Marketing Media Mix Models with Python & PyMC: a Case Study [PyCon DE & PyData Berlin 2024]

PyMC enables building and solving marketing mix models with Bayesian regression, providing uncertainty estimates for parameters

Marketing contribution analysis is more practical than attribution tracking - focus on aggregate channel impact rather than individual customer journeys

Ad stock transformation helps model the accumulated/persistent effects of marketing investments over time

Cross-correlation and Spearman correlation are useful for analyzing relationships between marketing spend and sales KPIs, especially for non-linear patterns

Setting appropriate priors is crucial - use domain knowledge to constrain parameters (e.g., half-normal distribution for ensuring positive returns)

Model validation should be done out-of-sample with test sets to avoid overfitting

Baseline effects (brand strength, seasonality) need to be separated from marketing channel contributions

Linear regression alone is insufficient due to:

• High error with high promotion levels
• Potential negative baselines
• Coefficient instability

Channel timing is critical - investments should align with when customers are most likely to convert

Privacy changes (cookie deprecation) will make attribution tracking increasingly difficult, reinforcing the need for contribution-based modeling approaches