The future is modeled

During the past years, businesses have been experiencing a deep digital transformation that is provoking a profound change in their marketing measurement landscape.

As a consequence, the boundaries that used to separate the use cases for econometric marketing response models and digital attribution are narrowing.

Advanced Marketing Mix Models are now more crucial than ever to respond to a converging and multi-technique approach, attracting Multi-touch attribution (MTA) and experiments to its orbit. Advanced MMMs allow to measure online and offline media investments, which are also resilient to signal-loss scenarios and are privacy-compliant.

6 keys to have in mind when building Advanced Marketing Mix Models

1. Consider modeling your funnel in full

Advanced MMMs are very flexible, it is possible to measure any response variable: from brand awareness to website visits, first-time buyers, sales in both physical and online stores, among others. MMMs’ omni-channel capabilities are like no other methodology in the measurement landscape. Building a complete ecosystem of connected models, trying to reflect the entire Consumer Journey, gives advertisers a more complete picture of business dynamics, even if the modellers have to guarantee a good level of automation to ensure not causing an impact on the speed of decision making. The balance between completeness and agility is one of the most important decisions to be made before starting any MMM project.

2. Customize the solution to suit your business.

MMMs offer a remarkable added value: the combination of business and media perspective in one unique model, with the possibility of a global and combined understanding of structural and dynamic drivers of growth. For this reason, advanced MMMs have to be adapted to the specific context of each advertiser. The advertisers should be focusing on building models that reflect your business reality through media (Ad-stock effects and diminishing return curves, etc.) and non-media variables (Pricing, Promotions, Competitors, new products releases, new business units, direct searches, etc.). The architecture of this construction has to be solid and robust in terms of methodology (following certified standards), but the decoration and the furniture is completely a customized solution.

3. Unlock the power of AI-automated code to increase scalability and reduce human bias

On one hand, reducing human bias is one of the biggest concerns of traditional MMMs. On the other hand, the modelling cadence has to be increased to weekly or daily results to accelerate decision making. Advanced MMMs should leverage automated code and techniques that are scalable and objective. Artificial Intelligence techniques such as those within Facebook’s Nevergrad open-source platform, optimize the way to explore over possible model solutions reaching optimum results faster and in a more precise way than traditional algorithms. And Facebook’s project Robyn, the open-sourced R code for semiautomated MMM using Machine Learning techniques, pursues a similar objective: increase the scalability and the objectivity of the analysis.

4. Build fast models that are flexible and adapt to future scenarios.

To improve model predictability, modellers have to work on the penalization of highly correlated explanatory variables. Regularization techniques, such as lasso or ridge regression help to improve the predictive performance of MMMs, avoiding overfitting and helping on variable mis-selection. Therefore, providing more flexible models that allow deeper levels of detail and more robust results, with a better balance between analytical and business ingredients.

5. Explore the relationship between experiments, attribution and MMMs

Advanced MMMs can be calibrated with experiments and attribution results to ensure a consolidated solution containing incremental results. There are three different ways to make these solutions converge into MMMs:

1. Constraining MMM models with experiment and/or attribution results via priors, guiding modelling towards solutions that reflect the experiments’ effect size;
2. Selecting MMM models that minimize the distance between model predicted results and results from experiments or attribution and
3. Validating MMM results by just comparing the predicted contribution for a channel compared to its experiment or attribution results within a certain period of time.

6. Focus on actionable results.

Advanced MMMs should be automated enough to allow continuous modelling that describes the performance of your marketing in almost real time. Finding balance on data granularity levels here is key (e.g. Modelling different types of campaigns like prospecting vs. Retargeting versus modelling each campaign separately). Finally, integrating a marketing budget optimizer with the ability to apply custom restraints that suit each advertiser’s business, can provide great recommendations about how to switch budgets in the media mix for the upcoming cycles.