Marketing Mix Modeling (MMM) is Anti-Fragile
Sometimes Bangalore traffic is not bad after all. Especially if one can finish a podcast while in transit 😅.
I was hearing a podcast featuring Nassim Taleb (Link in comments). In it he was discussing the idea of Anti-Fragile (also a great book by the same name)
I couldn’t but connect this concept to MMM because just few days ago one of our client asked – “why do you say you need variation in the data for MMM to work?”
Taleb said in the podcast “Some things benefit from shocks; they thrive and grow when exposed to volatility, randomness, disorder and stressors”
He calls these systems Anti-Fragile.
He also provided an interesting analogy.
“It is better to lift 1000 pounds once than to lift 1 pound 1000 times.”
Why?
Because systems often learn from meaningful shocks, not from endless repetition of tiny shocks.
📌How is MMM Anti-Fragile?
Most systems prefer stability but MMM does not. In fact, MMM needs variation in the data to work properly.
A simple example:
If a company spends $10k on a channel every single week/month, the dataset becomes extremely stable. From a business standpoint this may look disciplined.
But statistically it creates a problem. The model never sees what happens when spend changes.
Statistical models learn relationships from changes in variables. Remember your Regression 101? “1 unit change in Independent variable causes a x unit of change in dependent variable”
The fundamental concept underpinning this is ‘variance’.
If spend never changes, the model never observes how the KPI reacts to that change.
Another example:
Lets take another example – Imagine a company spent $10k one week, $8k the next, $25k thereafter, then pause for a week and then resumed.
This pattern of deployment has lot of variance and helps discern ‘what happened to sales (or any KPI) when the changes in spends happened’.
I view MMM has a collection of many tiny experiments (incrementality experiments).
Each spike, pause or budget shift creates information that helps the model identify incremental contribution.
MMM thrives in environments where marketing teams are actively optimizing budgets rather than keeping them static.
And no you don’t have to have million dollar spends on each channel. This ‘made up rule’ was put by some vendors to give them an indication whether you the client can afford to foot a bill of $100k-150k for one MMM model.
Gone are those days. With more MMM vendors entering and more transparency on what effort it takes for MMM, the cost is no longer inflated. At least not with us.
But coming back to the topic, Variation is not a weakness for MMM. Rather it needs it. The more variation in the data, there is more to discern.
P.S Image of Kintsugi. Kintsugi is a traditional Japanese art form where broken pottery is repaired using gold.
Instead of hiding the cracks, the repair highlights them. Similarly requirement of variation in MMM makes it standout in a good way.
