SCENARIODIFF: Multimodal Time Series Forecasting with Scenario Guidance

Researchers have developed SCENARIODIFF, a hierarchical contextual reasoning framework designed for multimodal time series forecasting. This framework addresses the challenge of incorporating textual information, such as news articles or reports, to improve forecasts, particularly when future trends are influenced by external events not yet reflected in historical numerical data. Existing multimodal forecasting methods often either directly ask large language models (LLMs) to predict values or implicitly fuse text and time series, making it difficult to interpret or control the influence of contextual information. SCENARIODIFF overcomes this by structuring contextual data into three distinct levels: a Historical Context Agent extracts evidence, a Scenario Agent generates qualitative descriptions of future scenarios, and an Anchor Guidance Agent provides sparse anchor points for event-relevant future periods. These structured signals then condition a Multimodal Diffusion Transformer. An additional component, Anchor Blended Sampling, refines generated trajectories locally without requiring model retraining. Experiments on the Time-MMD benchmark demonstrate that SCENARIODIFF is particularly effective in domains driven by specific events, highlighting the value of explicit, hierarchical scenario guidance for more accurate multimodal time series forecasting.