Strategic Inventory Placement in Large-Scale Multi-Echelon Networks : A Sampling-Based Approach

The placement and sizing of safety stocks in supply chains pose a challenging optimization problem. State-of-the-art multi-echelon inventory optimization models, such as the guaranteed-service approach, are non-linear and depend on statistical, time-series-based approaches that require distributional and parametric assumptions. We propose a data-driven, non-parametric and distribution-free approach for safety stock planning in multi-echelon inventory networks that utilizes historical demand and feature data. We extend data-driven optimization in inventory control from newsvendor models to multi-period and multi-echelon problems. Our model accommodates general, acyclic multi-echelon networks and simultaneously determines safety stock allocation and sizing by setting cost-optimal base stocks for all stages under consideration of service requirements. By developing a mixed-integer programming formulation and a Benders decomposition method, we offer a novel methodological approach to a well-studied problem that can be solved with commercial mathematical programming solvers. We also provide a probabilistic analysis of the data-driven performance relative to an oracle solution when sample data is limited. We show that the mixed integer programming approach is scalable by solving 38 large-scale supply chain benchmark networks with assembly, distribution, and general structures