Wang, Shiyu (ByteDance), Zhong, Xinyue (ByteDance), Li, Jiawei (Uppsala University), Mo, Baichuan (MIT), Ye, Zhou (ByteDance), Jin, Ming (Griffith University)

SCOPE-MoE: Supply Chain Forecasting with a Pretrained MoE-Based Large Time Series Model in E-Commerce

Scheduled for presentation during the Regular Session "S39b-Data-Driven Optimization in Intelligent Transportation Systems" (FR-EA-T39), Friday, November 21, 2025, 13:30−13:50, Coolangata 3

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Abstract

Rapid expansion of E-commerce intensified demands for accurate and scalable Supply Chain Management (SCM). The core of the SCM is demand and logistics flows forecasting. However, traditional forecasting models struggle with challenges such as data sparsity, long-tail distributions, and complex business scenarios. We propose SCOPE-MoE, an mixture-of-experts (MoE) transformer pre-trained on curated and balanced datasets from an E-commerce platform to address these limitations. Through strategic data sampling and preprocessing, SCOPE-MoE achieves robust performance even with limited historical data. SCOPE-MoE achieves zero-shot forecasting across four key time series types and consistently outperforms fully retrained baselines. In a 10-month A/B deployment at a leading E-commerce company, it reduced MAPE by 39.4%, improved order fulfillment by 9%, and cut inventory costs by 12%. Fine-tuning required only 45 minutes per epoch, demonstrating high computational efficiency. To our knowledge, this is the first real-world application of a large time series model at scale in E-commerce SCM, highlighting its practical and technical significance.