Open source Marketing Mix Modeling (MMM) tools are great for democratizing MMM. But they will never provide you with an accurate MMM model off the shelf.
Why?
Because of Flintstones curse (see Ridhima’s post – link in resources).
But that does not mean we can’t appreciate some innovative methods in these open source MMM tools.
Robyn may not be perfect but as somebody in a post said “it has flashes of brilliance”.
One of the those flashes of brilliance is – Decomp RSSD.
Robyn markets it as business fit metric. While others call it controversial (check out Mike Taylor’s tweet in resources).
๐ So it Decomp RSSD controversial? Does it optimize for politics?
In my opinion Decomp RSSD has some parallels with Bayesian MMM priors. But it induces far less politics than Bayesian priors do (more on this in future posts).
Staying on the topic, Decomp RSSD is like an alarm that beeps loudly when your model veers off too much from the historical marketing spends. Now historical marketing spends does carry a lot of institutional knowledge.
No brand would be crazy enough to pump thousands or millions of dollars on a marketing channel if it was losing them money big time. Yes this is an assumption, but from our experience, we feel by and large this holds true.
So as a starting point, it is not crazy to specify a model that more or less mirrors what worked in the past.
๐ So how is this different from Bayesian prior setting in MMM?
Priors are pre model, Decomp is post model
Well for starters, Bayesian prior setting is a blind guess. The priors are set not by looking at the data but by guesswork on what the attribution coefficient should be or in some cases, by virtue of what coefficients one historically got through previous model building efforts.
Whereas in Decomp, the metric is being optimized by looking at the data.
๐ The argument of Decomp digressing from the following two purposes of MMM:
โช To gauge if you are overspending or underspending on a marketing medium and by how much.
โช Accurate attribution to variables that caused the change in sales (or any KPI)
Yes, there is merit in the above argument.
Much like Bayesian paradigm confines us to what is possible, Decomp confines us to historical spends as a yardstick.
๐ So is Decomp RSSD not useful.
Not exactly. Decomp provides guardrails against extreme results that diverge from current business understanding. But in doing so it also has the risk of confining one to historical spends as yardstick.
At Aryma Labs we have our own proven methods that is rooted in belief of “letting the data talk” without constraining it too much.
However we do use Decomp RSSD in a novel way.
Will explain how exactly to use Decomp in a useful way in next post. Stay tuned !!
Resources:
Marketing Mix Modeling’s Flintstones Curse :
https://www.linkedin.com/posts/ridhima-kumar7_marketingmixmodeling-marketingeffectiveness-activity-7175842925279412225-3Nbh?utm_source=share&utm_medium=member_desktop
Mike Taylor’s tweet on Decomp RSSD
https://x.com/hammer_mt/status/1460583843698810884?s=20
