Nexus: Multi-Agent Framework for Time Series Forecasting with LLMs

A recent study presents Nexus, a multi-agent framework designed for forecasting time series by incorporating unstructured contextual data, including news and events. While traditional Time Series Foundation Models (TSFMs) are proficient in recognizing numerical patterns, they overlook important textual signals from the real world. On the other hand, large language models (LLMs) used for zero-shot forecasting display variable effectiveness. Nexus breaks down the prediction process into distinct phases: identifying macro and micro temporal variations and merging contextual data before producing a final forecast. This method enables adaptation to seasonal trends and unpredictable, event-driven information without depending on external statistical references or uniform prompting. The research, showcasing the significant capabilities of current LLMs, is available on arXiv with the identifier 2605.14389.