which helps provide
reasonable bounds when estimating models.