MMM Re-Discovered! By Facebook’s Dimple Dinesh

By Dimple Dinesh, marketing science partner, MENA, Facebook.

No one could have predicted what 2020, the most awaited year ever, had in store for us, from a global pandemic to profound tremors within the digital media industry. The clamour around stricter data protection laws and anti-tracking policies has hit our industry in extraordinary ways. As a result, and the same way we are adjusting to a Covid new normal, we must start gearing up for a “digital new normal.”

This digital new normal is two-fold. Firstly, businesses can no longer spend their ad dollars sub-optimally and report incorrect ROI numbers. We need to attribute conversions to the right media channels and stop basing decisions mostly on intuition or outdated data. Secondly, due to the pandemic, consumer behaviour has changed dramatically. Consumers are spending more time online now than they ever have before, which compels businesses to reach them in the right place by possibly considering different media budget splits. Businesses’ longevity will hinge heavily on first-party data to mitigate the new normal risks.

A measurement study gaining attention in line with these trends is the marketing mix modelling (MMM). MMM is a data-driven statistical analysis that quantifies the incremental sales impact and ROI of marketing and non-marketing activities. It enables marketeers to justify budget allocations and optimise future budgets.

Highly resilient and using on first-party data, MMM is not impacted by the changes happening today. It does not rely on individual people, IDs, or consumer journeys, but aggregates all the data in time series.

However, MMMs are heavily reliant on data quality and accuracy, which implies significant drawbacks: to build them is time-consuming, results are often outdated, and they lack sufficient detail and granularity. But these flaws and uncertainties can be addressed and the traditional ways of running these studies can be modernised so that this silver lining of opportunity can be seized, driving innovation. MMMs are here to stay.

So, here are five ways we could adopt contemporary methodologies for MMM:

  • Shorter models: MMM providers must quickly deliver the most up-to-date media efficiency results that will allow for a close-to-ongoing media-mix optimisation. Shortening the modelled period from the standard ~2/3 years could be a promising way to make MMMs more capable. These shorter models should contain greater granularities (e.g., at DMA or store level).
  • Recency variables: Media platforms change over time. A single variable over a ~3+ years may not precisely capture the next dollar’s true impact historically and looking forward. We believe there is substantial value in splitting variables for recency. For example, instead of considering an entire media X for 3 years, we could split this variable into media X – year 1-2 and separate the recent year media X – year 3 as a different variable.
  • Customized decaying effect: Every media channel’s impact wears off after a period once a consumer watches the ad. The decaying effect of a media channel is called an ‘ad stock’. Applying traditional decaying effect means that a media channel’s impact wears off by a set percentage each week. But in the ever-changing digital ecosystem, the way consumers interact with different media channels changes quite rapidly. It is beneficial to test innovative decaying transformations for each media channel, respectively, and not use one-size-fits-all methods.
  • Different modelling approaches: MMM providers in the market are using different modelling approaches to drive innovation. This keeps the wheel of modernization spinning. For example, in most cases, a traditional MMM considers the impact of all variables that influence sales to be static over the modelling period. But this may always not be true, and some variables could be dynamic in nature. State-space modelling is one of the modelling types within the regression family that could be explored. Using the state space modelling approach helps increase the models’ accuracy, overcome data irregularities, and is better suited for model automation.
  • Calibrated with experiments: MMMs are inherently complex studies. The insights generated from MMM models obviously depend on the value of the parameters found by running times series modelling. But how can we make sure these models are accurate?
    • One method of validation is to compare the incremental value showcased in the model (incremental cost per KPI) with the platform’s experimental tests. For example, on Facebook, we run lift studies to qualitatively compare the lift study results with the model’s incremental value reflected for Facebook.
    • Using results from lift studies to choose models. Compare the cost per incremental KPI from a study, such as conversion lift, to the MMM’s contribution and choose the model that minimizes the difference.
    • The most rigorous and harder way to implement is to incorporate the lift studies results into the MMM models.

In conclusion, MMM is a resilient, insightful tool that can help businesses optimize their marketing spends. We need to innovate and transform our methodologies to re-discover it and adapt to changes. In the immortal words of Don Draper from Mad Men, “Change is neither good nor bad. It simply is.”