Nexus: An Agentic Framework for Time Series Forecasting
Time series forecasting is not just numerical extrapolation, but often requires reasoning with unstructured contextual data such as news or events. While specialized Time Series Foundation Models (TSFMs) excel at forecasting based on numerical patterns, they remain unaware of real-world textual signals. Conversely, while LLMs are emerging as zero-shot forecasters, their performance remains uneven across domains and contextual grounding. To bridge this gap, we introduce Nexus, a multi-agent forecasting framework that decomposes prediction into specialized stages: isolating macro-level and micro-level temporal fluctuations, and integrating contextual information when available before synthesizing a final forecast. This decomposition enables Nexus to adapt from seasonal signals to volatile, event-driven information without relying on external statistical anchors or monolithic prompting.
Rather than relying on a single monolithic model to directly approximate the forecasting mapping, Nexus decomposes the forecasting task into three distinct, logical stages: Contextualization, Dual-Resolution Forecast Outlook Generation, and Forecast Synthesis and Calibration. By systematically breaking down the problem, the framework first structures the raw multimodal context, then projects future outlooks reasonings across different forecast resolutions, and finally utilizes a Forecast Synthesizer Agent to merge these perspectives into a final forecast. This multi-agent system allows Nexus to dynamically synthesize qualitative insights with historical trends, producing robust numerical predictions as well as explicit interpretable reasoning.
We show that current-generation LLMs possess stronger intrinsic forecasting ability than previously recognized, depending critically on how numerical and contextual reasoning are organized. The utility of LLM backbones for time series remains contested: systematic ablations of several LLM-based forecasting methods show that removing or replacing the LLM component often does not degrade performance, suggesting that much of the gain may come from patching, attention, or task-specific adaptation rather than linguistic pretraining itself. Nexus is positioned within this emerging direction: rather than relying solely on one-shot numerical extrapolation, LLM-based forecasting systems can benefit from explicit reasoning from state-of-the-art LLMs, diagnostic feedback, and iterative calibration over prior temporal evidence.
In this paper, we introduced Nexus, a novel multi-agent framework designed to tackle the complex challenge of multimodal contextual time-series forecasting. By decomposing forecasting into structured stages — Contextualization, Dual-Resolution Forecast Outlook Generation, and Forecast Synthesis and Calibration — Nexus helps manage the complexity of processing long sequences of numerical data combined with unstructured text. Nexus dynamically synthesizes broad macro-level trajectories with granular, event-driven micro-level catalysts, producing highly accurate numerical predictions alongside explicit reasoning. Through rigorous zero-shot evaluations on real-world Zillow real estate and Stock market datasets, Nexus consistently outperforms both time-series foundation model (TimesFM-2.5) and strong Chain-of-Thought LLM baselines.
Lines of inquiry this paper opens 24
Research framings built by reading the notes related to this paper — the questions it feeds into.
What makes understanding genuine rather than merely producing correct outputs? What capabilities and tradeoffs distinguish diffusion from autoregressive language models? How do different adaptation mechanisms compare on efficiency and capability preservation? How should agents learn and use user preferences during conversation? How do language models organize knowledge across layers and representations? What planning tasks require structured scaffolding beyond LLM generation? Do foundation models build genuine world models or task-specific shortcuts? Does reasoning learned in one domain transfer to other fields? How can iterative computation overcome fundamental limitations in model capability? How does training data composition shape what models learn? Why do benchmark scores systematically misrepresent true model capabilities? When and how should systems adaptively retrieve information during reasoning? How do feedback loops and human judgment shape AI research autonomy? What determines when reasoning models help or hurt task performance? Can language models objectively evaluate the quality of their own work? How does polished presentation create false authority for AI outputs? How does explicit structure affect the depth and consistency of language model reasoning? Can personalization increase user trust without amplifying privacy risks? How does decomposing complex reasoning into subtasks affect performance and safety? How can AI assistance improve human judgment without creating overreliance? Can behavioral steering through simulation guide LLM improvements without costly retraining? How reliable is self-consistency as a self-training mechanism? What determines whether LLM recommenders outperform collaborative filtering approaches?