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

In October 2024, JD Retail's supply‑chain technology team won the Daniel H. Wagner Prize for innovative inventory selection and allocation techniques, highlighting the impact of data‑driven optimization on e‑commerce fulfillment.

The paper first describes the two‑tier distribution network (regional distribution centers – RDCs and front‑mile distribution centers – FDCs) and the challenges of selecting SKUs for limited‑capacity FDCs while minimizing sales loss and transfer costs. It formulates the selection problem as an integer program with binary variables xi (SKU inclusion) and yo (order fulfillment), subject to capacity and fulfillment constraints.

Two heuristic selection algorithms are introduced:

ML‑Top‑K : predicts SKU sales using a hybrid model that combines a time‑convolutional network (TCN) and a multilayer perceptron (MLP) with promotion features, then selects the top‑K SKUs by predicted volume.

Reverse‑Exclude : iteratively removes low‑impact SKUs based on order influence fi until only K SKUs remain, ensuring remaining orders can be fully satisfied.

The allocation problem is modeled as a multi‑period stochastic inventory transfer task. An end‑to‑end learning framework integrates demand prediction, safety‑stock/target‑stock generation, and simulation modules. The model uses multi‑task learning with RNN/Transformer components to output optimal transfer quantities for each RDC‑FDC pair while minimizing total sales loss and transfer cost.

Experimental results on real JD data show that the ML‑Top‑K and Reverse‑Exclude algorithms improve local order fulfillment rates by 0.54 pp and 0.27 pp respectively, and a combined approach yields a 2.2 pp gain. The end‑to‑end allocation algorithm achieves a 58.59 % FDC fulfillment rate, surpassing baseline methods by over 1 pp while keeping loss rates comparable.

Deployed in production across all eight JD RDCs (covering 5‑13 FDCs each), the system reduces inventory holding and capital costs by tens of millions of RMB annually, saves over a hundred million RMB in transfer costs, raises inventory on‑hand rate by 0.85 %, and improves local order fulfillment by 2.19 % and 211‑order share by 1.44 %.

These results demonstrate the practical value of combining machine learning, operations research, and simulation for large‑scale supply‑chain optimization.