What is calibration in MMM?

A model trained on history alone can only know what the history shows, and for a channel that always ran at the same level, that is not much. Calibration brings in evidence the history does not contain: a conversion lift study, a geo experiment, a documented benchmark. The model is anchored to that evidence so its estimate for the channel starts from causal ground.

In a Bayesian marketing mix model there is no separate calibration step; calibration is what priors are. An experiment's point estimate and its standard error become a distribution on that channel's return, and the model starts there. The data then updates it. If the history agrees, the posterior narrows; if the history disagrees, the posterior moves and the disagreement is worth a conversation.

Attribution data can also inform priors, but only after debiasing. Last click over-credits bottom-funnel channels and platform ROAS is inflated, so raw attributed returns are never used directly; they inform the starting point together with what the team knows.

Odins takes experiment results into the model directly, the customer's existing lift and geo studies included, and designs new tests where the model's range is too wide to act on. The nuances matter: a test measures one channel in one period and one population, so timing, what was measured and who was exposed are all read before the result becomes a prior. A test designed with the model in mind gives a number the model can use; one designed without it often does not.

• Prior: What the model believes about a channel before it sees the data. The part of MMM that decides how much the data can teach.
• Incrementality test: A controlled experiment that measures what one channel causes, by comparing exposed and unexposed groups or periods.
• Triangulation: Combining MMM, attribution and experiments. The question is where: inside one model, or in a dashboard afterwards.
• Geo experiment: Run the activity in some regions and not others, then compare. The classic causal test, when the market is big enough to split.
• Bayesian marketing mix modeling: A model that starts from what you already know and lets the data correct it, with a range on every answer.

Go one level deeper: Triangulation: MMM, attribution and experiments. The technical view: informed priors. Or go back to all terms in the glossary.