Demand Patterns Choose Demand Models: Agentic Demand Forecasting with Amazon Connect Decisions

Existing demand forecasting systems require rebuilding in response to business changes, and valuable on-the-ground adjustment data often fails to be reflected in the systems. To address this, Amazon launched Amazon Connect Decisions (ACD), applying 30 years of accumulated supply chain expertise.

ACD leverages Agentic AI to diagnose demand patterns per SKU and dynamically select the appropriate forecasting model. In the diagnosis phase, it applies the SBC classification method based on ADI (Average Demand Interval) and CV² (Squared Coefficient of Variation) to categorize demand into four patterns: Smooth, Intermittent, Erratic, and Lumpy. Particularly in highly seasonal industries like fashion, considering the Product Lifecycle (trend shifts) is essential to improve forecast accuracy.

In the measurement phase, it uses WAPE (error magnitude) and Bias (error direction) as a pair instead of MAPE. This prevents division-by-zero errors and distinguishes between systematic and random errors, enabling accurate diagnosis of model bias.

In the prescription phase, it applies techniques suited to the pattern (such as Croston-type methods, ETS/ARIMA, etc.) and incorporates external signals like promotions. ACD automatically orchestrates this entire process in every cycle to generate the optimal forecast at the item level.