Meridian
Bayesian Marketing Mix Modeling for Retail
A Bayesian MMM in production, built on Google's Meridian and fitted on 137 weeks of real point-of-sale revenue.
- Brands converged
- 8/8
- R² on the largest brand
- 0,98
- Weeks of real revenue
- 137
The problem
Without causal measurement, media budget gets split by history and instinct. Platform dashboards attribute sales to themselves, and nobody knows how much of the revenue was incremental.
How it works
One pipeline per brand: SAP S/4HANA and Google Ads flow into BigQuery, get transformed into weekly series of spend, revenue and controls, and feed eight independent Bayesian models — one per brand — across 12 media channels and 16 regions of Chile. All eight converged, and they retrain weekly on GPU.
What it does
- 01
One model per brand
Eight independent models instead of one group average: each brand responds differently to each channel.
- 02
Real revenue, not proxies
Fitted on 137 weeks of point-of-sale revenue from SAP — not on platform-reported conversions.
- 03
Weekly retraining
Models retrain weekly on GPU, so the media response is never last quarter's.
Related projects
- AndreaAI Consumer Simulation for Marketing
- DelphineML Demand Forecasting for Footwear Retail
- CerebroCompetitive Pricing Intelligence for Retail
Last updated: August 2026
Run a retail operation and recognize the problem? The advisory side of this work lives at Menlo & Oak — and you can always write to me directly.