Marketing Mix Modelling
Running your first Marketing Mix Model
Actionable results build trust, inform strategies, and unlock growth potential.
Actionable results build trust, inform strategies, and unlock growth potential.
Marketing mix modelling (MMM) is a statistical technique that estimates your marketing channels’ effectiveness. Analysing all of your marketing data produces a model which shows the ROI and point of diminishing returns of your media spend.
This knowledge, in turn, enables you to make better decisions about where and how to spend your next dollar, or to reduce spend to maximise ROI or grow your brand.
MMMs can also identify the impact of other factors, such as promotions, supply chain interruptions and competitor activity, on your marketing. This knowledge helps you better plan and forecast campaigns, understanding your strategic context.
This blog is a starter’s guide to marketing mix modelling. It discusses data gathering, modelling, outputs and expectations, and answers the critical question: what next?
MMMs analyse your data to provide an in-depth understanding of your marketing performance. They follow a simple process:
The first step is to ingest all your relevant internal marketing data, including spend and conversions, and external data sources like economic activity, supply chain disruptions, and competitor intelligence.
Acquiring this data, especially historical data, can be a significant challenge for marketing teams without engineering support. We recommend building automated connectors to pay-per-click (PPC) platforms and internal source-of-truth datasets. They’re worthwhile investments, but they require co-ordination across marketing, engineering and data teams.
Understanding what exogenous data may be valuable to your model necessarily involves sitting with stakeholders to understand the factors that impact conversions across the business and the best way of representing them as a time series.
For example, many people understand that you can add weather data to a model, but the key is to input the most relevant data. For instance, ‘temperature by day’ may not be as valuable to your analysis as ‘days over 35degC’ or ‘weeks with more than five days of rain’.
Beyond that, you’ll want to consider how to group advertising channels. What channels will be the most important for modelling ROI? An MMM is most helpful in identifying where extra spend can be profitably employed, so we recommend thinking of the different audiences inside of each channel.
For example, you should distinguish between Google Brand and Google Generic campaigns. These campaigns have very different characteristics in terms of how much spend can be deployed, and where the point of diminishing returns is. Establishing such groups means separating the spend for your brand and generic campaigns into separate datasets.
Once acquired, your data needs to be standardised (with everything in the same format) and normalised (with everything at the same level of granularity, e.g. daily, weekly or monthly).
You can export standardised data from PPC ad platforms. However, gathering data from offline campaigns such as TV and radio can be challenging, as they deliver proprietary reports that must be re-formatted into daily values.
It is important to consider currency, tax and the impact of ‘bonus’ media when adding together spend from all different sources to ensure all data comparisons will be ‘apples to apples’.
Collecting, formatting and validating the data once is not enough; MMMs require regular updates. Automated data feeds allow your model to provide ongoing insights to assist decision-making rather than becoming an out-of-date review of the last 12 months that arrives 3 months too late.
Building an automated export from internal data sources and PPC channels is critical, as is establishing a regular cadence of spot reporting from your media agency for offline campaigns.
With all your data gathered and regularly updated, you should keep in mind two potential complications:
Even if you have a well ordered suite of data, you must double-check that all your inputs add useful information to your model. Throwing in a massive number of sub-channels is a classic mistake that will present a surprisingly low-error model that utterly fails to predict the future.
We call this ‘overfitting’, where we give the model too many inputs for the data we are modelling. For example, if we have 730 days of data, we should avoid inputting more than 20 sub-channels.
The risk of overfitting doesn’t simply mean leaving data out; instead, you should also combine channels that are too similar or too small to be helpful.
You should be wary of using different channels that are too highly correlated (for example, you may always buy CH7/CH9/CH10 in the same campaign). MMMs are statistics, not magic, so distinguishing individual effects won’t be possible if two channels are always simultaneously turned on and off. It’s best to add them together and treat them as one channel.
The only way to unpick these individual effects is with experimentation.
Once you have your data and your initial outputs, you can use these to create a marketing mix model that shows the optimum channel mix, budget allocation, ad-stock decay and more.
Ultimately, an MMM predicts a channel’s ROI by estimating its diminishing return curve and ad-stock decay to predict your historical conversions.
There are two main approaches to building these models: ‘Frequentist’ and ‘Bayesian’.
No matter your approach, the model’s assumptions and configurations should be transparent and open to discussion. New inputs and changes in the market will reflect in your model, so monitoring how the model performs over time is crucial.
Bimodal uses a transparent Frequentist approach as we believe this removes sources of bias and results in a credible model in the hands of users more quickly.
Fundamentally, there will always be a few possible models that credibly explain your results. With carefully chosen and validated inputs, there will only be a small number of results that make sense.
It can be handy to examine the features shared between different credible models and to identify where they diverge. These differences are often ideal questions (e.g. is Facebook Prospecting over-invested or knocking it out of the park?) to address with an experiment.
All models are wrong, but some are useful.
— Box
The bottom line is that the model needs to be accepted by the people who will use it: media buyers. Your first model will be the least trusted, as it will be unproven and poorly understood.
Rather than starting a protracted argument about the nature of marketing itself, we recommend choosing a model as a starting point and then committing to experimenting and improving it to achieve the best possible accuracy.
This approach allows your team to engage in a sophisticated way with your model’s outputs, which have some level of uncertainty. Instead of claiming that your model is indisputable, presenting it as a series of opportunities bearing experimentation will be much easier and more effective.
One stakeholder’s pet channel may be a colossal waste of money, but we can test this instead of simply asserting it.
Ask yourself: would your team believe your model is perfect, even if you could make it that way the first time? Rolling out an iterative, test-and-learn approach is much more credible and will, over time, build trust and confidence in its results.
In the simplest terms, an MMM shows you where to increase or reduce spend across your channels. An MMM is ideal for identifying opportunities to drive more revenue and increase ROI.
There are three key factors to consider here:
An MMM’s critical output is each channel’s diminishing return curve (or spend response curve). Across any marketing effort, higher levels of spend eventually experience diminishing returns.
Intuitively, every channel can bear a maximum spend before diminishing returns set in. Consider a TV campaign that made up 50% of all ads in every ad break for four weeks. Doubling this campaign’s spend, so it comprises every ad in every ad break, would not be likely to double its effectiveness.
So, the diminishing return/spend response curve estimates the ROI of any given level of investment in a specific channel in your model.
Adstock decay estimates how long a dollar spent on a channel influences the market.
Intuitively, some ads achieve greater memorability or encourage a longer-term behaviour change than others. This variability means spending in these channels may have longer-term impacts than expected.
Consider the difference between a highly produced TV ad and a simple banner ad with a discount message. The banner ad with a discount reminder may cause a customer to click through immediately, but it’s unlikely they will still think about it two weeks later.
On the other hand, TV ads may stick in the mind of a customer over several days. You can probably remember TV ads from childhood but likely struggle to remember the last display ad you saw.
This information is critical in making good decisions regarding marketing spend. It is essential to consider a channel’s adstock decay when trying to understand the timeframe to expect results from a campaign and how long to run experiments.
A budget optimisation report summarises the optimal adjustments to make to your media mix based upon the MMM. It allows you to see how specific adjustments impact each channel. This knowledge helps better plan specific channels, as well as forecast next quarter budgets and media strategies.
A budget optimisation report should help you deliver actionable media schedule-level recommendations on allocating your budget —increases and decreases— across every individual channel. When starting out, it’s a good idea to choose a single channel that represents a big opportunity for improvement as your first adjustment.
Once you have an idea for what channel to adjust, you should test it. Designing a geo-holdout experiment (similar to an A/B test but conducted by excluding certain geographic areas from your campaign) can help you validate your model’s accuracy and the exact impact of your changes.
With these optimisations in place, your first Media Mix Model is complete — congratulations!
Once your MMM is up and running, the next priority is to test, test and test again. Constant refinement through regular refreshing of data and incrementality testing allows you to improve your model to ensure it’s always as relevant, accurate and as reliable as possible.
Keeping your model updated once it’s up and running is critical, as long as you have a regular (even automated) flow of all of your marketing data, the MMM should be updated to include it. In practice this means an update every week or month.
Thus, your model will continually evolve, just as your business, marketing budget, communications channels and target audience will evolve. It is particularly useful to understand how different channels are adapting over time to your changing environment.
The most potent tool to improve an MMM is an incrementality test. Good experiments help calibrate your MMM, ensuring predictions are in line with a ground truth. They can validate identified opportunities, demonstrate the value of provided insights and prove ROI in new channels.
Several methods under this banner include the already-mentioned geo holdout tests, difference in difference tests, or matched market tests. Each of these attempt to understand the contribution of a single change to your overall marketing ROI.
These experiments can be time-consuming and resource-intensive, but the results are generally worthwhile. They provide direct measures of different tactics’ effects and can be adapted to test a wide range of ideas. You can read about running your first incrementality test here.
Our platform offers marketers a smarter way to make decisions. Ready to uplift your marketing effectiveness?
.jpg)
