MMM dataset: 156 weeks to check before modelling.
Answer:
An MMM dataset is not a table of spend and sales. It must align time grain, economic units, media calendars, price, promotions, availability, distribution, events and stable definitions. Its data sheet determines what the model can distinguish and what it will confound.
Method and decision:
Reconcile currency, taxes, returns, zones, products, weeks and accounting calendars for every source. Profile minimum, maximum, coefficient of variation, zero weeks and correlations by channel. Document price changes, promotions, breaks, openings, competition and tests that affect results or calibrate the model. Hold out a recent period before estimation and prohibit its use in learned transformations. Authorise modelling only when the dataset provides variation, coverage, traceability and controls sufficient for the intended decision.
Worked example:
Paid social and video were increased together for 31 weeks and reduced together for 27 weeks. Their spend correlation is 0.94. Even with no missing values, the model can attribute the effect to either channel depending on priors. Individual contributions vary by 60%, while their sum is stable at 11%. Permit a grouped social-plus-video interpretation or create variation through a test; reject separate allocation based on these unstable coefficients.
Limits:
The Noroa dataset is simulated and represents no real market. Statistical quality does not guarantee all confounders were observed. A long series cannot repair a definition break or missing variation.