What is Marketing Mix Modeling?
Sep 1, 2023
Introduction
Marketing Mix Modeling (MMM) is a type of statistical analysis that provides valuable insights into the effects of marketing activities on sales performance. MMM gives marketers a comprehensive view of which channels are driving business growth by applying advanced statistics and machine learning. When combined with experimentation, it provides a strong foundation to understand the incremental value of marketing investments.
In this blog post, we'll explore the ins and outs of MMM, from what it is, how it works, what data is required, and how it can be used to optimize your marketing investments. By the end of this blog post, you'll have a complete guide to MMM that'll help you go beyond the basics and make better decisions about your marketing strategy.
Why consider Marketing Mix Modeling?
As marketing leaders, you're likely trying to answer the following questions:
What's the return on marketing investment?
Where should we invest the next dollar?
How much of marketing's contribution is truly incremental?
How do we adjust investments in real time?
How much should we be investing in marketing?
How confident are we about hitting our business goals?
Many marketers rely on touch-based attribution to answer these questions. However, as we discussed in this article, this approach is highly limited and unreliable.
This is where Marketing Mix Modeling comes in. It's one of the only marketing measurement methodologies that are based on proven statistics. Through MMM, we can gain insight into which channels drive the highest ROI, and forecast future performance of campaigns and initiatives.
There are three reasons it's better than touch-based attribution to answer these questions:
Universal. MMM can tackle indirect marketing channels (e.g., paid social, video) and offline channels (e.g. TV) in addition to digital direct response (e.g. search engine advertising).
Future proof. Unlike touch-based attribution, MMM doesn't suffer from the impact of privacy changes.
Unbiased. MMM is free of bias towards any specific marketing channel, instead of touch-based attribution, which favors channels that generate digital touches.
How does Marketing Mix Modeling work?
At it's core, MMM is an advanced statistical analysis. In recent years, machine learning techniques have further improved their strength and accuracy.
MMM starts by interpreting one to two years of historical data to find a statistical correlation between each marketing channel and your chosen target metric (e.g. sales). Depending on whether you have data on other variables (e.g., price, competition, or macro-economic data), it can also find the correlation between those variables and your target metric.
It then splits your target metric into the baseline (sales without any marketing investments) and the estimated incremental impact from marketing. It does so at both the overall level and at the channel/strategy/campaign level depending on the granularity of the source data. Finally, it creates a statistical model that enables both historical exploration and predictive forecasting.
This model enables you to compare marketing's estimated incremental impact in one period to another period (e.g., month over month). It also enables you to compare various scenarios of marketing spending in future months.
There are four steps in building an MMM:
Choosing the right target metric
Collecting the right data to build the model
Choosing the right period for your input data
Building and interpreting the results of your model
1. Choosing the right target metric
The data you use for MMM will have a much bigger impact on the results you get than the model itself. The target metric is the most important choice of all. If you want to get the best results, make sure to focus on the target metric.
Your target metric
Should be directly influenced by marketing activities
Should follow marketing activities as soon as possible
Should be how you measure marketing's impact on the business
Businesses with longer sales cycles or slow velocity where there's a long lag time between marketing activity and the eventual sale/conversion, it's important to pick a target metric that's relatively higher up in the funnel. This recommendation ensures that you remove any potential noise from your model (e.g., the effect of the product onboarding or sales/distributor experience).
You can still track the conversion rate from this higher-funnel-metric to the eventual sale and report on that separately.
Furthermore, it's important to understand the underlying trends of the data you are using, as well as any seasonal or cyclical patterns, to ensure the model is accurately capturing the data.
For a consumer mobile business, a good target metric is the number of users who install your app. A bad target metric is the number of transactions made by a mobile user 60 days after their account is created.
2. Collecting the right data
Once you've picked a target metric, you need to start collecting all the marketing and other variables that may have an impact on that target metric. The data you use for MMM will have a much bigger impact on the results you get than the model itself.
Consider all your paid, owned, and earned marketing channels. E.g., advertising (online and offline, brand and performance), events, organic social, email, content marketing, sales outreach, partners.
Start with one input metric per channel, but consider merging or splitting them as you iterate. This decision should be based on the underlying marketing strategy. For example, it's possible you have both audience building (reach focused) and conversion (activation-focused) ads on Facebook. In this case, it's useful to split those into separate input metrics.
Collect two data points for each input metric - impressions and costs. We recommend using impressions as the core input metric. Costs are used for all the return on investment calculations. In some cases, e.g., where impressions data isn't unavailable, it's possible to use cost data in the place of impressions data.
Collect at least two years of data with daily or weekly aggregation. Monthly aggregations reduce the number of data points and will necessitate a longer period (four to five years).
In addition to your marketing input variables, consider including other variables that might impact your target metric. For e.g., a financial services firm might want to include interest rates and an ecommerce firm might want to include an index of prices. Think deeply about potential relationships between those variables and your target metric before including or excluding them.
For advanced businesses with multiple geographies, products, or brands, it's worth developing multiple models to get a more fine grained understanding of marketing investments and their impact on your target metric.
In conclusion, accurate, comprehensive, and up-to-date data is vital for effective Marketign Mix Modeling. In today's environment, with a modern data stack, it's possible to automate data pipelines to ensure near-real-time data collection.
3. Choosing the right time period
Choosing the right time period is another important detail to consider while building a marketing mix model.
Begin with all your data and think about narrowing it down. Is there a particular timeframe where the transformation in the target metric was likely due to elements unrelated to marketing?
Consider whether to train on data at daily or weekly granularity — if possible, try both and compare metrics of model fit. A weekly rollup means fewer points for training, but if mid-week variation isn’t a factor for your business, it may also help your model to converge more reliably.
4. Building, interpreting the results of, and improving your model
Once you've collected all the data, it's time to start building the model. Given the technical nature of this step, it's best left to a data scientist or a marketing analyst with prior experience with Marketing Mix Modeling. Having said that, with recent advances in this technology, it’s more accessible than ever before. So, feel free to try it out on your own using the steps below.
Even before the actual process of model building starts, you'll want to run a variety of checks on your source data to ensure the input data lends itself well to this exercise. For example, you want to check for multi-collinearity, which means two or more of your marketing channels are highly correlated. Multi-collinearity makes it harder to detect the differences between the channels that are correlated themselves and might require you to group those channels before modeling. Another example is conducting outlier detection and removal.
We're finally ready for the modeling. There are various data science techniques and open source libraries (Google's LightweightMMM, Meta's Robyn, Uber's Orbit) that can be good starting points. In a future blog post, we'll share more about the pros and cons of various open source libraries. Paramark has an end to end solution build on top of several open source libraries.
Assessing the performance of a marketing mix model is critical for gaining an accurate understanding of its effectiveness. We recommend using a variety of ways to interpret how well the model "fits". Fit is a way to understand how well the model predicts outcomes. Mean absolute percentage error (MAPE) is a good yardstick for this assessment. It represents the average of the absolute percentage errors of each entry in a dataset to calculate how accurate the forecasted quantities were in comparison with the actual quantities.
Another best practice, which we follow at Paramark, is to use a holdout or “our of sample” period where you test the model’s prediction against the actually observed data. This allows you to calculate a MAPE for that holdout period and test whether your model has high or low predictive power.
Statistics and data science professionals commonly look at R-squared as another way to understand fit. However, it’s our opinion that R-Squared is not as important for MMM as MAPE. It’s possible to have a good model regardless of how high or low the R-Squared value for that model is. This is because R-Squared is meant to explain the variance of a data set instead of the predictive ability of the model.
The process doesn't stop there. If your MAPE is relatively high, we recommend studying the results and the raw data to find ways to improve your model.
Review your target metric with a cross-functional team of marketing, sales, and finance teams to brainstorm other ways to explain this observation.
Look for patterns in your residuals, which are calculated by excluding your predicted outcome metric from the actual outcome metric.
Do they spike around the holidays?
Are they higher for days where you spend more / less on a given channel?
Come up with ideas to test
Adding extra variables that can explain the residuals might help
Merging / splitting media channels
Once you've built the model, the predictions from the model can also be used to plan scenarios and create forecasts. We recommend refreshing the model every week or month as you consume new data.
Finally, experimentation is a great way to secure additional data to fine tune your model. See more on that in the section below.
How is MMM related to experimentation?
It's important to understand that Marketing Mix Modeling generates correlational insights. In this context, correlation refers to the mutual relationship or connection between your marketing investments and your business outcomes.
MMM produces an estimate of incrementality. It's a hypothesis of a potential causal relationship, which refers to the extent to which your marketing investments cause your business outcomes.
Experimentation, on the other hand, is truly causal. We recommend combining experimentation with MMM to get a more robust and precise understanding of incrementality.
Read more about marketing experimentation here.
How do you get started with MMM?
As we've outlined here, Marketing Mix Modeling requires a specialized approach and methodology.
While it’s entirely possible to take a Do-It-Yourself (DIY) approach, it requires significant investment and commitment to build and maintain. If you go down that path, consider the open source libraries mentioned in this article, and hire a data scientist to lead the execution along with your marketing analytics and marketing leaders.
Marketing Mix Modeling is a core pillar of Paramark's solution. If you'd like to explore how you can benefit from it, please reach out.
