CastFlow: Dynamic Agentic Framework for Time Series Forecasting
CastFlow introduces an innovative framework that employs a dynamic agentic method for time series forecasting utilizing large language models (LLMs). In contrast to traditional static generative models that predict future values from past data in one step, CastFlow divides the forecasting into distinct stages: planning, action, forecasting, and reflection, all enhanced by a memory module. This structure facilitates the extraction of multi-view temporal patterns, the acquisition of contextual features through multiple rounds, iterative refinement of forecasts, and the creation of ensemble forecasts. It effectively overcomes the limitations of current LLM-based approaches, which struggle with temporal pattern extraction and context acquisition in a single round. This research is available on arXiv with the identifier 2604.27840.
Key facts
- CastFlow is a dynamic agentic forecasting framework.
- It uses large language models (LLMs) for time series forecasting.
- The framework organizes forecasting into planning, action, forecasting, and reflection.
- It includes a memory module to support the workflow.
- CastFlow enables multi-view temporal pattern extraction.
- It supports multi-round contextual feature acquisition.
- The framework allows iterative forecast refinement.
- It incorporates ensemble forecasts.
- The work is published on arXiv (2604.27840).
Entities
Institutions
- arXiv