Data‑Driven Inventory Selection and Allocation Algorithms for JD Retail Supply Chain

In October 2024 the JD Retail Supply‑Chain Technology team received the prestigious Daniel H. Wagner Prize for its innovative inventory selection and allocation techniques, highlighting the impact of data‑driven optimization in e‑commerce logistics.

The paper describes the two‑tier distribution network composed of Regional Distribution Centers (RDC) and Front‑line Distribution Centers (FDC). FDCs have limited capacity and must be stocked with a carefully chosen set of SKUs to maximize local order fulfillment while minimizing stock‑out loss and transfer cost.

Two sub‑problems are formulated: (1) the SKU selection problem, which decides which products to store in each FDC under a capacity constraint, and (2) the daily inventory transfer problem from RDC to FDC, which seeks the optimal replenishment quantities over multiple periods.

For the selection problem the authors propose two heuristics: ML‑Top‑K , which predicts SKU demand with a hybrid TCN‑MLP model and picks the top‑K items, and Reverse‑Exclude , which iteratively eliminates low‑impact SKUs until the capacity limit is reached. Both methods are shown to improve local order‑satisfaction rates.

The allocation problem is tackled with a novel end‑to‑end learning framework that jointly performs demand forecasting, safety‑stock/target‑stock generation, and simulation‑based decision making. The framework uses multi‑task learning, recursive neural networks, and a simulation module to minimize total sales loss and transfer cost while maintaining interpretability.

Extensive experiments on real JD data demonstrate that the combined selection algorithms raise local order‑satisfaction by up to 0.54 percentage points (≈100 k orders per day) and that the end‑to‑end allocation algorithm achieves a 58.59 % FDC fulfillment rate, surpassing baseline parameter‑search (57.55 %) and linear‑programming (57.12 %) methods.

After deployment, the system now serves all eight RDCs (each covering 5‑13 FDCs). Reported operational gains include a reduction of inventory holding and capital costs by tens of millions of RMB annually, a saving of over 100 million RMB in transfer costs, a 0.85 % increase in inventory on‑hand rate, and a 2.19 % rise in local order‑satisfaction, benefiting tens of millions of orders.

In conclusion, the paper presents a comprehensive, data‑driven optimization pipeline for large‑scale supply‑chain management, demonstrating the practical value of combining machine‑learning prediction, operations‑research modeling, and simulation in real‑world e‑commerce environments.