Agentic Time Series Forecasting

Position: Beyond Model-Centric Prediction — Agentic Time Series Forecasting

Overview

Conventional time series forecasting frames prediction as a model-centric, static, single-pass problem — fit a model, run inference, report metrics. This framing is fundamentally inadequate for the complexity and adaptivity demanded by real-world deployments.

This position paper argues for a paradigm shift toward Agentic Time Series Forecasting (ATSF): organizing forecasting as an iterative agentic workflow where systems perceive context, plan strategies, act through tools, reflect on outcomes, and remember experience — evolving continuously rather than remaining frozen after training.

Framework

ATSF is built around five interconnected cognitive components:

Component	Role
Perception	Adaptive extraction of task-relevant signals from raw, noisy inputs
Planning	Dynamic decomposition of forecasting objectives; revised as new information arrives
Action	Autonomous tool interaction; forecasting is one action among a broader action space
Reflection	Iterative self-evaluation and prediction revision without external supervision
Memory	Hierarchical accumulation of patterns, strategies, and failure cases across instances

Implementation Paradigms

We propose three concrete paradigms spanning the interpretability–adaptability spectrum:

• Workflow-Based Design — Structured cognitive pipelines (DAGs / SOPs); high interpretability and stability
• Agentic Reinforcement Learning (AgenticRL) — RL applied to the decisions surrounding forecasting; enables autonomous strategy discovery
• Hybrid Agentic Workflow (AgentFlow) — Workflow structure with localized RL at decision points; balances stability with adaptability

Authors

Mingyue Cheng, Xiaoyu Tao, Qi Liu, Ze Guo, Enhong Chen

University of Science and Technology of China

Related Work

A curated list of papers on agentic time series forecasting, maintained alongside this position paper.

Paper	Authors	Venue	Links
Cast-R1: Learning Tool-Augmented Sequential Decision Policies for Time Series Forecasting	Xiaoyu Tao, Mingyue Cheng, et al.	arXiv Feb 2026	arXiv · Code
CastMind: An Interaction-Driven Agentic Reasoning Framework for Cognition-Inspired Time Series Forecasting	Xiaohan Zhang, Mingyue Cheng, et al.	arXiv Nov 2025	arXiv

Citation

@article{cheng2026atsf,
  title   = {Position: Beyond Model-Centric Prediction -- Agentic Time Series Forecasting},
  author  = {Cheng, Mingyue and Tao, Xiaoyu and Liu, Qi and Guo, Ze and Chen, Enhong},
  journal = {arXiv preprint arXiv:2602.01776},
  year    = {2026}
}