How open-source MMM frameworks democratize MMM usage

This article is co-written by Mathieu Lepoutre and Arnaud Parent.

‍

Navigating vast and complex information in a privacy-conscious world is paramount for making well-informed and highly operational decisions. Marketers face these questions: Which marketing activities yield the highest ROI? Where is the marketing budget best used? How do different marketing drivers, including paid, owned, and earned media channels but also promotions or pricing, work?  Do external factors, such as competitor activities or economic conditions, influence marketing performance? What is the short-term impact on sales as well as the longer-term tail related to brand equity?

‍

The limits of attribution

Traditional Multi-Touch Attribution (MTA) techniques fell short of meeting key needs mainly due to their inability to accurately capture the multi-channel / online and offline nature of customer journeys. They tend to overly focus on the very short-term impact of digital interactions, neglecting the impact of offline marketing efforts and external factors like seasonality or economic shifts. MTA assumes that all conversions are attributed to a media touchpoint, primarily digital media, which is a fundamentally flawed assumption, even for advertisers operating exclusively online with digital conversions. In many cases, a significant portion of the impact does not actually come from media efforts.

‍

Moreover, MTA approaches are severely hampered by data privacy constraints and the shrinking availability and reliability of granular user-level data, rendering them increasingly ineffective and meaningless in today’s rapidly changing marketing landscape, even with attempts to mitigate this through features like Consent Mode or Enhanced Conversion.

This is where MMM comes back into play, providing a holistic evaluation of each marketing component's effectiveness, considering the complex brand environment. It optimizes the marketing mix to maximize return on investment, enabling companies to make more informed and strategic marketing decisions. The growing availability of advanced data analytics, data science skills, and machine learning tools has enhanced MMM's precision and scalability, making it a more valuable and widely accessible asset.

‍

The benefits of open-source libraries

Furthermore, the primary and most significant advantage of MMMs lies in their predictive capabilities. Marketing decision-makers can simulate various hypotheses and scenarios before a campaign launch, enabling them to optimize strategies and have an immediate impact on results.

However, significant concerns persist for many advertisers before committing to operationalizing these techniques:

• Complexity: MMM requires expertise in statistical analysis, data processing, and domain knowledge, making it inaccessible for companies without a strong analytical background. Besides, building and validating MMM models is a time-consuming process, often taking several months to complete.
• High costs: Engaging specialized firms or commercial measurement solution vendors for MMM projects can be prohibitively expensive, often limiting these efforts to a few flagship brands and leaving many other brands in the portfolio without the necessary resources.‍
• Lack of transparency: Outsourced MMM-based systems remain black boxes, with advertisers having limited visibility into the methodologies and assumptions used, which hinders process adoption and trust in the results.

As a result, the comprehensive services provided by third-party MMM vendors, who typically manage the entire process from discovery and data collection to modeling and insights/recommendations, are facing significant challenges.In-housing some or all components of a bespoke marketing effectiveness measurement and optimization platform is gaining momentum among major advertisers and appears now as a realistic alternative for an increasing number of marketers. For those still in the consideration phase and/or when adopting a SaaS MMM solution from specialized providers seems less relevant due to limited customization, potential data integration issues, or high ongoing subscription costs, a critical question is how the recently released or upgraded open-source MMM packages can act as a catalyst for change.

‍Key open-source MMM packages include some that are directly provided by major digital platforms:

> Meta’s Robyn: A publicly accessible codebase for semi-automated MMM that leverages machine learning techniques to accelerate the modeling process, minimize analyst bias and subjectivity and produce more controllable and scalable models. Additionally, it integrates seamlessly with advanced, user-friendly Meta tools like Nevergrad, facilitating efficient optimization of budget allocations.

>Google’s Meridian: A recently released set of Python libraries that provides data science teams with the foundational tools to explore MMM-style measurement using established and reliable techniques, representing a new generation of solutions by Google following LightweightMMM. This offers advertisers a significant advancement, as its fully Bayesian core allows for highly flexible model specification, estimating all parameters with uncertainty rather than relying on modeler assumptions. Additionally, it extends beyond traditional mainly by enabling the estimation of a geographical hierarchical model and - naturally expected from Google - improving the assessment of Search's role in effectiveness and integrating reach and frequency measures to optimize video advertising decisions. Learn more about this feature in this article: Why use reach and frequency instead of impressions in Marketing Mix Models? 

These packages clearly provide advanced features, seamless integration with Meta or Google ecosystems, and robust security. However, some advertisers seeking greater flexibility, customization, and independence often consider turning to community-developed code packages as an alternative approach. As an example, we can mention

> PyMC-Marketing: An open-source MMM solution developed by a team of PyMC Labs researchers and a community of experts, built on top of PyMC, a widely-used probabilistic programming library for constructing Bayesian models. While it is a powerful and flexible tool for MMM, its complexity, computational demands, and reliance on Bayesian methods make it best suited for experienced data scientists proficient in Python and Bayesian statistics. The package requires a significant investment in time and resources to set up and use effectively, and it may not be the best choice for users looking for automated, easy-to-use solutions.

‍