Bayesian Media Mix Modeling for App Install Spend

We decline media mix work below roughly two hundred and fifty thousand dollars a month across at least three channels with eighteen months of weekly history behind it. Under that, the credible intervals on the channel coefficients come out wider than the budget moves you would make from them, and you have paid for a model whose honest answer is a shrug. The first package exists to establish that before anyone commits to a build, and it ends in a recommendation not to build one more often than you would expect. How the model is built: Bayesian, in Meridian or PyMC-Marketing depending on the shape of your data, weekly, with geometric adstock and Hill saturation fitted per channel rather than assumed from a template. Priors are the part that decides whether the model is useful at all. We set them from experiments you have already run: geo holdouts, conversion lift studies, ghost ad tests. That way the model is calibrated to measured incrementality rather than to whatever correlation your spend curve happens to carry. Where no experiment exists for a channel, the prior stays wide and the output says so, instead of letting the fit manufacture a certainty nobody earned. The app-specific corrections: Installs are not the outcome, so the model runs against revenue with the install-to-revenue lag encoded from your own cohort curves. An app earning most of its lifetime value across months two through six cannot be modeled on the week its installs landed. Apple Search Ads branded terms are separated from everything else, because leaving them blended is how existing brand demand gets billed to whichever channel happened to be spending. Organic is a modeled outcome with its own drivers, not the residual bucket everything unexplained falls into. Exogenous drivers people leave out: For a crypto app, the dominant weekly driver of installs is usually the asset price, not your media. A model without a price regressor hands the acquisition team credit in a bull month and blame in a flat one, and both readings are wrong in ways that outlast the model. For beauty, retail promotional calendars and the two large seasonal sale windows outweigh most media variation in the fourth quarter. We encode these as controls and show the fit with and without them, so you can see how much of last year's story was demand rather than buying. Validation and delivery: Out-of-sample holdout on the final weeks, plus a comparison of the model's read against every incrementality test run since the training period ended. You receive the model in your own repository with its data pipeline, a written interpretation of each channel's response curve including the point where it stops paying, a scenario tool that shows intervals alongside every point estimate, and a working session where your finance and channel leads argue with us about the result. What we do not do: We do not let the model overrule an experiment. Experiments are the ground truth the model is calibrated against, so when they disagree the model is wrong and gets refit. We do not deliver a point estimate without its interval. We do not hand over an optimizer meant to run unattended between refreshes, because a six-month-stale model recommends last season's allocation with complete confidence. And we do not build MMM for a team that intends to stop measuring per channel: this reconciles attribution and testing, it does not replace either. Who this is not for: Apps with one meaningful channel, where a holdout test answers the question for a fraction of the cost. Teams below the spend and history thresholds. And anyone hoping a model will settle an internal argument about whose channel is best, because the intervals will not be narrow enough to win it and we will not narrow them to help.