The trouble with writing about mix modelling is that one can descend into cliché-land very quickly, end up stating the obvious over and over again. So let me very quickly put that obvious stuff behind me –this article is about how you can know the impact of your marketing spends, and then spend where it matters.
It starts with a Target
Like most things, it starts with a target. What is it you want to model? Sales? New customer acquisitions? Brand Recall? In most cases it has to be one of these so take your pick, and frame your target statement. Something like “model Sales!”, or if you wish to be more wordy – “how do I optimize my marketing mix to maximize Sales”. The important thing – whether you use 2 words or 10, you’ve decided your target.
Sure about that?
So sure, you’re modelling Sales. All India? Region wise? For a product category or across all Products? Think through these, and quickly now, as you could easily end up with trivial or absurd results without this input.
You might, for instance end up refining your target to “Model Sales for the West Region”.
Get your data together
If you’re going to do mix modelling, be aware that this will be 70% of the time spent. No kidding. Getting your data across digital, print, outdoor, promotions, radio etc can really take time. You’ll typically need a couple of years of week level data. No wait – the more granular the better, and further back you go the better.
Don’t sweat it
If getting all the data is a problem, don’t sweat it beyond a point. It is far better to get going than waiting for the day to come when your data is perfect (never), and everything you lack gets accounted for in the model as a base “unexplained” Sales, which still gives you plenty of insight. Really, just knowing how little you know can also be useful right?
What’s the data you’ll probably use:
• Digital spends, by source – Google/ Facebook et al, the easiest data to get your hands on
• Spends on Email/ SMS sends – Netcore and other similar sources
• Spends on Justdial and other similar channels
• Print spends – full page/ half page by publication (or publication category – newspaper/ magazine)
• Radio/ TV – spends or GRPs
• Outdoor spends
• … you get the drift
Should the data be super granular? Or aggregated up to spend category? You need to experiment with this, we’ve found the sweet spot to be somewhere in between.
Now add to that data:
Think beyond spends, add data to indicate seasonality, promotion weeks, pricing changes, new product launches, store openings, events and sponsorships etc. The more you add the better.
The modeling process may seem like a bit of a black box, but it’s really very very nuanced. A crucial aspect to mix modelling is setting the carryovers right. You place an Ad today, the impact carries over beyond the day of the ad, but it also plateaus over time! And this is different for an Email blast, a radio spot and a Print Ad. Very logical, but gives rise to quite a few headaches for the modeler.
The “model” itself usually refers to a regression model, a multivariate one. The modeling process would typically cycle through hundreds of models, improving the model quality (or, reducing the model error), till there’s no improvement possible. Genetic algorithms are often used to “evolve” the models and eventually select the best fit. In plain English – the algorithm develops and selects it’s best output for you to consider.
THE OUTPUTS YOU GET
A key output is the forecast that shows you that given the marketing plan, you can now forecast Sales. You’ll typically get a forecast on past data/ holdout sample and get to see how accurate the model is. How accurate you want it to be however is a business call. We’ve heard “over 90%”, we’ve also heard “anything is better than what we’ve been doing so far!”.
The Brand Painting
This is one killer output: What is that Base sales that would have happened without the marketing initiatives, and how do different marketing actions lead to it adding up to the final actual number. Spin off benefits are things like the relative spend vs. impact of each marketing initiative.
Forecasts and Optimizers
The final touch is the ability to use the output to optimize. The goal seek can be phrased as 1) given a Sales target how much do I need to spend and where should I spend it or 2) I have this much money to spend, how should I spend it to maximize sales.
End Note: One would imagine mix modeling is something all marketers routinely do, but we’ve found that it is not a common practice in most industries. Seeing the outputs of a mix model has been a “big reveal” to a number of clients, and they suddenly get manic gleams in their eyes, goosebumps, and develop a sudden fondness for saying “mama mia!” over and over again. And it’s real, not something that looks only cool on demo data sets but falls flat when Indian brands throw their data at it – warts and all. Here’s to more precision in your marketing!