SERAF Enhances Time Series Forecasting with Multimodal Retrieval

Time series forecasting models often benefit from identifying and leveraging historical patterns. Recent advancements in Retrieval-Augmented Generation (RAG) have inspired approaches to retrieve relevant historical time series segments to improve predictions. However, relying solely on numerical similarity can be insufficient, especially when dealing with non-stationary data where underlying patterns might shift over time. To overcome this limitation, researchers propose SERAF (Semantics-Enhanced Retrieval-Augmented Time Series Forecasting). This multimodal framework performs a dual retrieval process: it considers both the time series' numerical similarity and the semantic similarity of their self-generated textual descriptions. This allows SERAF to retrieve two complementary sets of historical patterns and their corresponding future outcomes. These retrieved numerical and semantic insights are then selectively and jointly used to guide future predictions. Experiments across seven real-world datasets demonstrate that SERAF effectively bridges the gap between numerical and semantic views of time series, outperforming state-of-the-art baselines.