How JD’s Data‑Driven Inventory Selection Boosted Fulfillment Efficiency

Background and Motivation

In e‑commerce, fast order fulfillment and high customer satisfaction depend on efficient supply‑chain management. JD Retail leverages a two‑tier distribution network of Regional Distribution Centers (RDC) and Front‑line Distribution Centers (FDC), where over 90% of self‑operated orders are delivered within 24 hours.

Problem Definition

The core challenges are (1) selecting a limited set of SKUs for each FDC (inventory selection) and (2) determining daily stock transfers from RDCs to FDCs (inventory allocation) to maximize order fulfillment while minimizing stock‑out loss and transfer costs.

Inventory Selection Model

The selection problem is formulated as an integer program that chooses up to K SKUs per FDC to maximize the number of orders fully satisfied locally. Direct solution is infeasible due to NP‑hardness and massive SKU scale.

Heuristic Algorithms

ML‑Top‑K : Predicts future SKU sales using a hybrid model (TCN + MLP) and selects the top‑ K SKUs by predicted volume.

Reverse‑Exclude : Iteratively removes the least impactful SKUs based on order frequency until only K remain, guaranteeing that remaining orders can be fulfilled.

Inventory Allocation Model

The allocation problem considers stochastic demand over multiple periods, a single RDC serving several FDCs, and aims to minimize total sales loss and transfer cost across a planning horizon T. Traditional linear programming is too slow for real‑time decisions.

End‑to‑End Allocation Algorithm

A multi‑task learning framework integrates demand forecasting, safety‑stock generation, and simulation‑based decision making. It uses RNNs, MLPs, and Transformers to output target inventory (TI) and safety stock (SS) levels, then simulates daily sales to evaluate loss and fulfillment metrics.

Experimental Evaluation

Using JD’s real data, the proposed algorithms were compared against existing methods:

ML‑Top‑K and Reverse‑Exclude improved local order satisfaction by 0.54 pp and 0.27 pp respectively; combined they achieved a 2.2 pp gain.

The end‑to‑end allocation algorithm raised FDC satisfaction to 58.59 % (vs. 57.55 % and 57.12 % for baseline methods) with comparable loss rates.

Deployment and Impact

The algorithms have been integrated into JD’s intelligent supply‑chain decision‑support system, covering all eight RDCs and their associated FDCs. Since launch, the system has reduced inventory holding and capital costs by tens of millions of RMB annually, cut transfer costs by over one hundred million RMB, increased inventory on‑hand rate by 0.85 %, and boosted local order satisfaction by 2.19 %.

Conclusion

The study demonstrates that data‑driven, optimization‑oriented techniques can substantially improve large‑scale e‑commerce supply‑chain performance, offering a blueprint for similar operations worldwide.