Marketing Mix Model

Here is a quick brain dump on marketing mix models and why a Bayesian approach is useful for this task.

Caveat: My experience is that most people working in digital marketing don’t have time, capacity or know how to do (Bayesian) modeling, they are busy writing copy, twiddling buttons or looking at dashboards someone else built.

I saw on social media that there is an interesting workshop on Bayesian uncertainty quantification. I saw that they say that they don’t cover MMMs. Err what are MMMs? So I thought I would write a quick post on what is a marketing mix model and why a Bayesian approach is useful for this task.

Perhaps when I’m done with my backlog I’ll do a worked MMM example in PyMC as I had already done one for one of my projects as a Capstone Project in one of the courses in the Bayesian Statistics Specialization I took recently on Coursera.

What is a marketing mix model?

Product, Price, Place, and Promotion

Back in the day I learned about Marketing Mix form reading one of Philip Kotler’s many tomes on Marketing Management. If I recall correctly marketing Mix refers to the brands or products that a firm offers to its customers. The marketing mix is often referred to as the “4 Ps” of marketing: Product, Price, Place, and Promotion. But the wikipedia article suggest that there are Seven or even Eight P for different marketers.

The marketing mix is a framework that helps businesses plan and execute their marketing strategies effectively.

This leads to the question of what is a marketing mix model?

Marketing mix models (MMM) are statistical models that estimate the impact of various marketing tactics on sales and other key performance indicators (KPIs). They help businesses understand how different marketing channels contribute to their overall performance, allowing them to optimize their marketing spend and strategies.

In reality though we are talking about when we say MMM model is inferring the the response of different media channels on a KPI like sales.

What does a Bayesian marketing mix model add?

The Bayesian approach to marketing mix modeling offers several advantages over traditional methods:

1. Updating with New Data: Bayesian models can be updated as new data becomes available, allowing for more dynamic and responsive marketing strategies. This is particularly useful as online marketing provides lots of data where consumer response can be fast-changing and agility can translate into competitive advantage. In contrast non-Bayesian models tend to be static which causes model drift and requires retraining validation etc.
2. Uncertainty Quantification: Bayesian models provide a natural way to quantify uncertainty in the estimates. This are known unknowns that are essential for decision-making, as it allows marketers to understand the confidence they can have in the results and make more informed choices. Non-Bayesian models give a best estimate but can’t quantify uncertainty in a way that is useful for decision-makers.
3. Incorporation of Prior Knowledge: Bayesian models require the integration of prior knowledge or expert opinions into the analysis. This is more work but often turns out to be an advantage This is particularly useful when historical data is limited or when certain marketing channels have known effects based on previous campaigns.
4. Flexibility: Bayesian models can easily accommodate complex relationships and interactions between marketing channels.
5. Model Checking and Validation: Bayesian methods provide a framework for model checking and validation. We can use posterior predictive checks to assess how well the model fits the data and identify potential areas for improvement.

The Bayesian approach also has many other bit and pieces of machinery which can be incorporated into the workflow.

• E.g. we can use empirical bayes to set priors or hyper priors based on the data. We can avoid the tunnel vision of point estimates and instead look at the full posterior distribution which can let us avoid making bad decisions. MCMC methods also provide diagnostics to check model fit and convergence.