Explainable AI Forecasting and End-to-End Inventory Management in JD's Smart Supply Chain

On February 14, Gartner announced the 2024 Gartner Power of the Profession Supply Chain Awards, with JD Group winning the Supply Chain Technology Innovation Award, the only Asian company among finalists that also included Google, Cisco, MTN Group, and Allina Health.

JD's retail technology team continues to explore innovations, applying end‑to‑end inventory management and explainable AI to achieve faster inventory turnover and more efficient supply‑chain decision‑making.

1. Super‑hard problem in supply‑chain decision: How to improve prediction accuracy while ensuring explainability?

Sales forecasting is a key component of supply‑chain decisions; its accuracy directly impacts inventory control, financing, production planning, and market strategy. Traditional statistical time‑series models are giving way to more complex machine‑learning algorithms that can handle higher‑dimensional data and capture nonlinear relationships.

Forecasting has evolved through three stages: traditional statistical methods (e.g., ETS, ARIMA), machine‑learning methods (e.g., XGBoost, LSTM, Transformer), and hybrid algorithms that combine both.

Current machine‑learning models are often black boxes, lacking interpretability, which hinders their adoption in supply‑chain practice. Hybrid algorithms attempt to improve both accuracy and interpretability but suffer from low scenario matching and reliance on statistical assumptions.

For JD's supply chain, replenishment decisions require not only high accuracy but also high explainability to gain business trust, especially for head‑item stock‑outs.

The main technical challenges are: (1) Existing algorithms under high explainability constraints have poor accuracy; (2) Existing hybrid algorithms have weak global fitting ability and low compatibility with supply‑chain scenarios.

JD's intelligent supply‑chain team proposes a new explainable forecasting technique—a hybrid algorithm that builds a universal explainable framework, dramatically improving both interpretability and accuracy.

1. Explainable prediction process and results greatly increase user trust

The new technique outputs predictions as a composition of demand factors (baseline, promotion, marketing, etc.) and uses causal inference on large‑scale order and promotion data to trace how each factor influences the forecast, thereby guiding more accurate replenishment decisions.

2. A generic algorithm combining decomposition and ML (W‑R algorithm)

The W‑R algorithm creates an interpretable additive model by weighting a decomposition‑based component with a deep residual network, preserving time‑series interpretability while leveraging global information for higher accuracy. It operates in two stages: an initial decomposition module (e.g., prediction = baseline + promotion + marketing) and a ML adjustment module that refines component weights and residuals.

Future work will iterate on full‑process explainability, automatic attribution diagnosis, and plan explainability.

2. End‑to‑End Inventory Management Strategy and Model Design

Inventory management is critical; inaccurate decisions cause stock‑outs or excess inventory, leading to lost sales or high storage costs.

Traditional “predict‑then‑optimize” (PTO) separates forecasting from optimization, causing information loss and sub‑optimal decisions, especially under JD's massive SKU count and volatile demand.

JD's team introduces an end‑to‑end inventory management technique based on a Multi‑Quantile Recurrent Neural Network (MQRNN) that directly predicts optimal replenishment quantities from historical sales, procurement, and supplier performance data.

The model first estimates optimal replenishment per order time using a dynamic‑programming‑based decision model, then builds feature libraries from sales, lead‑time, order cycle, and initial inventory, trains the MQRNN, and finally predicts sales, lead‑time, and order quantities to output the replenishment decision.

The input consists of five feature groups (demand‑forecast features, product attributes, supplier lead‑time, inventory cycle, initial stock level) and outputs three items: the final replenishment decision, demand forecast, and supplier lead‑time forecast.

Since deployment, these technologies have improved forecast accuracy by 7%, increased in‑stock rate by 2%, reduced inventory turnover time by nearly two days, and saved billions of yuan in holding costs, achieving over 85% automation in the auto‑replenishment system.

JD, serving nearly 600 million active users with over 10 million SKUs, leverages these advanced supply‑chain technologies to maintain a 95%+ same‑day or next‑day delivery rate, a 30‑day average inventory turnover, and a 97%+ in‑stock rate.

For detailed technical information, refer to the papers: https://doi.org/10.1287/mnsc.2022.4564 and https://arxiv.org/abs/2212.06620.