KairosAgent: Agentic Time Series Forecasting with F...

KairosAgent: Agentic Time Series Forecasting with F... | AI Research

Key Takeaways

KairosAgent is a new framework designed to improve time series forecasting by combining the analytical strengths of Large Language Models (LLMs) with the num...
Cross-domain multimodal time series forecasting is a challenging task, requiring models to integrate precise numerical comprehension, cross-domain semantic understanding, and effective multimodal fusion.
Existing approaches either build Time Series Foundation Models (TSFMs) from scratch or leverage pretrained Large Language Models (LLMs).
However, TSFMs often overlook semantic understanding and lack the ability to perform future-oriented semantic reasoning, and LLMs struggle with numerical comprehension and accurate quantitative forecasting.
To overcome these limitations, we propose KairosAgent, a novel agentic framework for multimodal time series forecasting, including an LLM-based reasoner and a TSFM-based forecaster.

Paper AbstractExpand

Cross-domain multimodal time series forecasting is a challenging task, requiring models to integrate precise numerical comprehension, cross-domain semantic understanding, and effective multimodal fusion. Existing approaches either build Time Series Foundation Models (TSFMs) from scratch or leverage pretrained Large Language Models (LLMs). However, TSFMs often overlook semantic understanding and lack the ability to perform future-oriented semantic reasoning, and LLMs struggle with numerical comprehension and accurate quantitative forecasting. To overcome these limitations, we propose KairosAgent, a novel agentic framework for multimodal time series forecasting, including an LLM-based reasoner and a TSFM-based forecaster. KairosAgent unifies textual reasoning and numerical forecasting by dynamically invoking analytical tools to enhance the numerical understanding and semantic reasoning capabilities of LLMs. The reasoning results are subsequently fused into the TSFM pipeline, enabling more accurate and reliable future predictions. To further improve the reasoning, we curate a large-scale corpus of high-quality trajectories, alongside a reinforcement learning from forecasting paradigm with multi-turn refinement and turn-level credit assignment. Experiments demonstrate that KairosAgent achieves superior zero-shot forecasting performance while maximizing the utility of pretrained LLMs and TSFMs, presenting a promising direction for efficient and interpretable time series agents. The project page is at this https URL .

KairosAgent is a new framework designed to improve time series forecasting by combining the analytical strengths of Large Language Models (LLMs) with the numerical precision of Time Series Foundation Models (TSFMs). While traditional models often struggle by either ignoring semantic context or failing at complex numerical calculations, KairosAgent bridges this gap. It uses an agentic approach to reason about the "morphology"—or the qualitative shape and pattern—of future data, which is then integrated into a specialized forecasting model to produce more accurate and reliable predictions.

Bridging Reasoning and Forecasting

The framework operates through a two-stage process. First, an LLM-based reasoner acts as an agent that interacts with a set of analytical tools to examine historical data. Instead of simply looking at raw numbers, it identifies trends, periodic patterns, and volatility. It then summarizes these findings into a "morphology description." Second, this description is fed into a TSFM-based forecaster as a semantic guide. By using a gated fusion mechanism, the forecaster can adaptively incorporate this high-level reasoning into its numerical predictions, ensuring that the final output is grounded in both statistical evidence and logical deduction.

Training for Better Performance

To ensure the model learns effectively, the researchers developed the T-STAR corpus, which contains over 40,000 high-quality reasoning trajectories. The training process is split into three distinct phases: 1. Supervised Fine-Tuning: The reasoner is trained to correctly use analytical tools and generate accurate morphology descriptions. 2. Multimodal Alignment: The forecaster is trained to interpret these descriptions as useful priors for numerical prediction. 3. Reinforcement Learning: The model is refined using a "turn-level" credit assignment method. Unlike traditional methods that only reward the final result, this approach evaluates the contribution of each individual reasoning step, allowing the agent to learn which specific actions lead to better forecasting accuracy.

Results and Capabilities

Experiments show that KairosAgent achieves superior zero-shot forecasting performance compared to existing models. By separating the task into semantic reasoning and numerical forecasting, the framework avoids common pitfalls like numerical hallucinations—where models generate plausible-sounding but mathematically incorrect data—and "black-box" statistical fitting. This makes the agent’s decision-making process more interpretable, as users can see the reasoning behind the predicted future patterns.

Key Considerations

KairosAgent is designed to be generalizable across diverse domains, such as climate, economy, and energy. By focusing on qualitative morphology rather than forcing the LLM to output raw numbers, the framework maintains high numerical accuracy. The use of turn-level rewards is a significant shift in how these agents are optimized, as it provides a more granular learning signal that helps the model improve its intermediate logic rather than just guessing the final outcome.