Marketing Mix Model on the rise again (2)

6 min readJan 31, 2024

(Continued from Part 1)

Marketing Mix Modeling (MMM) has seen a resurgence in popularity due to privacy regulations and other changes. However, the context and environment are different now than they were then.

Modernized MMM

It’s been a while since MMM has seen a revival, and technology has evolved, as well as the level of familiarity with data experienced by people in various fields. Today, when we look at new and improved examples of MMM, we see the following characteristics that stand out

Incorporating machine learning and AI into MMM

Modern MMM leverages machine learning algorithms and AI to analyze data, build predictive models, and conduct simulations within the traditional MMM framework. There are also attempts to incorporate LLMs such as ChatGPT and Bard into MMM, for example, to refine market data or analyze scenarios.

Real-time analysis and response

Today, statistical modeling and machine learning algorithms have shorter execution times and more data than ever before, so MMM can be used for shorter-term simulations and analyses that were previously only possible using quarterly or annual data. Traditional attribution analytics can also integrate into MMM to quickly gain more comprehensive insights into the customer experience.

Leverage more data

Today, you have access to a wider variety of online customer behavior data (even when anonymized), including data from mobile, more external data, and the ability to leverage unstructured data. This allows you to understand and optimize your marketing strategy in more detail.

Analyze and simulate different scenarios

You can simulate various marketing scenarios to evaluate performance and make decisions.

Wider application

Whereas MMM was used primarily in the B2C sales industry, today it is possible to design and utilize MMM in a variety of marketing in various domains.

Open source MMM in big tech companies

MMMs can also utilize external solutions, but while it may be surprising to think of the large-scale solutions of yesteryear, there are MMMs that have been released as open source libraries.

As restrictions were placed on the use of personal information such as cookies, big tech companies that actively utilize advertising were the most affected, so they moved quickly to respond. They also researched MMMs and released them as open source. Google’s LightweightMMM and Meta’s Robyn are two examples of this.

Google’s LightweightMMM

LightweightMMM is a Python library for MMM that applies Bayesian modeling to MMM to improve the accuracy and precision of analysis. Bayesian modeling in MMM gives you the flexibility to handle uncertainty and variability in your data. The need to understand probability distributions is always a bit of a barrier for those new to Bayesian, but once understood, it can lead to better predictions and insights than traditional one-answer linear models.

(Source: https://github.com/google/lightweight_mmm/)

Robyn from Meta

Robyn is an open-source MMM library available in R and Python that uses Meta’s time series decomposition library Prophet and optimization library Nevergrad, which are widely used in time series analysis.

It supports hyperparameter optimization and various optimization methods, as well as various regression models and data transformations, so that each company (advertiser) can use MMM appropriately.

Meta has shown with Accenture that when it is difficult to use customer identifiers, MMM can be used to successfully perform multi-channel impact analysis without individual identifiers.

As an aside, you can see that the same Bayesian and time series decomposition used in Google’s and Meta’s time series analysis libraries is reflected in MMM.

(Source: https://github.com/facebookexperimental/Robyn)

Dispelling myths about MMM

If you’re a marketer who’s used MMM before, you may have a few misconceptions. And for those who are new to it, there are the usual misconceptions that come with any new tool. As with all tools, MMM can be helpful, but it’s not a silver bullet, so don’t expect too much and don’t trust too much.

Oversimplification and generalization

As with any data analysis, it’s tricky to interpret MMM results in terms of a simple “cause” and “effect,” and to think of the results as a single point, and to judge them as right or wrong.

Misunderstanding the effect of digital marketing channels

You might think that MMM is a product of the past, and that the effectiveness of digital marketing channels can’t be fully reflected in it. However, this is a problem that can be solved with good data collection and proper use of MMM, and Robyn mentioned above that there is no problem analyzing these different channels.

Misunderstanding the flexibility of the model

You might think that MMM is a static system and that the formulas and models don’t change. However, it is very flexible and can be adapted to suit your company and data. This is especially true if you utilize open source libraries.

Slow

Traditionally, MMM has only been used for long-term analysis and it has been too huge and slow, so you might think it’s not suitable for today. However, modern MMMs compensate for this drawback in a number of ways.

Overconfidence in universality

It’s easy to think that MMM results can be applied to any situation or market.However, every situation, environment, and data set is ever-changing, so it’s important to understand the context and background before interpreting these results.

Things to consider when adopting MMM

As with many services or systems that utilize data analytics, there are certain things to consider when implementing MMM, and there are better approaches.

Start small

Before making a big, company-wide commitment to MMM, start small with a small scope or a few marketing areas, fill in the gaps, and expand if the results are good. It’s also a good idea to iterate in small increments. MMM can be started as small as a library, so you don’t have to make big changes right from the start.

Collaborate with experts in your field

MMM is a process of interpreting complex results using a lot of data, so it’s important to collaborate with experts in each field. If you don’t have data experts, marketing experts, and domain experts working together, you may not be able to interpret the results properly or use the right models.

Clean your data

It’s important to make sure you have all the data you need and that it’s of good quality.As with any data analysis, the quality of your results will depend on the quality of your data, and garbage-in-garbage-out is the norm in this field.This is especially true for MMM, which requires analyzing time-series data, where failure to catch data corruption can lead to different patterns.

Buy-in from decision makers and stakeholders

It’s important that decision makers are aware of this, as it’s what they’ll be referring to, but it’s also important that stakeholders are fully informed and buy-in is sought as goals are set, data is collected and utilized differently, and information may change.

Epilogue

It’s an era of chaos for marketing analytics, but the goals of “maximizing performance” and “optimizing resources” remain the same, and data analytics can go hand-in-hand with them, just in different ways. Marketing performance analysis or results decomposition and optimization using time series data analysis has been going on for a while now, albeit in small ways, and it’s time to give it a systematic shape and bring it to the forefront.

And while MMM may not be new to you yet, the sooner you start adopting and adapting to it while we’re still in a period of disruption, the easier it will be to adapt to the changes and different environments ahead.