Mexican CPG manufacturer (salsas, anonymized)
Hierarchical Demand Forecasting for a Mexican CPG Manufacturer
Route-level demand forecasting across ~16 distribution routes — champion models selected against real history per route, delivered as bilingual apps the team runs itself.
- month forecast horizon, with 95% confidence intervals per client
- 1–12
- month forecast horizon, with 95% confidence intervals per client
- rows of retailer POS data (Walmart Retail Link + HEB sell-out) turned into basket analysis
- 2M+
- rows of retailer POS data (Walmart Retail Link + HEB sell-out) turned into basket analysis
The starting point
A Mexican CPG manufacturer in the salsas category sells through ~16 route-based distribution territories, plus major retailers. Planning ran on aggregate history and gut: no forward view at the route, SKU, or client level, and the sell-out data its retail partners provided — Walmart Retail Link, HEB — sat unexploited.
The method
We built hierarchical time-series forecasting on the internal route sales history — región → estado → ciudad → ruta → marca → SKU → cliente — with an automated champion-selection loop per route: ETS, ARIMA variants, seasonal naïve, and ensembles competed on held-out months, and the lowest-error model won production. (Prophet was tested and dropped — the classical models beat it on error.) Hierarchical reconciliation kept route-level forecasts summing to the totals leadership plans with. A separate workstream mined Walmart Retail Link 10-minute POS data (~2M rows) and HEB daily sell-out for market-basket analysis.
The result
The team now plans against a forecast that names its own error instead of a gut number. Deliverables, all bilingual (ES/EN): a working forecasting app — pick a route and client, get a 1–12-month forecast with 95% confidence intervals — a model-performance dashboard, auto-refreshing Spanish-language client reports, multi-sheet forecast workbooks, operational recommendations (safety stock near 20% of mean demand, top-client focus, route optimization), and basket-analysis decks from the retailer data.