In supply chain demand forecasting, data is scattered and stored in various enterprise nodes, forming data islands. The independent operation of prediction algorithms by each node is not coordinated, resulting in algorithm fragmentation. The interaction between the two reinforces the amplification of prediction bias, the intensification of the bullwhip effect, and resource mismatch. Traditional centralized modeling is difficult to implement due to data privacy and commercial confidentiality constraints. Federated learning, as a distributed machine learning paradigm, allows models to train collaboratively without moving data, breaking data silos through cross-node parameter aggregation, and unifying algorithm logic and alleviating algorithm fragmentation through iterative mechanisms of global and local models. Embedded differential privacy and security aggregation technologies, while protecting sensitive enterprise information, achieve joint modeling, simulation validation shows that this path reduces prediction error by 22% to 30% within a controllable accuracy loss range, thereby providing a decentralized solution for supply chain demand forecasting that balances both security and collaboration.