Demystifying Market Mix Modelling
A plain-English guide to how modern Market Mix Modelling actually works - and what separates a useful MMM from an expensive distraction.
Market Mix Modelling (MMM) has been around for fifty years, but most marketers still describe it as a black box. That is a problem, because the technique is having a real renaissance - driven by the collapse of third-party cookies, the limits of attribution, and the simple fact that boards want one number for how marketing pays back.
This piece strips MMM back to first principles, then describes what "modern" actually means.
What an MMM really is
An MMM is a regression. You take weekly (sometimes daily) sales or another KPI, and you decompose it into the things that drove it: media by channel, price, promotion, distribution, weather, competitor activity, seasonality, and a base level. The output is a set of coefficients you can translate into ROI, response curves, and saturation points.
The maths is not the hard bit. The hard bit is the data engineering and the judgement about which variables to include, how to transform them (adstock, diminishing returns), and how to handle the inevitable multicollinearity between channels that all spike together at Christmas.
What "modern" MMM adds
Three things distinguish a modern build from the slow, annual exercise of the 2010s:
- Bayesian priors. Instead of letting the model freely choose any coefficient, you constrain it with what you already know - geo experiments, platform lift studies, historical norms. This dramatically reduces the "TV ROI of 47" results that used to embarrass the practice.
- Faster refresh. Code-first pipelines (Meridian, Robyn, internal frameworks) let you re-run monthly rather than annually. That moves MMM from a board-deck artefact to a planning tool.
- Tighter scope. Good modern MMM resists the temptation to model everything. Pick the decisions the model needs to inform, scope the channels and KPIs to those decisions, and accept that the rest is noise.
How to know if yours is working
Three quick tests:
- Out-of-sample fit. Hold back the last 8–12 weeks. Does the model predict them within a sensible band?
- Stability. Run it with last month's data and this month's. Do the channel ROIs swing by 40%? If so, the model is fitting to noise.
- Decision change. Did the last MMM cycle actually move spend? If the answer is "we noted it but did not act", you have an analytics deliverable, not a decision tool.
MMM is not a magic trick. Done well it is the most honest answer to "what is our marketing actually worth", and the cheapest way to stop wasting the next budget cycle.