TimeHF introduces a novel approach for creating billion-scale time series models guided by human feedback, addressing critical limitations in existing forecasting systems.

• Combines large-scale neural networks with human preference learning for superior forecasting performance
• Achieves remarkable zero-shot generalization across diverse time series tasks
• Demonstrates significant accuracy improvements over current state-of-the-art methods
• Successfully deployed in supply chain management at JD.com with real business impact

This research represents a breakthrough for engineering teams working with complex time series data, offering a scalable architecture that can be applied across industries from manufacturing to retail forecasting.

TimeHF: Billion-Scale Time Series Models Guided by Human Feedback