Model

A glimpse under the hood of Sellforte's powerful Marketing Mix Modelling platform methodology

Model

Sellforte's methodology

MMM in a nutshell

With historical media and sales data of 2-3 years, the model learns the causality between marketing activity and sales uplift and attributes the sales to the given activity, and thus, media uplift and ROI can be calculated.

Activity can be…

  • Media investment or an Affiliate
  • Promotion or a discount

…which has time-to-time variation in investments, media metrics, timing, length, media mix, or other things that let the model separate multiple activities from each other.

MMM in a nutshell

Bayesian MMM

Sellforte’s approach to Marketing Mix Modelling (MMM) is a Bayesian hierarchical modeling methodology that has been the gold standard for MMM since 2017 when Google’s research team decided to sum up the problems with earlier methods in MMM.

Bayesian hierarchical models enable the usage of priors in modeling which helps to manage the noise of the data, but also the model to learn from data when there are statistically significant differences e.g. TV channel Y has higher ROI than TV channel X.



Bayesian MMM

Model calibration

Sellforte's modeling results are always calibrated with the best available data.

The industry standard is to calibrate MMM results with lift tests, but quite often when we start, there is no such information available. Thus, Google Analytics attribution and Ad platform attribution results are used to calibrate the MMM results for those specific mediums as often the best information for a given platform sits in their analytics.

Model calibration

Model validation

Sellforte platform includes a section model validation where you can find Insample MAPE (Mean Absolute Percentage Error) and R2.

These numbers will tell you how well the model predicts future sales based on the trained model.


Model validation

FAQ

Bayesian hierarchical modeling has been chosen for the methodology because it's the Golden standard for MMM according to Gartner and it works well with extremely big and extremely small data sets.

Also, it allows usage of informative priors as believe not all information cannot possibly be founded by models, so it needs to be given additional information to be able to find uplifts, that create insights that will be useful in everyday life.

Want to see how results look like in the Dashboards?


Read more about Dashboards