Bayesian MMM Explained: Adstock, Saturation, Priors and Calibration

Bayesian marketing mix modeling has gone from a niche statistical technique to the default approach for serious in-house MMM programs in just a few years. If you are evaluating Bayesian MMM, this guide walks through what it actually is, how the components fit together, and what changes about the way your team makes media decisions when you adopt it.

What does Bayesian mean here

A Bayesian MMM model treats unknown quantities, like the ROI of paid search, the adstock decay of TV, or the diminishing returns curve of display, as probability distributions instead of single numbers. You start with a prior belief about each parameter, the model observes your data, and the posterior distribution captures what you should believe given both the prior and the evidence.

In practical terms, this is the difference between a model that says paid social ROI is 3.2 and a model that says paid social ROI is most likely between 2.4 and 4.1, with a 90 percent credible interval. The second answer is more honest, more actionable, and harder to oversell.

Frequentist vs Bayesian MMM

Adstock: media has a memory

Adstock is the assumption that advertising effects do not vanish the moment the campaign ends. A TV ad seen this week may still drive sales next week. The most common formulations are geometric adstock, where the effect decays by a fixed percentage each period, and Weibull adstock, which allows the decay shape to be more flexible. PyMC-Marketing supports both.

• Geometric adstock: simple, well understood, often a good default for digital channels
• Delayed adstock: introduces a peak some weeks after spend, useful for channels with awareness lag
• Weibull adstock: more flexible decay shape, can fit channels that behave non-geometrically

Saturation: diminishing returns are real

Saturation models capture the fact that the first 100k spent in a channel often works harder than the next 100k. The standard functional forms are the Hill function and the logistic function. Both produce an S-shaped or concave response curve where marginal returns flatten as spend grows.

• Without saturation, the model implicitly believes you can scale a channel infinitely without diminishing returns
• With saturation, the optimizer can find the inflection point where additional spend starts to underperform
• Saturation is what turns an MMM model into a useful budget planning tool

Priors: where your knowledge lives

Priors are the most misunderstood part of Bayesian MMM. People sometimes think priors mean the model is biased. The opposite is true: priors are the place where your existing knowledge becomes explicit and challengeable. A prior is a statement like before seeing the data, we believe the ROI of TV is somewhere between 0.8 and 2.5, with a most likely value around 1.5.

If your team has historical experiments, prior MMM results, or industry benchmarks, those go into priors. The data then updates those beliefs. The result is a model that respects what you already know without ignoring new evidence.

Lift tests and calibration

MMM and incrementality testing are often presented as alternatives. They are not. They are complementary. A geo-experiment or a holdout test produces a clean, causal estimate of a single channel at a single moment. MMM produces a comprehensive, ongoing view of all channels but is more reliant on assumptions. The combination is stronger than either alone.

• Use lift tests to set tight, informative priors on calibrated channels
• Use MMM to extrapolate to channels and periods you cannot easily experiment on
• Refresh priors as new experiments come in

Uncertainty quantification

Because the output of a Bayesian MMM is a distribution, you can answer questions a frequentist model struggles with: what is the probability that paid social ROI is greater than 2? What is the credible interval around the optimal budget? How likely is it that the optimizer recommendation is at least 5 percent better than current spend?

Posterior interpretation

Once you have a fitted model, the typical outputs are channel contributions over time, ROI per channel with credible intervals, response curves, posterior predictive checks, and budget recommendations. Each of these can be reported with uncertainty bands rather than single numbers, which dramatically changes the conversation in stakeholder meetings.

Diagnostics that matter

• Convergence checks: R-hat, effective sample size, divergences from the NUTS sampler
• Posterior predictive checks: does the model recreate observed sales patterns
• Holdout performance: how does the model do on time periods it did not see
• Sensitivity to priors: do reasonable changes to priors produce reasonable changes to outputs

From model to budget decision

A model that is never used to change behavior is a model that does not exist for the business. The full Bayesian MMM workflow has four steps: build the model, validate it, use the posterior for scenario analysis, and run optimization with risk constraints. SIMBA wraps all four steps into a single product so the team does not have to glue them together.

See the Bayesian MMM page for how SIMBA exposes priors, calibration, and uncertainty in the platform, and the marketing mix modeling software page for the broader product workflow.