Joël Bühler, Head of Marketing Intelligence & Technology at SMG Swiss Marketplace Group, gives us an insight into Marketing Mix Modelling, why Lego helps to understand it, and how it is applied at SMG.
MMM is often an application of time series regression models, usually with two specific characteristics: First, one tries to account for a time-lagged effect: This is because marketing expenditures are expected to influence consumers’ decisions at a future point in time. Second, marketing expenditures or contacts, which are an important input to the model, are also transformed to account for the effect of diminishing marginal utility.
The Lego Principle
A Structural Time Series Model is used as the basis for MMM at SMG. Many of these “time series” share common features, such as a general upward or downward trend, repeating patterns, or sudden increases or decreases. Structural approaches to time series explicitly account for these features by representing an observed time series as a combination of components.
Thus, in these models, decomposition into individual, interpretable components (such as trend and seasonality) is central. The analogy to Lego building blocks is, therefore, quite obvious. But what exactly are these components?
The 4 Building Blocks
- Trend – This component is closely related to brand equity or the factors that contribute to the continuous strengthening or weakening of the company’s success.
- Seasonality – These are influences that depend on the weekly or annual calendar. An example is the well-known “summer slump”, or the different user behaviour during the week vs the weekend.
- External and internal factors – Unlike seasonality, this component is less predictable and not periodic. This includes, for example, whether the weather has been good or bad, competitors have launched a new campaign, or significant product changes have been made.
- Marketing spend – Probably the most interesting component. More specifically, this is about the effect of marketing spend, as this is quantified in the metric of the defined KPI.
However, which factors should finally be part of the model depends on the target. Only strongly delayed effects are difficult to quantify. Example: If a user decides to use a product today because of an ad, that user may generate constant value over the next 2 years. To map this value in MMM and attribute it to marketing activities is extremely difficult.
MMM at SMG Swiss Marketplace Group
At SMG, a Bayesian Structural Time Series Model is used. The time series components are extremely close to Facebook’s Prophet model. More can be read about this in The American Statistician (72,1), Forecasting at Scale by S.J. Taylor and B. Lethman. However, SMG uses the Python package PYMC3 for modelling. Modelling the time-lagged marketing effects and saturation are the “secret ingredient” in it.
Basically, better performance is observed with digital channels with the use of Marketing Mix Modelling. In this article, we were able to give a little insight into the topic and how it is applied at SMG. Many thanks to Joël Bühler for the insight.
For those who want to know more:
Broadbent, S. (1979). One way TV advertisements work. Journal of the Market Research Society, 21(3), 139-166.
Liu, Y., Laguna, J., Wright, M., & He, H. (2014). Media mix modeling–A Monte Carlo simulation study. Journal of Marketing Analytics, 2(3), 173-186.
Taylor, S. J., & Letham, B. (2018). Forecasting at Scale. The American Statistician, 72(1), 37-45.
Jin, Y., Wang, Y., Sun, Y., Chan, D., & Koehler, J. (2017). Bayesian methods for media mix modeling with carryover and shape effects. https://research.google/pubs/pub46001.pdf
Joël Bühler, Head of Marketing Intelligence & Technology at SMG Swiss Marketplace Group LinkedIn