Marketing Mix Modeling • 6 min reading time

How to dramatically speed up your Marketing Mix Modeling

Ted Lorenzen - published on March 8, 2024

I learned Marketing Mix Modeling - what MMM is, how to estimate one, and what to do with it - from very experienced practitioners a long time ago. So long ago that we had a server room and an IT guy that would occasionally walk over and peer at the blinking lights to see what was going on in there. I had a desk phone with voicemail I was expected to check and a stack of business cards and . . .  

In those halcyon days of yore, we’d block out an MMM project as a 16-week endeavour and assumed it would take us 12 once the client got us all the data files. (Unless those files were all pdf we had to manually transcribe, in which case it went more slowly.) 

Fast forward to this month, I saw a MMM consultant colleague get a model project completed in a week; a week where he had quite a few other things going on. I think we all agree that waiting a week for MMM results is much better than waiting 16! But in case he’s not available and -- honestly -- even if he is, there are two keys to speedy MMMs. Both are process related and achievable by any team managing marketing spends. Those keys are continuous data ingestion and proper MMM tooling. 

 Continuous Data Ingestion 

The largest bar in the Gantt chart in an old style MMM project was data prep. In a 12-week project we would block the first 4 weeks as data prep and would consistently need the first 6. Some of this was due to using data collection processes built around the needs of accounts receivable and not MMM. I fondly remember waiting an extra month (a full four months after close of the quarter!) to get spend for the last modeled quarter from an agency. I asked why it took so long and her response was: "We don't bill until 90 days post quarter and the bill is for the total spend, not split by brand. We don't have resources to disaggregate the expenses until after we bill all our clients.”  

If you want a fast turnaround time on MMM, you need all the partners involved to hold hands and agree to frequent rapid turnaround reporting at or below the granularity needed for modeling. Waiting to build that data until MMM is set to start is wasting a lot of time! 

Having a weekly or monthly data update process makes starting an MMM project much faster and empowers a marketing manager to check spend pacing vs budget or year ago totals. Teams can understand what media has occurred and what is planned. Clients in old-style MMM engagements often used the model data as their marketing data warehouse, but in the world of the ‘modern data stack’ where firms spend huge percentages of their total budgets on social media, search engine marketing and display networks, it is simple to pull marketing activity regularly and view it in a business intelligence tool.  

We at ScanmarQED have prebuilt connectors, data normalization rules and dashboards to get a marketing team from zero to a marketing data business intelligence with just a few days of supported setup. Working from a prebuilt, always-on data warehouse takes 6 weeks out of your MMM timeline! 

MMM Tooling 

The second longest Gantt bar in an old style MMM project was 4 weeks long and was labelled “Model building”.  In my office (with the server room) we used custom software running in a custom compute environment for model estimation. It was awesome: handled huge data well, hardly ever broke even on numerically intractable problems, and could be scheduled to run all weekend long.  

Buuut, the kinds of questions that come up in model development are: “Should we use monthly dummies or Fourier series terms for seasonality?” and “Hmmm, this Facebook advertising switched creative and audience in June, does it look different if we split Facebook into two?” and “Programmatic and non-programmatic display are nearly collinear, what if we add them together?” 

With each question, the lists of models to test would get bigger and bigger and we poor consultants stayed up later and later, rearranging the input data and coming up with notes on which model iteration was for which question and WAITING for the models to run and then extracting results so we could visualize fit and compute key marketing contributions to know if an iteration was useable or not. That was how we worked through a research program of models. Ultimately, we would consider trade-offs between most predictive, truest to the data, acceptable to the client and bless some model as THE FINAL MODEL. The key point is that it was many steps from “I have a question” to “I know if that is a good model.” 

A modern MMM tool combines all those steps into a seamless workflow. It can test models very rapidly (10,000 a minute, even), with heuristics applied to choose the best models from the possible models. A modern tool also tightens the loop between model estimation and visualizing outputs. An analyst using a modern MMM tool can ask “What if I just . . .?” and see the results immediately. With a tool like this, e.g. our very own strataQED  MMM workbench, model building time can shrink to a week – even if an analyst occasionally gets up to have a coffee break.  

What will we do with all of that time back? 

So, on the old timeline, data prep was 4-6 weeks. Model building was 4 weeks. But remember I said we blocked off 12 weeks? The remaining 2-4 weeks were used to socialize the model with all the stakeholders (which led to more model iterations) and produce the final model deliverables. 

How can you cut that down? Maybe you can and maybe you can’t. And maybe you shouldn’t because that’s time spent with your stakeholders, learning what decisions they need support for and how best to build forecasts, what-if scenarios, and optimal budget allocations – which, of course, strataQED’s ecosystem handily supports. 

If you can cut your data prep and model building time down to 1 week, you can spend 7-9 weeks with your stakeholders designing and running a robust MMM planning process instead of beating on your MMM vendor to get models completed. That’s a win worth a software license or two in my book! 


Picture of Ted Lorenzen

Ted Lorenzen

Director of Data Science