A practical approach to replenishment optimization with extended (R, s, Q) policy and probabilistic models

Area of Science:

• Operations Research
• E-commerce Logistics
• Supply Chain Management

Background:

• Effective inventory management is critical in e-commerce due to fluctuating demand and complex fulfillment networks.
• Existing optimization models often use simplified demand assumptions, failing to capture real-world uncertainties.
• Bridging predictive modeling with practical replenishment decisions is a key challenge.

Purpose of the Study:

• To integrate probabilistic demand forecasting with advanced inventory policy optimization.
• To extend the classical [Formula: see text] policy for distributed fulfillment and seasonal assortments.
• To develop a practical tool for e-commerce inventory optimization.

Main Methods:

• Developed the Zalando E-commerce Operating System (ZEOS) Inventory Optimization Tool.
• Unified one-shot inventory policy optimization with probabilistic gradient-boosting models (LightGBM).
• Adapted the [Formula: see text] policy for distributed networks and seasonal assortments.

Main Results:

• Achieved significant uplift in Gross Merchandise Value (GMV) and GMV after fulfillment costs compared to human and classical baselines.
• Maintained high operational availability ([Formula: see text]) and demand fill rate ([Formula: see text]).
• Probabilistic forecasts with percentile objectives and a 12-week horizon showed optimal performance.

Conclusions:

• The ZEOS tool effectively bridges probabilistic forecasting and policy optimization for e-commerce.
• The approach enhances efficiency, reduces costs, and improves profitability in dynamic retail.
• This research offers a pioneering solution for complex inventory management challenges.