Cascading Demand Forecasting for more than 140 stores
Stockouts and shrinkage across a network of 140 supermarkets.
Managing the weekly supply of thousands of items is a critical challenge. Underestimating causes stockouts and lost sales, while overestimating leads to severe financial losses from expired fresh produce. On top of that, promotional campaigns suddenly disrupt normal consumption and cause cannibalization between sister brands.
A cascading predictive engine built into supply planning.
Development of a two-stage predictive engine: a base model (Prophet) estimates the baseline trend and seasonality by store and item, while an advanced model (XGBoost) learns and corrects the non-linear impact of promotions and commercial cannibalization, generating pinpoint-accurate orders.
Measurable Impact
The engineering behind the impact
AI Logic and Models
Hybrid cascading model (Two-Stage Ensemble):
- Prophet: Isolates the pure time trend and baseline seasonality.
- XGBoost: Captures the non-linear complexity of promotional impact and cannibalization.
Data Ingested
- - Daily sales by POS.
- - Promotions calendar.
- - Holidays and seasonality.
- - Cannibalization matrix.
Architecture & Stack
A pipeline developed in Python and orchestrated in the cloud. It generates weekly orders for the 140 stores and loads them into the supply system 100% autonomously.
Want to replicate this level of efficiency in your operation?
We don't rely on theory. We analyze your current systems and design a roadmap to unlock the operating margin you're losing right now.