EventTSF: Event-Aware Non-Stationary Time Series Forecasting

Researchers have introduced EventTSF, a novel autoregressive diffusion framework designed to enhance non-stationary time series forecasting by integrating textual event data. Traditional forecasting models often rely on single-modality inputs, limiting their ability to capture complex dynamics in sectors like energy and transportation where external events significantly influence trends. EventTSF addresses two primary challenges: bridging the gap between discrete external events and continuous time series, and mitigating imbalanced denoising difficulties caused by uniform diffusion timesteps. The model employs an event-aware flow-matching timestep conditioned on event semantics to enable fine-grained multimodal interactions. Extensive experiments conducted on seven synthetic and real-world datasets demonstrate that EventTSF significantly outperforms twelve existing baselines. It achieves average performance gains of 41.3% in probabilistic forecasting and 27.5% in deterministic forecasting across all evaluation metrics. This advancement highlights the potential of combining natural language processing with time series analysis to improve predictive accuracy in dynamic environments.