Some weeks back a contact on LinkedIn, after reading my article on Google LightweightMMM, asked me for some questions on Marketing Mix Modelling and Attribution.
I was really busy in the last month and I was not able to reply properly but here we are.
I split the questions on MMM in two different posts.
I’ve encountered some difficulty in grasping the summary of the MMM model. Could you provide additional insights or recommend resources for a clearer understanding?
Each year marketing departments invest significant amounts of money in marketing activities. We could talk for hours about what all the goals of these activities are, but we can all agree about one key goal: increasing sales.
How? Marketing has the responsibility to build familiarity and recognisability, one path to follow is advertisements.
Oversimplifying, people hit multiple times by a message, consistent colours, font will remember and recognise a product more easily and will be more likely to buy it compared with all the other similar options.
How to evaluate the marketing effects?
One way is creating a mathematical model that approximates the buying process.
This, with an intrinsic level of uncertainty, can be done with Marketing Mix Models.
Despite the hype, despite big words like AI, Machine Learning etcetera, a Marketing Mix Model is just a slightly more complex version of a linear equation.
When we start studying physics during high school we model complex phenomena starting from basic equations, then step by step we add layers of complexity.
Take for example Hooke’s law: “an empirical law which states that the force (F) needed to extend or compress a spring by some distance (x) scales linearly with respect to that distance”.
It’s obvious that in reality things go slightly differently but it’s a reasonable approximation.
Marketing Mix Models try to do the same.
On one hand we have the inputs that are a representation of the marketing activities that reasonably influence the buyer behaviour (Price/Distribution/Impressions from YouTube Ads/Impressions from TV Ads) and on the other hand we have the value sales or number of products sold.
This is the fundamental goal of a Marketing Mix Model.
Then there is a second one, less easy to measure which is the channel’s diminishing returns.
I talked about them in detail here:
A promotional channel like Facebook or YouTube has a certain number of viewers, this means that when the advertising budget is low I can expect an approximately linear relationship between the increase in sales and the increase in the number of people who were influenced by the ads, but after a point we will start reaching a saturation level, which is understandable, because the people on the platform are not infinite.
This change is called diminishing returns and we can model it. It’s a phenomenon that happens also in other fields and if you want to know more you can check the diminishing return Wikipedia page.
If we are able to approximate the diminishing return function that influences each media channel then we are able to optimise the media spend. This is mainly an operational research job and differential calculus. As far as I know, only the Robyn library developed by Meta leverages Nevergrad library, an adaptive optimisation algorithm for minimising functions.
What’s the output of a Marketing Mix Model?
Unit or Value sales decomposed into different factors.
Imagine yesterday your e-commerce daily revenues were £ 10,000. A Marketing Mix Model will try to decompose the factors that, reasonably, determined the £10,000 sales.
- £500 from YouTube Promotion
- £500 from Facebook Promotion
- £250 from Google Paid Search
- £750 from Seasonal Factors (weather/season/other)
- £2,000 from Price
- £1,000 from Organic Search
- £5,000 from the “Base”
The base represents a chunk of sales that would happen without any incremental activity.
How to decompose sales in a Marketing Mix Model?
We need historical data. At least 1-2 years and based on that it is possible to build a regression that relates historical sales data with all the activities conducted in the past: from media activities to price promotions.
Weekly or daily data in my MMM?
It depends but in general I would prefer weekly data.
On ecommerce data I have not a straight answer, but I still prefer weekly data. They can help smooth out the random fluctuations and build more robust models.
Some people prefer daily data because it will lead to better R2, I am not a huge fan of this, because in order to improve the R2 you will take some spurious correlation that doesn’t have any causal relation.
This is a hard problem in MMM that I illustrated in an earlier post, based on a Google Research:
Monthly data on the other hand is not enough to catch the media impact.
If you liked the article and you want to get in touch with me please fill the form below or add me on LinkedIn.
