Traditional demand forecasting treats each SKU as an independent time-series, often leading to chronic overstocking and stockouts. The article argues that supply chains are inherently networked, where a demand shift for one product affects others through shared plants and storage locations. This structural insight is missing from conventional models.

Using a real FMCG dataset with 40 SKUs and 9 plants, the author demonstrates a Graph Neural Network (GNN) approach. Unlike isolated models, a GraphSAGE architecture lets each SKU aggregate signals from its neighbors, capturing both spatial relationships and temporal dynamics. This spatio-temporal modeling fundamentally changes how demand signals propagate through the network.

The experiment evaluates forecasts using WAPE (Weighted Absolute Percentage Error) and Bias. A simple naïve baseline produced a WAPE of 0.86, meaning 86% of volume was misallocated, despite zero bias. The GNN model aims to lower WAPE by understanding operational context, moving beyond accuracy to actionable, financially sound forecasts for supply-chain planning.