Bayesian App MMM Calibrated Against Geo Holdouts

Two years of weekly data, or the model tells you what you already believe. We build media mix models for app publishers with at least 104 weeks of spend history across six or more country-level markets, and we calibrate the channel coefficients against geo holdout tests rather than against last-click reports. Below that history, and without one clean holdout to calibrate on, the credible intervals come out wide enough to justify whatever budget the person commissioning the model already wanted. How the model is built: Hierarchical Bayesian regression in PyMC-Marketing, one partially pooled level per country, weekly grain. Geometric adstock with a fitted decay per channel, Hill saturation on spend. Controls for price changes, App Store and Play feature placements, platform seasonality and the Italian August collapse, which moves domestic install volume further than any paid channel we have modelled. Apple Search Ads brand and non-brand enter separately, because pooling them fits a saturation curve to a term that has none. Post-ATT, SKAN conversion values are used as a sanity check on the model's iOS estimates and never as an input. Calibration: Each channel's prior comes from a measured lift test where one exists. Where none exists we say so on the coefficient itself, in the report, and we name the test to run next and roughly what it costs to run. A Meta coefficient that has never been checked against a geo holdout is an opinion with error bars, and we will not present it as anything else. What you get: A written read of the marginal return per channel at current spend, the point where each channel stops paying, and three budget scenarios given as revenue distributions rather than point estimates. The model code and data pipeline are handed over as a repository you own, so your own analyst can refit it. A one-day working session with your growth and finance leads to argue with the results before they harden into a plan. Refit cadence: Quarterly. Monthly refits on weekly data change nothing except the story, and we have watched teams re-plan four times in a row off noise. What is not included: Attribution setup, MMP configuration, SKAN schema design, and dashboards. We do not build a self-updating model wired into a BI tool. Media mix modelling answers a budgeting question asked four times a year; putting it behind a daily refresh creates a number people check every morning and act on every week, which is worse than having no number. Who this is not for: Publishers spending under roughly 150,000 euro a quarter on user acquisition. Single-channel advertisers, where the model has nothing to separate. And anyone who needs the answer to be that the current mix is correct. If your spend is recorded at campaign level with no country breakdown, the model cannot separate geography from channel, and you will hear that on the scoping call rather than after the invoice. Data we need: Weekly spend by channel and country, installs and in-app revenue on the same grain, subscription or licence revenue kept separate from ad revenue, and dates for pricing changes, store placements and app releases. A warehouse connection or CSV, either is fine. How we are paid: Fixed fee. We never take a share of the budget the model reallocates, and we do not write the conclusions before the model runs.