Bayesian Marketing Mix Modelling with Holdout Calibration

Marketing mix modelling is only worth commissioning if the model can be wrong in a way you would notice, and that needs two things most engagements do not have: at least 104 weeks of weekly history, and enough deliberate variation in spend across that history for the model to separate one channel from another. If your budgets have been flat and always-on for two years, no mix model can tell you what brand search contributed, and no amount of Bayesian machinery changes that. We check both conditions before quoting, and we have declined this work more often than we have taken it. What we build: A Bayesian model in PyMC-Marketing, fitted on weekly spend, impressions where they exist, and a settled revenue series rather than platform-reported conversions. Adstock shape chosen per channel by posterior comparison instead of by convention, Hill saturation curves per channel, and control variables that actually move your category: price and promotion calendar, product availability, competitor share of voice, organic search interest. Seasonality is where generic models fall over in the Gulf. Ramadan moves eleven days earlier each Gregorian year, so a Fourier term keyed to the Gregorian calendar smears it into noise. We key it to the Hijri calendar and model the pre-Ramadan build, the shifted daypart pattern, the Eid spike and the post-Eid trough as separate effects. The July and August residency exodus from the UAE, Dubai Shopping Festival and the Saudi school calendar each get their own term. Without these, a Gulf model attributes a seasonal collapse to whichever channel happened to cut spend that month, and someone acts on it. Calibration is the part we will not skip. Platform-reported return is not evidence and we do not use it to set priors. The ROI prior for each channel comes from a geo holdout or a platform lift study, either one you have already run or one we run first. We report the posterior interval rather than a point estimate. A channel whose ninety percent interval spans 1.1x to 4.6x is a channel you should test, not a channel you should scale, and we write it in those words. What you receive: The fitted model and the repository that produced it, response curves per channel with the saturation point marked and the uncertainty drawn, an allocation under your own constraints, and a written statement of which channels the model cannot see. We refit quarterly. We do not refit weekly; a model that moves every Monday is being read as a dashboard, and this is not a dashboard. What we refuse: We do not run mix modelling below roughly 150,000 dollars a month of media across at least three channels. We do not include a channel with no spend variance in the fitting window; it is unidentifiable, and including it produces a confident number that means nothing. We do not reconcile the model against your last-click report or force the two to agree — they measure different things, and forcing agreement destroys both. We do not accept a media plan as an input and return a justification for it. Who this is not for: Teams who need an answer inside a quarter with no test history to calibrate against. Single-channel advertisers. Anyone commissioning a mix model because attribution stopped flattering their programmatic line; the model will usually make that worse, and you should know it before you pay. Working languages are English and Arabic. Documentation is delivered in English; we present in either.