What KPI can we model?

KPIs for MMM should be defined per each analysed period. It should be something that we measure weekly or monthly.

Most common KPIs are:

sales volume
sales value

The best results we get modelling sales either by volume of by value. More often we model sale volume. Because then we can also include price into the model.

Other KPI that can be considered

awareness
store traffic
webpage traffic

We can model different KPIs than sales but it depends on the brand and values that we have. Awareness is first of all difficult to measure and values usually do not change much over time.

Adstock functions in MMM

Adstocks represents how long and how strong ads have an impact over time. It means how long is affect KPI, after the ad is displayed. Two main groups of functions for adstock are:

geometric: adstock_i = ad_i + 𝞭 * adstock_i-1
convolution: adstock = 𝜮conv ⨂ ads

Main pros and cons of adstock functions

function	pros	cons
geometric	easier to calculate	ad effect lasts forever
convolution	defines how long ad effect lasts (how many days or weeks)	more difficult to calculate

Please check the example; it is easier to understand with on live example

Exaples calculated for ad spend week1 = 10 and week2 = 5

convolution function:

parameters	conv₁	conv₂	conv₃
parameters	1	0.5	0.2

calculation

week	x [$]	adstock	*x_iconv₁ + x_i-1conv₂ + x_i-2conv₃**
w1	10	10	= 10 * 1 + 0 * 0.5 + 0 * 0.2
w2	5	10	= 5 * 1 + 10 * 0.5 + 0 * 0.2
w3	0	4.5	= 0 * 1 + 5 * 0.5 + 10 * 0.2
w4	0	1	= 0 * 1 + 0 * 0.5 + 5 * 0.2
w5	0	0	= 0 * 1 + 0 * 0.5 + 0 * 0.2

MMM vs. linear model

MMM is not just a linear model; we have to make a few adjustments

Adstock – we assume that sometime the effect lasts after the contact with the ad:
- adstock depends on medium type
- in the case of TV ads, it may be a few weeks; for an online ad about a week
Saturation – there is a limit; by investing in media, we cannot increase a demand infinitely:
- in case of media investment 2 ∗ 2 ≠ 4
Possible lag – some actions may take some time to affect the KPI:
- it may have an effect a week or a day after contact with an ad
Predefined constraints:
- own action – affect our KPI positively
- competitors – affect our KPI negatively; of course, there are exceptions,
  - some competitors’ activity may influence the whole category
  - sometimes, people don’t know which brand is promoted in the ad

Why to do MMM?

yes to: statistics and data

no to: cookies and polls

MMM is the best way do give budget recommendation having many marketing channels, which are not connect. Because:

no cookies data required
do not require unique user data
not based on declarative survey
based on actual numbers and statistical models

Panel data give some insight but they do not cover all the media. It is possible that in our case panel cover all of our activities. Still it is a sample data. If your brand is small and there is not a lot of activities it may not be captured.

Declarative survey are a great tool for example to understand brand awareness. But they have issues. People usually don’t remember what and where they saw in terms of media and ads. Or they think that they remember.

It is difficult to attribute the effect of marketing activities from online ads, radio, press or on-site. Even within online space it is difficult to measure effectiveness e.g. of ads and influencer marketing. The most objective way to create statistical model aka MMM (Marketing Mix Model)

What data do we need?

KPIs:

ads:

other promotions:

holidays:

For a good model, we need to collect as much information about factors that impact our KPI as possible.

Most common elements are:

KPI (sales volume, awareness, e.t.c.) – one KPI that we want to estimate
factors influencing KPIs
- own actions: media spending per channel, price, special promotions, sale, newsletters
- competitors activity
- holidays (summer break, XMAS)

For MMM, we need data over time, each factor with the same frequency.

The most popular data granularity used in MMM are:

2-year historical, weekly data – gives the best result
3-year historical, monthly data

It may look like the more detailed data, the better. But for some media is impossible to give, e.g., daily data. That’s why there has to be some kind of compromise. Usually, it is weekly data.

What questions does MMM answer?

What is the ROI of my marketing activities? How much value each $ spent bring?
How to optimize the marketing budget?
- across media – how much should we spend on each channel, what is the optimal level of spend
- over time – how to distribute ads during the year considering all the activities, holidays, and competitors’ activities.
What are the base and incremental sales? How much sales are generated by marketing activities, and what are the organic (not affected by marketing) sales?
Which ad performs better? How long does the effect last? Comparison not only media channels but also specific campaigns.
What is the competitors’ impact? How much and when do we lose sales because of competitors’ activity?

What is MMM?

Marketing Mix Modelling
statistical analysis to estimate the impact of various marketing activities
used for marketing mix optimization:
- calculation of ROI for each channel/campaign
- forecast/plan future activities

Statistical model allow us to calculate/estimate the impact of marketing activities based purely on number, not human judgement.

MMM works on time series data. That means that in order to create a model we need past data (e.g. last 2 years weekly data about sales and marketing spendings). That also means that having a model we can forecast/estimate future level of KPIs based on planned activities.

Not everything can be quantified but for sure we can quantify spendings on adds.

How does the statistical model work, and what is the benefit?

1. historical data

weekly data with ads spent and KPI (sales)

date	online ads	newsletter	sales
week1	15$	1	120
week2	15$	0	150
week3	10$	0	140
…	…	…	…
week53	15$	0	130

2. model – formula

sales = 100 + 2 * (online ads) + 10 * newsletter3. future plans

3. future plans

we know media plan for upcoming weeks

date	online ads	newsletter	sales
week54	15$	1	?
week55	15$	0	?
week56	10$	0	?

having a model we can forecast the result (sales)

date	online ads	newsletter	sales
week54	15$	1	140
week55	15$	0	130
week56	10$	0	120

e