Meta Robyn

Robyn is Meta's open-source MMM, built in R with a Python port. Instead of Bayesian inference it uses ridge regression, a regularized regression that shrinks unstable coefficients, wrapped in an evolutionary hyperparameter search (Nevergrad) that tries thousands of adstock and saturation settings and returns a set of candidate models rather than one answer.

That last part is the workflow surprise: Robyn hands you a Pareto front of plausible models and expects a human to pick one using business judgment and its model-selection plots. Teams expecting a single number find this unsettling; statisticians consider it honest.

Known gotchas from the documentation: the hyperparameter search is compute-hungry, seasonality is delegated to Prophet, and uncertainty is not expressed as posterior distributions the way Bayesian frameworks report it.