Dump your Marketing Mix Modeling today!
Listen to the Audio Version here
MMM has a long history in marketing. Pioneers like Unilever started using it in 1919. Yes, that was over a hundred years ago. Surprisingly, the core principle of the methods has not changed. The AI age has now brought an answer: Causal AI Deep Learning models.
Recently, a European brand asked me to review their current approach to marketing mix modeling. So I joined a meeting with an established media agency and the head of analytics explained their approach to me.
From a statistical perspective, it was all well thought out and a carefully selected approach. A statistics professor would probably be pleased with this work. It produced significant and somehow plausible results. For me, it was the first time I learned first hand how other agencies do MMM. Everything was great from a textbook perspective.
Except that I would do it completely differently. Only that I think it was logically flawed. Only that it could only produce true results by accident.
"Your MMM may be state of the art, but still fundamentally flawed"
Here are a few flaws I noticed:
Casual (instead of causal) thinking: The model tried to predict daily sales by using yesterday's sales and the days before. This is a typical technique developed for time series prediction (called ARIMA). It leads to a very high explanation (R2) of the results. Why is this? Because yesterday's sales correlate much better with today's sales than with last month's sales. The problem is that there is no causal relationship between yesterday's sales and today's sales. They correlate because yesterday's sales and today's sales have similar causal drivers. By using this ARIMA approach, you increase the chances that the model will fail to identify the true causal drivers of sales.
Contextual information: The brand sells a product that is bought as a gift, so demand increases before Easter and Christmas. The model has simply marked the days of Easter and Christmas. Unfortunately, customers buy before these events - when the model would have no idea about the holiday season. So it is a good practice to mimic in the data what consumers have in mind, like "how many days is the gifting event". When we look at it later, it turns out to be THE main driver of demand and sales. Understanding the domain and context can suddenly skyrocket the explanatory power of an MMM model.
Long-term branding effects: Same-day sales are only part of the impact of an ad. People may buy tomorrow, or some may buy a month later. But the ad increases the likelihood that a customer will remember the brand when he is in need or in urge for the product category. That is why whenever possible MMM should model not only sales but brand consideration. If not, you're missing about half of your impact and underestimating your ROI by a factor of two. (Write it down: "We - are - missing - half - the - impact!")
Write it down: "We - are - missing - half - the - impact!"
How is this possible? I can only guess: We tend to work in silos. Data scientists vs. marketers. In essence, we need both - ideally in the same person - to build meaningful models. Then there is the bystander effect I talked about in a previous article (see 10x Leadership). It takes courage to stand up and say, "What we've been doing for decades is flawed. Let's fix it."
Do you know this legend? A vain emperor is tricked by two swindlers who promise him an invisible suit that only the wise can see. The emperor parades through the city in his "new clothes," but everyone is too afraid to admit they can't see anything - until finally a child cries, "But he's not wearing anything!”
It needs courage to be that kid. But it can be very rewarding.
What about an AI-powered MMM tool?
Guess what, there are Martech vendors offering MMM tools based on AI, they say. Sure, this sounds like an improvement, and it probably is.
But there are two BUTs:
1. AI is not causal AI. It is not designed to find the cause of results. It is designed to find a formula that makes the best prediction on learning data. This has major drawbacks: Model Drift (model performance fades in production), Discrimination (unfair claiming of causes), Biased attribution of effects. (for more on this, I recommend reading my latest book "Think Causal Not Casual" available at Amazon & co.)
2. Modeling is only half the game. The examples above (Casual not Causal, Context, Brand) show that it is not just about the algorithm, it is about using the right meaningful data and using it in the right way.
"If you encounter AI-based MMM, check it twice"
Fixing MMM with Causal AI and Deep Learning
One of the reasons why most MMMs still use traditional statistics is this: For MMM, we have to work with very few data points. Often we work with weekly data, which gives you about 100 data points for two years. When the MMM focuses on the last year with daily data points, that's still not a lot of data.
Traditional statistical modeling finds the parameters of a known formula. The strength of AI is to find a complex unknown formula in the data. But that requires more data points.
It DID require more data. Over time, modeling methods have been developed that use a clever trick to work with very few data points. Deep learning is an example. We have models that sometimes have more input variables than data points. The trick is to use multiple output variables (for example, today's sales, but also the next few days) to feed the algorithm with more information. This is important because one thing is clear: you can't pull information out of thin air.
"AI doesn't need a lot of data anymore, thanks to recent innovations."
With AI, you can now learn and surprise yourself. For example, in this pharmaceutical project, we found that providing product samples works, but that the effect diminishes to the point where it actually prevents sales instead of increasing them.
My kids just told me yesterday that during the European Soccer Championship, the advertiser "Cool Blue" annoyed them so much with preroll ads on Youtube that they literally hate this brand. These anomalies happen, and with conventional MMM we are blind to them.
Same with interacting variables. MMM assumes that variables are independent. But you can bet that in the gifting season, the in-season ads are more impactful in driving sales than in the off-season. With conventional modeling, you simply measure the mean and miss the seasonal leverage.
In a recent case, Causal AI-based MMM proved that the campaign had 3x the impact that conventional MMM had found. There are many reasons for this. One key point is that Causal AI took into account indirect medium-term effects as well as long-term branding effects.
3 reasons to scrap your MMM
Here are three good reasons to pilot an MMM with Causal AI & Deep Learning:
GROWTH: Drive more short-term and long-term (brand) sales with the same campaign budget, because you build on more realistic findings.
WHY: You learn more about what drives results, e.g. creative power, seasonality, audience match, depending on the data you are able to feed in.
RESPECT: Become a respected voice in the organization. Leaders who dare to be more innovative, and then can show results, achieve what most laggards never will: respect and a greater voice.
Test MMM with Causal AI & Deep Learning and compare it side-by-side with what you have. This innovative test, learn, repeat mindset puts you on an exponential path.
This is how you 10x your impact.
p.s. If you don't know where to start, let me know. At SUPRA we are conducting free assessment calls this fall. Within one hour, we can reliably confirm the feasibility and potential of your MMM upgrade.