Go to ICLR 2025 Conference homepageTF-score: Time-series Forecasting using score-based diffusion model
Diffusion models have emerged as powerful generative models, capable of synthesizing high-quality images by capturing complex underlying patterns. Building on this success, these models have been adapted for time-series forecasting, a domain characterized by intricate temporal dependencies. However, most existing works have focused primarily on empirical performance without sufficient theoretical exploration. In this paper, we address this gap by introducing a generalized loss function within the diffusion-based forecasting framework. Leveraging this foundation, we introduce TF-score, a score-based diffusion model designed to capture the interdependencies between historical data and future predictions. Extensive experiments across six benchmark datasets show that TF-score consistently surpasses leading baselines, including prior diffusion-based models. Furthermore, we extend existing guidance sampling strategies into a our score-based formulation, achieving performance gains across multiple datasets while providing a detailed analysis of the trade-offs involved.