Guide: Calculating Informative Priors for MMM

This guide describes how to calculate tighter, more informative priors for PyMC-Marketing MMM based on data characteristics.

Important: The default wide priors (intercept: Normal(mu=0, sigma=2)) often lead to better-calibrated posteriors. Only use informative priors when you have strong domain knowledge or specific requirements. Overly tight priors can lead to overconfident (underdispersed) posteriors.

When to Use Informative Priors

• You have domain knowledge about expected ROAS ranges, decay rates, or effect sizes
• You have results from previous studies or models to inform your beliefs
• Business constraints require certain parameter ranges
• You're doing sensitivity analysis to understand prior impact

Step 1: Calculate Scaled Statistics

PyMC-Marketing uses MaxAbsScaler for target and channels (divides by max absolute value).