Interactive model check · synthetic data

Same sales. Different story?

Two models look at the same marketing history. Switch one assumption and see how the classic MMM’s channel estimates change, then compare them with River-1 and the known simulated truth.

Who gets credit for sales?

Total estimated contribution across the same 166 weeks

Running River-1…

Dots show posterior means; whiskers show 80% posterior intervals. Diamonds show the known synthetic truth. Figures are sales units, not ROAS.

Loading model results…

Same observed sales

—

MMM in-sample sales RMSE, in sales units. A good fit does not establish correct attribution.

Known promotion contribution

—

Total sales generated by promotion in this synthetic world.

What both models were given

Observed sales and the selected MMM’s fitted sales. Shaded stripes mark promotion weeks.

This is a comparison on the observed fitting period, not a holdout forecast. Both MMM fits use identical channel spend and sales. River-1 also receives spend and promotion for the next week, as its inference API requires a query period; it does not see next-week sales.

One deliberately constructed synthetic example, not evidence of accuracy on real campaigns. The precomputed PyMC-Marketing MMMs use four MCMC chains; River-1 is an amortized model and returns separate posterior draws. Neither a match between models nor this single comparison establishes causality.