Marketing Mix Modeling: specify the decision before estimating the model.
Answer:
A useful Marketing Mix Model does not start with a spend table or an algorithm. It starts with a decision: which budget, levers, range, horizon and constraints need to be informed? That specification determines the economic variable, grain, transformations, confounders, calibration experiments, validation and admissible uncertainty. Without a use contract, a model can predict history well yet be dangerous for allocation.
Method and decision:
Define budget, channels, units, periods, scenarios, constraints, update frequency and prohibited uses, notably extrapolation outside support and overly fine creative interpretation. Audit variation, correlations, breaks, missing data, price, promotion, distribution, seasonality and events, then build an identifiability matrix before estimating. Compare specifications, priors and windows; inspect residuals, contribution, adstock, saturation and parameters, rejecting economically impossible compensation. Calibrate against compatible experiments or quasi-experiments and publish intervals, use domain, version log and divergence from evidence. Allow MMM allocation only when specification, identifiability, decision validation and uncertainty meet the approved contract.
Worked example:
An MMM covering 156 weeks predicts sales with 7% MAPE and recommends moving €400k from video to search. Video and branded search correlate at 0.82; video spend has never been below €250k per quarter. Under specification A, video ROAS is 0.9 and search 3.4. Under a longer adstock prior and geographic calibration, video is 1.8 ± 0.7 and search 1.9 ± 0.5. Cutting €400k brings video to €120k, outside support. A bounded −€120k scenario has €96k expected gain, interval −€18k to +€205k, and €130k maximum regret. MAPE therefore does not validate the initial allocation. The contract turns it into a €120k guarded pilot, not automatic optimisation.
Limits:
Aggregate data separates synchronised activities poorly. Priors can stabilise estimates without creating new information. Structural change reduces the transportability of historical curves.