Meridian explicitly states that out-of-sample prediction metrics are only structural checks and cannot directly validate causal inference; correctly aligned experiments are needed to assess causal marketing effects.

Counterevidence & uncertainty

A model can predict held-out outcomes while assigning biased channel effects when confounding, controls or priors are wrong.

What would change the reading

Update if validated methodological evidence shows predictive fit alone identifies causal channel effects under the deployed data-generating process.

Primary routes

1. MB-S001https://developers.google.com/meridian/docs/post-modeling/model-fitOpen source record →