How We Made the Most of 150 Data Points

When it comes to marketing effectiveness, analysts often rely on digital advertising. In reality, though, marketing is a mix of digital and offline activity run in flights, while the available data is aggregated weekly and comprises just 150 rows over three years. Classical ML models under these conditions produce negative weights for effective channels, and time-series models answer the question "What will happen next?" rather than "What was driving sales?"

In this presentation, I'll use a real-world case study to demonstrate how to use Bayesian regression in PyMC-Marketing to build a causal, interpretable model of the contribution of TV, radio, OOH, and digital to sales. I'll also explore how to validate the model when data is insufficient by conventional standards, and what makes us confident in this model.