Decentralized retail networks struggle with two main problems: predicting customer demand accurately and sharing customer data while keeping personal information private, especially when stores and platforms are in different locations.  Federated learning is a new approach that lets teams train models to predict things without sharing actual data. Stochastic linear programming is a method used to make smart decisions about inventory and restocking when demand is uncertain.  This paper suggests a combined system where forecasts of customer demand, made using federated learning at each store, are used in a two-step planning model for managing inventory in a retail network that has multiple levels of distribution.  Building on recent advancements in federated learning for demand forecasting, supply chain planning under stochastic demand, and decentralized FL architectures, we formalize the interaction between the learning layer and optimization layer and demonstrate the approach on a synthetic yet realistic dataset calibrated to patterns reported in recent literature.  The experimental results show that using the proposed FL SLP integration leads to a reduction of 18 to 25 percent in average stock outs and 8 to 12 percent in total network costs compared to non-federated baselines that use either local models or centralized training with data pooling.  A statistical analysis using ANOVA confirms that these improvements are significant at the 5% level.  We end by talking about important things to consider when putting decentralized retail networks into use in the real world and by pointing out where future research could go.