How AI Is Transforming Supply Chains: Inside JD.com's Billion‑Scale Time‑Series Model

At the "Smart Canal, Smart Future" 2025 AI Innovation & Entrepreneurship Conference in Hangzhou, Professor Shen Zuojun, an academician of the Hong Kong Engineering Academy and chief scientist of JD.com’s retail supply chain, co‑chaired the AI+Smart Logistics and Supply Chain session and delivered a keynote on how AI technology can upgrade the supply‑chain industry.

He emphasized that building an intelligent decision‑making hub requires dynamic demand forecasting for precise resource matching, AI‑based risk perception for resilient response, and cross‑chain collaboration to break industry silos, forming a perception‑decision‑execution closed loop that redefines efficiency and drives a shift from linear cooperation to networked intelligent collaboration.

JD.com’s retail supply chain manages over 10 million SKUs across more than 1 500 smart warehouses, achieving over 90% same‑day or next‑day delivery for self‑operated orders, a performance enabled by an integrated intelligent inventory platform that combines planning coordination, demand forecasting, intelligent decision‑making, and risk perception.

The core of this platform is a precise time‑series forecasting capability. JD’s proprietary large‑scale time‑series model reaches industry‑leading accuracy on multi‑dimensional feature fusion, supporting intelligent product selection, dynamic allocation, and fulfillment optimization, thereby improving inventory turnover and creating a prediction‑decision‑execution‑feedback learning loop.

Traditional forecasting methods such as ARIMA, Prophet, LSTM, and TCN struggle with complex patterns and zero‑shot generalization, while large language models adapted for time series (e.g., GPT‑4TS, TimesFM) have yet to achieve breakthroughs due to scarce high‑quality datasets and RLHF adaptation challenges.

JD’s algorithm team responded by building the industry’s first billion‑parameter pure time‑series model, incorporating multi‑scale feature fusion and adaptive temporal attention, and introducing weak‑supervised pre‑training tasks tailored to supply‑chain scenarios. The model surpasses state‑of‑the‑art baselines, especially in zero‑sample cross‑domain prediction.

To train the model, the team assembled a 1.5‑billion‑sample high‑quality dataset and devised time‑series splitting, data‑ratio, and synthetic data construction methods. They introduced the PCTLM architecture, which processes data in patches with enhanced projection to capture cross‑patch information, and a grouped time‑position encoding attention mechanism. A novel RLHF framework (TPO) further improves prediction accuracy and generalization.

Beyond demand forecasting, JD built a data‑driven intelligent selection system using ML‑Top‑K, Reverse‑Exclude, and Hybrid Selection algorithms, raising local order fulfillment rates by 2.19% and high‑value order proportions by 1.44%.

The end‑to‑end allocation algorithm integrates demand forecasts, multi‑objective optimization, and simulation to generate high‑stock‑availability, low‑cost allocation plans, handling billions of variables in real time. Since deployment, it has cut warehouse holding costs by tens of millions of yuan annually, saved over a hundred million yuan in allocation costs, and increased stock availability by 0.85%.

These advances illustrate how AI is shifting supply‑chain value from pure efficiency gains to global resource optimization, reshaping business models and user experiences. JD.com is further collaborating with Tsinghua University to deepen research‑industry integration, focusing on intelligent prediction, dynamic scheduling, and multi‑level inventory coordination, aiming to open its innovations to the broader industry and drive high‑quality economic development.