Bayesian Media Mix Model Calibrated to Geo Tests

The number this produces: Incremental contribution per channel, with a credible interval, at the spend level you are actually running. A model that reports paid social returned 3.1 is not decision-grade. A model that says the incremental return sits between 1.4 and 2.9 and that the curve flattens above 90,000 EUR a week tells you what to change on Monday and how confident to be while changing it. How it is built: Weekly data, never daily. Geometric adstock and a Hill saturation term per channel, estimated in PyMC. Priors are load-bearing rather than decorative: where you have run a geo holdout or a conversion lift study, its result becomes the prior on that channel's coefficient, which is the difference between a model that agrees with your experiments and one that quietly argues with them. Controls for price, distribution, promotion, weather where it matters — energy switching and heating degree days are not independent — and the fixture calendar where it matters more. Validation is out of sample: the final thirteen weeks are held back and the model has to predict them before anyone is shown a chart. What has to be true before we take the work: We want 156 weeks of weekly spend by channel. 104 is the floor. Below that the adstock and saturation parameters are not identified and any number we hand over is a prior wearing a posterior's clothes. Spend also has to have moved. If a channel has run within ten percent of the same weekly figure for two years, the data contains no information about its curve, and the honest recommendation is to run an experiment rather than pay us to model an absence. What is delivered: The model itself — code, priors and specification — in your repository rather than ours. Response and saturation curves per channel. A budget allocator that respects your contractual minimums instead of pretending they can be switched off. A written account of the model's failure modes: which channels are weakly identified, which coefficients are carried by a single quarter, and what would have to happen for the conclusion to reverse. A quarterly refresh protocol your analysts can run without us. Where we part company with standard practice: Brand search is treated as incremental only when a brand search holdout has been run. Without one it enters as a control, not as a media channel, because a model that credits brand search with conversions it merely witnessed will recommend spending more on it indefinitely. We have rebuilt three Robyn models in which the winning channel changed with the random seed; that is not a modelling result, it is a lottery with a plot. We also refuse to stack a model on top of last-click attribution and present their agreement as validation. They share the same bias, so of course they agree. Not for you if: Your reporting cannot produce weekly spend by channel and market without someone assembling it by hand each time. You need the answer this month. You cannot run a geo test at any point in the next two quarters, without which this is curve fitting with confident error bars. Or the model is wanted to close an internal argument that has already been settled; we will produce what the data supports, including on the occasions when that is the inconvenient answer.