From U.S. Army to MMM expert at McDonald’s – Alex Gaski
Alexandra Gaski is a Marketing Data Science Analytics Professional. She has nine years experience within data science in organisations like FTB Chicago, Boeing, the U.S. Army and now McDonalds.
How did you start using Marketing Mixed Modeling (MMM)?
I actually started out my career in analytics while I was in the U.S. Army working in Military Intelligence. Working with data in the military is very different but looking for patterns and using disparate data sources is still similar no matter where you work.
This made me quite agile in my way of thinking and today it’s easy for me to adapt to different kinds of data really well.
At the end of the day, all data is based on human behavior, right? So, for me, it wasn’t a huge leap to go from data in the military to working with FCB Chicago with clients like Boeing, Jack Daniels, and even PACAAR Parts (they are a B2B supplier of big rig truck parts) to name a few.
How are you using MMM today?
Today I’m leading the marketing mix project at McDonald’s US. When a company like McDonald’s has a significant amount of ad spend and advertising dollars going out into the world it is really important for us to know what we actually get back on our investment.
When I came to McDonald’s in 2017 they were still just dipping their toes into MMM and had not gotten the level of granularity they needed to really make help make better media decisions.
So for the last two and a half years, I have focused on rebuilding the marketing mix model from the ground up. We had to completely revamp the data inputs and pull in new data in order to get the results that we wanted and make better and more data-informed decisions.
To do this we partner with analytic firms that specialize in building marketing mix models.
We’re getting very detailed in our findings as we run the model quarterly now. Before we used to run the model annually and it has been quite the challenge to do the analysis quarterly! We really had to work and push hard to automate as many things as possible to get the data quarterly.
Since we are now running it consistently to measure performance as well as a planning tool, we can plan ahead of what we want to do as a marketing organization in 2021. It’s super helpful knowing what’s going to get us the greatest return for our marketing investment for the next fiscal period.
What is your best advice within data cleansing and data collection?
Data is really the heart and soul of your model. As one of my favorite professors used to say “Garbage in –garbage out” meaning, if you have the best algorithms or AI in the world but you have bad data, you still end up with a bad model.
I recommend a two- or three-year plan. Taking a look at the most granular variables that you want to get into the model and then map out what you can do in the next three years.
In some cases, it can take up to two years to get most of the variables in the model and then another year to increase the level of granularity of those variables.
I also recommend collecting the data at least quarterly, even if you only run the model every six months or once a year, getting in the habit of collecting that data from the beginning is going to be better for you in the long run.
In MMM you are creating a simulated world while trying to understand the market’s response to your advertising. So accuracy is important. It lets you know that you’re accurately portraying the market and the market’s response to your advertising.
Without something to control for all the variables, there is too much going on for the human brain to untangle what is really driving sales and foot traffic.
Are you also working with Multi-Touch Attribution (MTA)?
Yes, we use a limited form of MTA on a weekly basis to optimize a campaign while it’s still running. A lot of strategies out there mix MMM and MTA, but we don’t.
I believe that it’s important for our attribution to be predictive and validated with our MMM. Therefore, we make sure that our attribution results are correlated with our MMM results, but we keep the tools separate so that we can continue to use both to validate the other, creating a system of checks and balances with the final judge and jury of a program being the MMM as that model accounts for the most amount of variables and is proven to be predictive of our total sales.
I believe this helps us to make good and accurate decisions while a campaign is still inflight since we know that the attribution metrics, we are optimizing on are predictive of the final performance in our MMM.
What are the challenges with Multi-Touch Attribution?
It’s very difficult to get a single view of a person rather than a cookie view or a device ID and even more difficult to get a television device or a computer IP address.
So it’s not that MTA is problematic in itself. Multi-touch attribution is challenging in viewing a single customer and privacy regulations are making that even more challenging.
For Safari users, they don’t allow cookies to be tracked, without explicit opt-in. Facebook doesn’t allow cookies. Google will not send out any information from YouTube or Google. So there’s more than just cookie deletion. Cookies are becoming obsolete.
The upside with online sales is that you can look at the entire customer journey and do continuous tactical changes to optimize the journey based on where we lost the customer’s interest.
However, for McDonald’s most of our sales still happen at our physical stores and in cash, making attribution of our ads very difficult.
Also, for big-ticket purchases like buying a car, you might have a long customer journey with many different touchpoints but for us, the customer journeys may be relatively short and based on existing behaviors. Eating out at a Quick Service Restaurant can be a frequent or habitual occasion for some people so our real question is if our ads are bringing in additional visits or would those people have come in anyway?
To answer that question, we have had to test out different methodologies to account for these issues.
Validation is really key for us. We want to be able to proof that the models we are building are an accurate depiction of the market place. For MMM we do this with test and withholding methodologies to make sure our models are predictive of sales and count of transactions and with attribution, we routinely validate by running an analysis of the attribution results against our MMM results to ensure the attribution results continue to be predictive our MMM.
At the end day, we want to help inform our business partners with the best data to make decisions. When people know better, they do better, I sincerely believe that.