Skip to content

CastFlow: Learning Role-Specialized Agentic Workflows for Time Series Forecasting

Home / Papers / CastFlow: Learning Role-Specialized Agentic Workflows for Time Series Forecasting

CastFlow: Learning Role-Specialized Agentic Workflows for Time Series Forecasting

Authors: 潘柏凱, Mingyue Cheng, Zhiding Liu, Shuo Yu, Xiaoyu Tao, Yuchong Wu, Qi Liu, Defu Lian, Enhong Chen Date: 2026-04-30 Paper ID: openalex:2604.27840

Summary

CastFlow is an agentic framework that replaces the traditional one-shot LLM forecasting paradigm with a dynamic, multi-stage workflow involving planning, multi-view feature acquisition, and iterative refinement. By combining a frozen general-purpose LLM for reasoning with a fine-tuned, role-specialized LLM for numerical tasks, the framework leverages ensemble baselines to improve accuracy. The approach is optimized via a novel two-stage training process, achieving superior performance on diverse time series benchmarks compared to existing static models.

Key Contributions

  • Introduced CastFlow, an agentic forecasting framework that incorporates multi-view temporal pattern extraction and iterative forecast refinement.
  • Developed a role-specialized architecture separating general-purpose reasoning (frozen LLM) from domain-specific numerical forecasting (fine-tuned LLM).
  • Proposed a two-stage training paradigm using supervised fine-tuning and reinforcement learning with verifiable rewards (RLVR) to optimize agentic workflows.

Open Questions & Future Work

Key Concepts

  • agentic-time-series-forecasting: A paradigm for time series forecasting where LLMs iteratively plan, extract multi-view temporal features, and refine predictions through reflective workflows.

Archivist Review

The paper introduces a shift towards agentic, multi-stage workflows in LLM-based time series forecasting. I have generalized the framework name ‘CastFlow’ into a more reusable concept note for ‘Agentic Time Series Forecasting’. The open question regarding the trade-off between reasoning and numerical accuracy in such agents addresses a significant, non-boilerplate bottleneck.

Approved Concepts

  • Agentic Time Series Forecasting: It shifts the LLM forecasting paradigm from static, one-shot inference to iterative, multi-stage agentic workflows, enabling reasoning and reflection.

Approved Open Questions

  • Joint Reasoning-Numerical Accuracy Bottleneck: This captures the fundamental bottleneck in current agentic LLM-forecasting frameworks, specifically regarding the interaction between frozen reasoning models and fine-tuned numerical predictors.

Rejected Candidates

  • [concept] CastFlow (castflow) - paper_local: This is a specific framework name, better categorized under the more general Agentic Time Series Forecasting concept.

Metadata & Links

url
https://arxiv.org/abs/2604.27840
paper_id
2604.27840
paper_source
openalex
domain
nlp
tags
time-series-question-answering-tsqa
architectures
datasets
concept_slugs
agentic-time-series-forecasting
dataset_slugs
skill
TimeSeriesSkill
created_at
2026-05-03T06:02:20Z