Market Mix Modelling in R
Why aren’t more market mix modellers using R?Steve Cookson | May 2016
Market Mix Modelling in R
• My Background• Market Mix Modelling (MMM)
– What it is– Process & Output
• Why R is underused in MMM• 3 challenges to MMM that R can help with
My Background
• Started Market Mix Modelling in 2005• Currently
My Mango Data Science Radar
Data‐driven technique to measure the impact
of marketing
Market Mix Modelling (MMM) Branch of Econometrics
Economics + Statistics
What is Market Mix Modelling?
Gather Data
Market Mix Modelling (MMM) Branch of Econometrics
Economics + Statistics
Quantify drivers of a KPI – e.g. sales
Plug in financial data to calculate ROI
Budget Allocation
Stakeholder Questions
Example Stakeholder Questions• Was my new product launch additive to the business or did it cannibalise?
• How does the stocking issue in Q2 affect our year‐end forecast?
• I have an additional $1MM budget. Should I invest in Germany or Spain?
• Should I advertise Brand A or Brand B?
• Should I increase my % spend behind digital media?
• Should I stick with my new TV ad, or revert back to the old one?
• How did the cold weather affect our summer holiday campaign?
• Can I increase price or should I reduce?
What data do you think we need for an MMM project?
• Can include (but not limited to):– KPI : usually sales– Media and/or marketing– Competitors– Climate– Holidays– Macro trends – e.g. economy– Stocking measures / shocks– Price and promotions
• Characteristics:– Over Time : Weekly, (Monthly, Daily or even Hourly)– Regional (where possible)
MMM Methodology• Majority: Time Series, Multivariate Ordinary Least Square (OLS) Regression
– Explaining why a KPI like sales changes over time– Correlate multiple factors at once– Minimise the difference between the actual KPI and the modelled KPI– Look for stable models that make sense
• For equation fans: yt = + x1t + x2t + … + ixit + t
MMM Output• Modellers tend not to show the equation details to clients• Instead, we present the model graphically
-10
0
10
20
30
08-J
an-1
205
-Feb
-12
04-M
ar-1
201
-Apr
-12
29-A
pr-1
227
-May
-12
24-J
un-1
222
-Jul
-12
19-A
ug-1
216
-Sep
-12
14-O
ct-1
211
-Nov
-12
09-D
ec-1
206
-Jan
-13
03-F
eb-1
303
-Mar
-13
31-M
ar-1
328
-Apr
-13
26-M
ay-1
323
-Jun
-13
21-J
ul-1
318
-Aug
-13
15-S
ep-1
313
-Oct
-13
10-N
ov-1
308
-Dec
-13
05-J
an-1
402
-Feb
-14
02-M
ar-1
430
-Mar
-14
27-A
pr-1
425
-May
-14
22-J
un-1
420
-Jul
-14
17-A
ug-1
414
-Sep
-14
12-O
ct-1
409
-Nov
-14
07-D
ec-1
404
-Jan
-15
01-F
eb-1
501
-Mar
-15
29-M
ar-1
526
-Apr
-15
24-M
ay-1
521
-Jun
-15
19-J
ul-1
516
-Aug
-15
13-S
ep-1
511
-Oct
-15
08-N
ov-1
506
-Dec
-15
Sale
s (T
onne
s)
Week Ending
Brand XRegular Pr ice Stocking Market Seasonality & Compet it ion Advertising Promotions Volume
MMM Applications• Price Elasticity :
– For every % increase in price, by what % will my sales fall ?• Rule of thumb: if the price elasticity < 1, then raising prices will increase profits
• Return on Investment (ROI) : – For every £1 I spent, how much did I get back in sales (or profit) ?– Example calculation:
• From a market mix model, we know that advertising generated 5,000 units of sales, worth £100,000• We also know that we spent £300,000 on advertising• So the ROI = £100,000 / £300,000 = 0.33
• Budget allocation :
Revenu
e
Spend
Revenue A Revenue B Revenue C
MMM alternatives to R• Most market mix modellers don’t use R
• Either:
• Or MMM‐specific software:
• Or create their own software in‐house
Why not R ?• Possible reasons?
– Economics graduates using E‐Views and STATA ?– Licensing issues ?
• Key issue: speed– Many MMM data sets are still small enough to manipulate in Excel
– Click and point is faster than command‐driven model building• CommandR package is basic
– Churn out standard MMM‐specific charts
Can R help with challenges to MMM?
• Challenge #1‐ Explain the MMM results• Simply reporting ROI is not enough • We must explain why ROI has changed
• One way involves storing models and their results• Run (cross‐sectional) meta‐analysis of modelling results
• R can store modelled results as variables / values for later use• R is better suited to cross‐sectional modelling than existing MMM software
Can R help with challenges to MMM?
• Challenge #2‐ Forecasting and optimising
• Static reporting is not enough • Clients are looking for scenario planning tools as standard
• R (via Shiny) can provide a direct link between models and an interactive output
Can R help with challenges to MMM?
• Challenge #3‐ Alternative techniques +
• MMM already exists in parallel to Digital Attribution and Single‐Sourcemodelling
• These methodologies could merge in the future• Both use larger data sets than MMM and more complicated statistical techniques than Ordinary Least Squares (OLS)
• R can handle larger data sets / automated links• R is more versatile than existing MMM software – goes beyond OLS
Thank You
Questions?