Transcript Marketing Mix Modelling

Market Mix Modelling
Estimate the effectiveness of
investment in media
Agenda
• Business application of Marketing Mix
modelling
• A case study
• Strengths and weaknesses
• Brief introduction to more advanced
approaches: pooled regressions and structural
equations
Making BP’s media dollars work harder
• “Mindshare helped BP to make the most of their media
investments across the many states of the USA.”
• “BP engaged Mindshare to develop enhanced media
investment strategies to maximise sales and boost revenue
performance.”
• “Drivers of performance were quantified (e.g. media,
promotions, distribution, competitor effects) in seven USA
states, over three years”
• “Return on investment figures were calculated - both short
and long term - for 40 campaigns.”
Marketing Mix modelling
• Statistical methods applied to measure the impact of
media investments, promotional activities and price
tactics on sales or brand awareness
• Used to assist and implement a marketing strategy by
measuring:
– Effectiveness: contribution of marketing activities to sales
– Efficiency: short term and long term Return-OnInvestment of marketing spend
– Price elasticity
– Impact of competitors
MMM How does it work?
• A statistical model is estimated on historical data with sales as
a dependent variable and list of explanatory variables as
marketing activities, price, seasonality and macro factors
• The simplest and broadly used model is linear regression:
Salest    1  var1t  2  var2t  ...  t
• The output of the model is then used to carry out further
analysis like media effectiveness, ROI and price elasticity and
to simulate what-if scenarios
Factors that could drive sales
Advertising
TV
Radio
Print
Outdoor
Internet
Promotions
Sponsorships
Events
Price
Adv quality
Distribution
Merchandising
Competition
Seasonality
Weather
Economic
Demographic
Industry data
Salest    1  var  2  var  ...  t
1
t
Sales
2
t
MMM project process
Set out objectives
Data preparation
-Define scope
-Discuss data
availability
-Design data-warehouse
•Collect data
•Validate, harmonize
and consolidate data
•Present exploratory
analysis to client
Presentation
Model development
•Interpretation of
results
•Learning and
recommendations
•Estimation
•Diagnostics
•Calculate ROIs, Price
elasticity and response
curves
Case study
• An energy company SPetrol wants to evaluate the advertising
investments of its retail business in the US from 2001 until
2004.
• Client’s questions:
• How much have we made through advertising?
• What is the return on investments of our media activities?
• Which marketing drivers have had the greatest effect?
• What’s the influence of price on our sales?
• Are we optimally allocating our budget across products ?
Target variable
Advertising data
• The performance of TV and radio advertising is expressed in
terms of Gross Rating Points (GRPs) . A rating point is a
percentage of the potential audience and GRPs measure the
total of all rating points during and advertising campaign.
– GRPs (%) = Reach * Frequency
– Example: Let’s assume a commercial is broadcasted two
times on TV
1st time on air
2st time on air
25% of target
televisions are tuned in
32% of target
televisions are tuned in
GRPs
57%
Advertising data
• Spetrol has deployed 5 TV campaigns over the
sample with a total expenditure of 300 million $
• Each campaign lasted from 4 to 8 weeks
• Is there any relationship between sales and TV
advertising?
Carry over effect of TV
Carry over effect of TV
• The exposure to TV advertising builds awareness,
resulting in sales.
• ADStock allows the inclusion of lagged and non
linear effects
ADStockt ( )  GRPt    ADStockt 1
0  1
• Alpha is estimated iteratively using least squares.
The estimate is then validated by media planners
Advertising data
300 M
TV Spend
164 M
Radio
160 M
Outdoor
Below the line promotions
• It may include
– sponsorship
– product placement
– sales promotion
– merchandising
– trade shows
• Usually represented by dummies (variables
equal to 1 when a promotion takes place and
0 otherwise)
Below the line promotions
Sponsorship
World Rally
Championship
Sale promotion
Sale promotion
5% Discountt
Price
Seasonality
August seasonal dummy
5% Discountt
Peaks every year
in August
Sale promotion
Exploratory analysis
Scatter plot
32
Unit root test
Histogram and desc stats
Series: SALES
Sample 1 209
Observations 209
28
24
20
16
12
8
4
0
130000
140000
150000
160000
170000
180000
Mean
Median
Maximum
Minimum
Std. Dev.
Skewness
Kurtosis
154403.1
153960.2
183102.5
125997.0
9476.290
0.053546
3.456209
Jarque-Bera
Probability
1.912312
0.384368
Correlation matrix
Model development
Estimated equation
Salest = 167412 +
168* AdStock(GRPsTVt,0.75) +
161* AdStock(GRPsRadiot,0.35) +
166* AdStock(Outdoort,0.15) +
580* PromotionDummyt +
6507* Seasonalityt +
-12631* Pricet + Errort
Model diagnostics
• Model:
– Significant F-stat and high R-squared
• Variables:
– Significant T-stats
– Coefficients must make sense
– Variance inflation factor low
• Residuals:
– Normality (Jarque-Bera)
– Absence of serial correlation ( Durbin Watson,
correlogram)
Residuals diagnostics
16
Series: RESID
Sample 1 209
Observations 209
14
12
10
8
6
4
2
0
-10000
-5000
0
ˆ  y  yˆ
5000
Mean
Median
Maximum
Minimum
Std. Dev.
Skewness
Kurtosis
-2.31e-11
-66.11295
8049.987
-11378.69
3612.711
-0.158326
2.624286
Jarque-Bera
Probability
2.102443
0.349511
Durbin Watson = 1.69
DW>2 positive autocorrelation
DW<2 negative autocorrelation
Estimated factors contribution to sales
Fitted Salest = estimated Intercept = 167,412
Can be interpreted as Brand Equity:
•Volume generated in absence of any marketing
activity
•Indicator of the strength of the brand and users’
loyalty
Estimated factors contribution to sales
TV Contributiont(000’ Gallons) =
coefficient *Adstock(TV)t
Fitted Salest = 167,412 + 168* TVt + 161*Radiot +
166* OOHt + 580* Promotiont
Estimated factors contribution to sales
Peacks every year
in August
Peaks every year
in August
Fitted Salest = 167,412 + 168* TVt + 161*Radiot +
estimated Intercept
= Seasonaility
167,412
166* OOHt Equity
+ 580*= Promotion
t + 6507*
t
Can be interpreted as Brand Equity
Estimated factors contribution to sales
Fitted Salest = 167,412 + 168* TVt +
161*Radiot +
166* OOHt + 580* Promotiont + 6507*
Seasonailityt - 12631* Pricet
Negative price
effect
Marketing mix (sample output)
Estimated factors contribution to sales
Estimated factors contribution to sales
N
TotSalesContribution coeff   Factori
i 1
Estimated factors contribution to
revenue
N
Tot Re venueContribution coeff   Factori  Pricei
i 1
ROI
ROI 
TOT Re venueContr ibution
TOTCost
Does it really make sense?
TheDiminishing
more I invest in
media, returns
the more I sell
Response curves
NegExp  a  (1  exp(b  GRPs))
S  a  (1 /(1  exp(b  (GRPs  mean(GRPs))))
Taking into account
diminishing returns
Price elasticity
• Assumption: constant elasticity across the sample which
implies a linear relation between volume and price
• By using the coefficient of the regression, it is possible to
derive an estimate for price elasticity:
– Price coefficient = -12631
– Average price = 1.51 $
– Average volume sales = 154,000 Gallons
Avg Price
Elasticity 
* coeff  0.12
AvgSales
A 10% drop in price
increases sales by 1.2%
Dynamic price elasticity
Elasticity changes with price
200,000
Weekly Volume and $ Sales vis-à-vis price of 1.75L
180,000
Volume (9L Cases)
160,000
140,000
120,000
100,000
80,000
60,000
40,000
20,000
Elastic (>1): Demand is sensitive to price changes.
Inelastic (<1): Demand is not sensitive to price changes
30
29
28
27
26
25
24
Price (750 ml)
23
22
21
20.0
Volume
19
18
17
16
15
14
13
12
11
10
9
0
Estimated through non
linear regressions
Client’s questions
How much have we made through advertising?
• 1 billion $ driven by TV
• 500 million $ due to radio
• 200 million $ generated by Outdoor and
promotional activities
Investments in media generated 1.7
billion $ in revenue
Client’s questions
What is the return on investments of our media
activities?
For each dollar invested in TV you get 3.5 dollars
back
Client’s questions
What’s the influence of price on our sales?
A 10% drop in price
increases sales by 1.2%
Are we optimally allocating our
budget across products ?
Maximum
Marginal
Return
Optimal
GRPs
Over Optimal GRPs
Point of
Saturation
Sub –Optimal GRPs
Maximum
Average Return
Invest more in Radio
and less in OOH
Marketing Mix – Sample Output
Marketing mix (sample output)
45
Diminishing Returns
5000
35
4500
Promo TV Saturation
3500
3000
2500
Current
2000
Optimal
30
Weekly GRPs
4000
Weekly Sales
Carry Over Effect
40
1500
25
20
15
10
1000
5
500
0
0
0
20
40
60
80
100
120
140
160
180
Week1
Week2
Week3
Week4
Week5
Avg. Weekly GRPs
Diminishing Returns is the point were spending
additional GRPs does not results in additional
sales.
Simultaneous Effect
Volume
Carry Over Effect (Ad Stock) relates to the
residual effect of an ad.
Base/Seasonal
TV/Radio/Print
Direct Marketing
Time
Rates/Promotions
When all the components are layered on Base
sales, it is clear what drivers contribute to sales
and when and their Simultaneous Effect.
Pros and cons
• Simple and intuitive
• The outcome is backed by
qualitative expertise and in
field research
• Constructive way of running
different scenarios and
evaluating past
performance
• Better with granular data
• Very successful method –
high turnover
• Correlation doesn’t imply
causality
• Risk of spurious regressions
especially when modelling
in levels
• Model highly depends on
variables chosen
• Poor in forecasting
Spurious statistics
• A high correlation
between sales and TV
could mean:
Sales
Media
Income
– Either media causes
sales
– or sales causes media
– or a third variable causes
both sales and TV
What is the truth?
Non sense correlations
• Some spurious
correlations:
– death rate and
proportion of marriages
Corr = 0.95
– National income and
sunspots Corr = 0.91
– Inflation rate and
accumulation of annual
rainfall
• On the other hand, a
low correlation doesn’t
rule out the possibility
of a strong relation:
Corr = 0.0
•Correlations must support a theory
•Calculate correlations both in levels and differences
•Always look at scatter plots
What variables should have been
included?
New media
• Digital Marketing
– Display Marketing
– Search Engine Marketing (SEO & PPC)
– Affiliate Marketing
– Mobile Marketing
– Social Media
New media
• Data availability
– Impressions
– Clicks
– Post event activity
– Bespoke engagement metrics
• Example of a tracking centre:
– Double-click
Alternative methods
•
•
•
•
•
•
Linear regression
Logistic regression
Discriminant analysis
Factor analysis
Cluster analysis
Structural equations modelling
Pooled regressions
Sales
Local media
Nat media
Local Price
California
California
USA
California
+ ... + error
Nevada
Nevada
USA
Nevada
+ ... + error
Oregon
Oregon
USA
Oregon
+ ... + error
sa
Pooled regressions example
1. SalesCalifornia = c11*TVCalifornia +
c12*TVOregon+c13*RadioCalifornia +c14*RadioOregon +
ErrorColifornia
2. SalesOregon = c21*TVCalifornia +
c22*TVOregon+c23*RadioCalifornia +c24*RadioOregon +
ErrorOregon
 TVC
 SalesC   c11 c12
 Sales   c
O

 21 c22
c13
c23
Media effect is also tested across regions

c14   TVO   C 

 

c24   RadioC   O 


Radio
O

How advertising effects consumers?
Understanding:
– the process by which advertising affects
consumers
– How the effects of advertising are spread over
time
– The role of different media
– The role of competitors
The purchase funnel
• A basic process that
leads to the purchase of
a product consists in:
– Awareness – costumer is
aware of the existence of
a product
– Consideration – actively
expressing an interest in
the company
– Purchase
Awareness
Consideration
Purchase
Working on survey data
• A sample of the target
audience is interviewed
about brand awareness,
consideration and choice
• Research agencies provide
awareness, consideration
and purchase time series in
% terms
– i.e. A purchase of 10% means
that 10 out of 100 interviewed
people purchased the product
Testing the purchase funnel
Awareness
Media
Consideration
Purchase
Advertising first exercise its
influence on awareness. Via
awareness there is an effect on
consideration which drives the
consumer to purchase
Testing the purchase funnel
• Awarenesst=c11+c12*TVt+c13*radiot+c14*OOHt+error1t
• Considerationt = b1*awarenesst + c21 + error2t
• Purchaset = b3*Considerationt + b2*Awareness +c31 +
error3t
a1,a2,a3 must be insignificant to confirm theory
 1
 b
 1
 b2
 a1
1
 b3
 a2   Awart   c11
 a3    Const   c21
1   Purcht  c31
c12
c13
0
0
0
0
Const 
c14  
  1t 

TVt   


0 
  2 t 
 Radiot 
0  
  3t 
 OOHt 
Agenda
• Business application of Marketing Mix
modelling
• A case study
• Strengths and weaknesses
• Brief introduction to more advanced
approaches: pooled regressions and structural
equations
References