Time series forecasting is crucial in domains such as energy and transportation, where non-stationary dynamics are influenced by external events often described in textual form. Traditional forecasting models typically rely on a single modality, limiting their ability to capture complex contextual interactions between time series data and external events. EventTSF introduces an autoregressive generation framework that combines historical time series with textual event data to improve forecasting performance. The model tackles three main challenges: fine-grained synchronization between discrete textual events and continuous time series, temporal uncertainty from textual semantics, and misalignment between textual embeddings and temporal patterns at multiple resolutions. EventTSF employs autoregressive diffusion with flow matching to model temporal-event interactions, adaptively controlling flow matching timesteps based on event semantics to handle uncertainty. The denoising component uses a multimodal U-shaped diffusion transformer to fuse temporal and textual data efficiently across different resolutions. Extensive experiments on eight synthetic and real-world datasets demonstrate that EventTSF outperforms twelve baseline methods, achieving a 10.7% improvement in forecasting accuracy and 1.13 times faster training. This advancement highlights the potential of multimodal approaches in enhancing non-stationary time series forecasting by leveraging event-aware information. Future applications could extend to other domains where event-driven dynamics are critical, and further research may explore refining multimodal alignment techniques to boost performance even more.
👉 Pročitaj original: arXiv AI Papers
