Marketing-mix model and attribution read of a synthetic campaign dataset

Harbour & Hearth sells bedding, cookware and lighting online in the US, UK and Canada. It spends about US$9.6 million a year on media: paid search, paid social, YouTube, affiliates, podcasts, TV in the US, and email. Revenue is about US$58 million.

Every platform claims credit. Meta says it drove 41% of revenue, Google says 46%, and the affiliates network says 18%. Together they claim more than 100%. The last-click report in analytics tells a different story again, giving almost everything to branded search and email. The CFO, Daniel Hughes, has said he will cut the budget by 15% next year unless marketing can show which spend actually adds sales.

The head of growth, Priya Menon, wants an independent read using a marketing-mix model (MMM). Real data is confidential, so you will: - simulate 104 weeks of data for three countries with seasonality, promotions, price changes and seven channels, where you set the true effect, carryover and saturation of each channel; - produce the attribution numbers that platforms and last-click would report from that data, to show how they mislead; - engineer adstock and saturation features; - fit an MMM in Python (regression, Bayesian or a library such as Robyn or PyMC-Marketing); - validate it: does it recover the true effects you set, and how well does it predict a held-out period; - use the model to recommend a budget shift and explain its uncertainty.

Done looks like a reproducible notebook, a validation report and a five-page memo to Priya and Daniel. They will judge it on whether the model is honest about uncertainty and whether the recommendation would survive a sceptical CFO.