Marketing Mix Modeling With Geo-Test Calibration

A mix model needs about two years of weekly history and, more to the point, spend that actually moved. Variance is the first thing I check: if three of your channels have sat within ten percent of the same weekly budget for two years, there is nothing for the model to learn from, and you will hear that in week one rather than week twelve. Under roughly two million dollars of annual media, the same money buys better answers as geo experiments. How I differ: A mix model is a prior-driven Bayesian object, not a regression that happens to be in fashion. Priors come from your own lift tests where they exist and from published elasticities where they do not, and both are written down before anything is fitted, so nobody can retrofit a belief to a result they liked. I do not calibrate to platform-reported ROAS. That teaches the model to reproduce the double counting you hired it to escape, and it makes the output unusable the moment somebody checks it against an experiment. The build: PyMC-Marketing, or Meridian where your team already runs it. Geometric adstock for channels with short memory, Weibull where the decay is genuinely delayed. Hill saturation curves per channel. Seasonality through Fourier terms rather than fifty-two dummy variables eating your degrees of freedom. Controls that matter in these categories specifically: promotional calendar and price index for wellness brands, weather and technician capacity for local services, and the sports and event calendar for igaming, where one fixture list explains more weekly variance than any media plan in the room. Calibration: Geo experiments run alongside the model, designed with GeoLift, and their measured lift enters as an informative prior on that channel's coefficient. This is the step most engagements skip, and it is the step that makes the output survive a budget meeting. Validation is a time-based holdout rather than random k-fold, and I report out-of-sample error on the last thirteen weeks instead of fit statistics computed on the whole period. What you get to decide with: Response curves per channel with credible intervals. An allocation under your real constraints, including contract minimums, agency commitments and the channel your CEO will not cut. And marginal return at current spend, which is the number that answers whether the next hundred thousand dollars goes to search or to connected TV. Every recommendation carries its interval, because a point estimate without one is a guess with better posture. The retainer: Monthly refit on new weeks, a written read on what moved and whether it moved beyond the noise band, one experiment designed each quarter to shore up the weakest coefficient in the model, and a standing session with your finance counterpart so the model and the plan use one definition of spend. Not included: No creative diagnostics and no media buying. No attribution model. Mix modeling and multi-touch answer different questions, and running them side by side to see which wins is a reliable way to spend two quarters arguing. No weekly refit either: refitting weekly invites decisions made on noise, which is the opposite of the reason to build this. Who this is not for: Businesses with one dominant channel; you do not need a model to tell you what you already know, and the model will mostly recover your own assumption. Teams who need an answer this month. And anyone who wants a single ROI figure per channel with nothing attached to it, which I decline to produce, because it will be quoted for two years after it stops being true.