What is Bayesian marketing mix modeling?

A Bayesian marketing mix model treats every unknown in the model, such as how strongly a channel works and where it saturates, as a distribution rather than a number. Before the model sees any data, each distribution is a prior: what the company and the modeler already believe about that channel. The data then updates each prior into a posterior, which is the model's answer: a most likely value and the range around it.

The alternative, a classic regression, starts from nothing and lets the data decide everything. With marketing data that is a problem. There is often one data point per week to estimate roughly ten activity types, and a model given that little data and no starting point overfits: it explains the past beautifully and forecasts poorly.

The model is built before it looks at the data. Historical spend gives reasonable levels and a first guess at marginal return per channel. The digital platforms' own conversion and impression data give a saturation signal per channel. Structured interviews with the team add what the numbers cannot: if you raised this channel 25%, how much more response would you expect? Together these become the priors.

Training then treats the history as a set of experiments. Where the data disagrees with a prior, the posterior moves; where the data is thin or noisy, the prior carries more of the weight and the range stays wide. That is the point: the model says less where it knows less.

Two things follow. First, the model is stable. Two years of history sit behind each new month, so a monthly retrain moves the answers slowly rather than flipping them. Second, every recommendation carries a range, so a budget decision can be made at the level of certainty the business needs. A target that has to be hit with 80% confidence calls for a different budget than one that only has to be likely.

Being Bayesian is no longer a differentiator on its own. Google Meridian, PyMC-Marketing and most commercial vendors are Bayesian too. What separates models is how carefully the priors are set, especially for a single national market where the data is thin. That discipline is where Odins puts its effort. How Odins sets priors in data-sparse markets.

• 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.
• Posterior: The model's answer: the prior after the data has had its say, with a most likely value and a range.
• Confidence range: The range a Bayesian model reports around every number. A recommendation without one is hiding something.
• Marketing mix modeling (MMM): The statistical method behind everything else on this list: spend in, sales out, every channel on equal terms.
• Calibration: Anchoring a model to outside evidence, usually experiment results. In a Bayesian model, calibration is a prior.

Go one level deeper: Why Bayesian, and what it changes. The technical view: informed priors. Or go back to all terms in the glossary.