Position Paper · arXiv 2026

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

Mingyue Cheng, Xiaoyu Tao, Qi Liu, Ze Guo, Enhong Chen
University of Science and Technology of China

Abstract

TL;DR We argue that time series forecasting should be reframed as an agentic process — moving beyond static, model-centric pipelines toward systems that perceive, plan, act, reflect, and evolve through accumulated experience.

Conventional time series forecasting treats prediction as a model-centric, static, and single-pass problem. This framing proves fundamentally inadequate for complex, adaptive, multi-turn scenarios encountered in real-world deployments. We propose Agentic Time Series Forecasting (ATSF) — a principled framework that organizes forecasting as an iterative agentic workflow. Rather than optimizing a single model's predictive accuracy in isolation, ATSF enables forecasting systems to interact with tools, incorporate feedback from outcomes, and evolve through accumulated experience. We formalize the framework through five interconnected cognitive components — Perception, Planning, Action, Reflection, and Memory — and propose three concrete implementation paradigms: workflow-based design, agentic reinforcement learning, and hybrid agentic workflows. We conclude by outlining key research opportunities and open challenges to guide the community toward this new frontier.

Framework

ATSF CYCLE 👁 PERCEPT 🧠 PLANNING ACTION 🔍 REFLECT 💾 MEMORY
Perception adaptive input

Extracts task-relevant information from raw, noisy inputs. Adaptive rather than fixed preprocessing — determines what information matters for each specific context.

Planning dynamic strategy

Formulates forecasting objectives and decomposes complex tasks into subtasks. Revised dynamically as new observations or feedback arrive.

Action tool use

Executes decisions through autonomous tool interaction. Forecasting is treated as one action within a broader action space, enabling integration of diverse methods.

Reflection self-evaluation

Iteratively evaluates, interprets, and revises predictions. Supports self-judgment without requiring external supervision or ground-truth feedback.

Memory continual learning

Preserves reusable experience — patterns, strategies, failure cases — across instances. Hierarchical access at instance, task, and domain levels.

Contributions

01

Paradigm Formalization

First position paper to formally define the ATSF paradigm, establishing a principled conceptual framework for agentic forecasting.

02

Limitation Analysis

Rigorous characterization of why static, model-centric approaches fail in adaptive, multi-turn real-world forecasting scenarios.

03

Implementation Paradigms

Three concrete strategies spanning the interpretability–adaptability design space, with comparative analysis of mechanisms and trade-offs.

04

Research Roadmap

Comprehensive outline of open challenges and future research directions to guide the community toward this new frontier.

Implementation Paradigms

01
Workflow-Based Design
Structured Pipelines

Structures forecasting as predefined sequences of cognitive steps using directed acyclic graphs (DAGs) or standard operating procedures. Organizes the five ATSF components into interpretable, deterministic pipelines.

Interpretable Stable Rigid for unseen scenarios
02
Agentic RL
AgenticRL

Applies reinforcement learning to the decisions surrounding forecasting — objective planning, action selection, strategy revision — rather than to predictive models alone. Enables autonomous discovery of novel strategies.

Autonomous evolution Novel strategy discovery Training instability
03
Hybrid Agentic Workflow
AgentFlow

Integrates explicit workflow design with localized reinforcement learning at specific decision points. Balances the stability of structured workflows with the adaptability of learned strategies.

Balanced design Adaptable Architectural complexity

Paradigm Evolution

Era I
Statistical
ARIMA, Prophet
Era II
Machine Learning
XGBoost, LightGBM
Era III
Deep Learning
Informer, PatchTST
Era IV
Foundation Models
Chronos, TimesFM
Era V
LLM-Based
Time-LLM
Era VI
Agentic TSF
ATSF

Related Work

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

Xiaoyu Tao, Mingyue Cheng, Chuang Jiang, Tian Gao, Huanjian Zhang, Yaguo Liu
Reformulates time series forecasting as a sequential decision-making problem rather than a static input-output mapping. Introduces a memory-based state management mechanism that maintains decision-relevant context across multi-step interactions, combined with a tool-augmented agentic workflow for feature extraction, model invocation, and iterative refinement. Trained via dual-stage supervised fine-tuning and multi-turn reinforcement learning with curriculum learning.
Xiaohan Zhang, Tian Gao, Mingyue Cheng, Bokai Pan, Ze Guo, Yaguo Liu, Xiaoyu Tao, Qi Liu
Proposes an LLM-based agentic framework that reimagines time series forecasting through iterative, expert-like reasoning rather than single-pass regression. Integrates temporal features, domain knowledge, historical case references, and contextual information through a multi-stage workflow: context preparation, reasoning-based generation, and reflective evaluation. Achieves training-free forecasting while demonstrating superior performance across multiple benchmarks.

Citation

BibTeX
@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}
}