Time Series Forecasting Algorithm System in E-commerce: Practice and Applications at NetEase Yanxuan

This article introduces a comprehensive time series forecasting algorithm system developed by NetEase Yanxuan for e-commerce scenarios, implementing a complete "data-model-service" pipeline.

Background: Time series data refers to data describing changes in one or more features over time, such as daily product sales or stock prices. These algorithms are widely used in e-commerce for predicting future product sales to determine replenishment quantities and forecasting APP traffic for search and recommendation strategies.

Applications at Yanxuan: The system supports multiple business scenarios including: (1) Product sales forecasting for 100,000+ SKUs with 360-day predictions; (2) Inter-warehouse transfer volume forecasting by province and warehouse; (3) Warehouse order volume forecasting for 20,000+ warehouses with 60-day predictions including cold-start for new warehouses; (4) Hourly advertising traffic forecasting for 100 billion daily requests across hundreds of media platforms; (5) APP traffic forecasting for 7 major traffic entrances including search, recommendation, and product detail pages.

System Architecture: The system consists of three main components: (1) Data features including user data (attributes, behaviors), product data (characteristics, sales), business data (warehousing, marketing), and external data (holidays, weather); (2) Algorithm models spanning classical statistical models (Linear Regression, ARIMA, Holt-Winters, Prophet), traditional machine learning (XGBoost, LightGBM), and deep learning (DeepTCN, MASS, DeepAR, LSTNet, TFT, Informer); (3) Algorithm services built on Tornado framework with unified interfaces, distributed computing support, and the open-sourced Typhoon SDK.

Key Insights: The article emphasizes that model complexity does not guarantee better predictions. Traditional ML models like XGBoost currently achieve the best results in practice. Additionally, prediction interpretability is as important as accuracy for most business scenarios. The fundamental limitations include: (1) Future events may differ from historical patterns (e.g., live streaming commerce); (2) Unpredictable events like pandemics or promotions significantly impact forecasts but cannot be predicted by algorithms.