Amazon moves demand forecasting from specialized models toward a reusable forecasting foundation model

Amazon researchers Boris Oreshkin and Dmitry Efimov are presenting at ISF 2026 on the transition from specialized MQCNN demand-forecasting models toward a PFN-based foundation model that can support multiple forecasting use cases. A reusable forecasting foundation model could reduce the need to maintain separate models by product, market or planning use case, but it also raises questions about transferability, bias, local exceptions and governance. Planning teams should validate performance by demand pattern and horizon, compare results with local baselines, measure Forecast Value Add, preserve planner overrides and prevent standardization from hiding material regional or category-specific differences.