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

This paper proposes Agentic Time Series Forecasting (ATSF), a paradigm shift from traditional static, model-centric prediction to a dynamic, iterative process driven by agentic workflows that integrate perception, planning, action, reflection, and memory to enable adaptive and continual forecasting.

The Gist

Imagine you are trying to predict the weather for next week.

The Old Way (Model-Centric):
In the past, forecasting was like hiring a very smart, but rigid, calculator. You fed it all the historical weather data (temperature, wind, rain from the last 10 years), and it spat out a single number: "It will be 72°F on Tuesday."
If the calculator made a mistake, you couldn't really ask it why. If a sudden storm front appeared that the calculator didn't know about, it just kept giving you the same wrong answer. It was a "one-and-done" job. It didn't think; it just calculated.

The New Way (Agentic Time Series Forecasting):
This paper argues that we need to stop treating forecasting like a calculator and start treating it like a human detective or a smart project manager. This new approach is called Agentic Time Series Forecasting (ATSF).

Instead of just crunching numbers, the system acts like an "Agent" that goes through a cycle of thinking and doing, much like a human expert would. Here is how it works, using a simple analogy:

The "Smart Detective" Analogy

Imagine you are a detective trying to solve a mystery (predicting the future).

1. Perception (Looking at the Clues):

   • Old Way: The detective just looks at the raw police report.
   • New Way: The agent looks at the messy crime scene. It decides, "Hey, this noise in the data is just background chatter; let's ignore it. But this weird pattern in the traffic data? That's important!" It filters out the junk and focuses on what actually matters.

2. Planning (Making a Strategy):

   • Old Way: The detective immediately guesses the culprit.
   • New Way: The agent pauses and says, "Okay, to solve this, I need to check the bank records, talk to the witness, and look at the security footage. I'll do the bank records first." It breaks the big problem into small, manageable steps.

3. Action (Using Tools):

   • Old Way: The detective has no tools; they just guess.
   • New Way: The agent actually uses tools. It might call a "Statistical Model" tool to check trends, a "Search Engine" tool to find news about a sudden event, or a "Simulation" tool to see what happens if it rains. It picks the right tool for the job.

4. Reflection (Checking the Work):

   • Old Way: The detective writes the report and leaves.
   • New Way: The agent looks at its own guess and asks, "Does this make sense? Wait, I predicted rain, but the news just said the storm was canceled. My prediction is wrong. I need to rethink my plan." It critiques itself and fixes its own mistakes.

5. Memory (Learning for Next Time):

   • Old Way: The detective forgets the case once it's over.
   • New Way: The agent writes a note in its journal: "Next time I see a pattern like X, I should check the news first." It remembers what worked and what didn't, so it gets smarter every time it solves a case.

Why Do We Need This Change?

The paper says the old "calculator" method is failing us in the real world because:

• The world changes: Real life isn't static. A pandemic, a stock market crash, or a new policy changes everything instantly. A rigid calculator can't adapt; a detective can.
• It's a conversation, not a command: In real life, we don't just want a number; we want to know why and what if. The new agent can have a back-and-forth conversation, refining its answer as you give it more information.
• It learns from experience: The old models need to be retrained from scratch to learn new things. The new agent just updates its "journal" (memory) and keeps going.

The Three Ways to Build This "Detective"

The paper suggests three ways to build these smart agents:

1. The Flowchart Method (Workflow): You give the detective a strict checklist (First check A, then check B). It's reliable and easy to understand, but a bit rigid.
2. The Trial-and-Error Method (Reinforcement Learning): You let the detective learn by doing. It tries things, gets rewarded for good guesses, and punished for bad ones. It's very flexible but can be messy and hard to control.
3. The Hybrid Method (AgentFlow): This is the best of both worlds. You give the detective a general plan (Workflow) but let it learn and adapt specific parts of the plan on its own (Learning).

The Bottom Line

This paper is a call to action for researchers. It says: "Stop just building better calculators. Start building better thinkers."

By turning forecasting into an active, thinking process that can plan, use tools, check its own work, and remember the past, we can create systems that are much better at helping us make decisions in a chaotic, changing world. It's the difference between asking a robot for a number and asking a brilliant human consultant for a strategy.

Technical Summary

Based on the position paper "Beyond Model-Centric Prediction—Agentic Time Series Forecasting" by Mingyue Cheng et al., here is a detailed technical summary.

1. Problem Statement

The paper identifies a fundamental structural mismatch between traditional model-centric time series forecasting and the realities of real-world decision-making.

• Current Paradigm Limitations: Traditional forecasting is formulated as a static, single-pass supervised learning problem. It maps historical observations to future values using a fixed model architecture. While this has driven accuracy improvements through better representations and scaling, it suffers from:
  • Static Nature: Once inputs and horizons are defined, the process cannot adapt to new information or changing objectives.
  • Closed-Loop Operation: It lacks mechanisms to interact with external tools, domain knowledge, or complementary analyses.
  • Single-Pass Execution: Predictions are generated once as final outputs, lacking iterative self-correction, reflection, or revision based on uncertainty.
  • No Experience Accumulation: Knowledge from past forecasting instances is not explicitly stored or reused; adaptation relies solely on retraining parameters.
• The Gap: Real-world forecasting is inherently adaptive and multi-turn. It involves interpreting raw data, selecting features, hypothesizing, revising predictions based on discrepancies, and evolving strategies over time. Current models fail to capture this "process-level complexity."

2. Methodology: Agentic Time Series Forecasting (ATSF)

The authors propose ATSF as a conceptual reframing of forecasting. Instead of viewing it as a function $f(x) \to y$, ATSF views it as an agentic process composed of five core cognitive components:

1. Perception: The system extracts task-relevant information from noisy, heterogeneous, and unstructured inputs. Unlike fixed preprocessing, this is an adaptive cognitive process that determines what information is relevant based on context.
2. Planning: Before execution, the agent formulates objectives, decomposes complex tasks into sub-tasks, and determines high-level strategies. Crucially, planning is dynamic and can be revised based on intermediate outcomes or feedback.
3. Action: The agent executes decisions by interacting with a toolkit of external tools. This includes:
   • Invoking predictive models (statistical, deep learning, LLMs).
   • Performing statistical analyses or retrieving auxiliary data.
   • Note: Forecasting itself is treated as just one type of action within a broader action space.
4. Reflection: An internal self-evaluation mechanism where the system assesses the quality, plausibility, and uncertainty of its predictions. It identifies discrepancies and decides whether to trigger corrective actions or alternative strategies.
5. Memory: A mechanism for experience accumulation. It stores reusable knowledge (patterns, strategies, failure cases) in a structured format (e.g., knowledge graphs, embeddings) rather than just updating model parameters. This allows for cross-instance transfer and continual evolution.

Implementation Paradigms

The paper outlines three ways to instantiate ATSF:

• Workflow-Based Design: Structures forecasting as a predefined sequence of cognitive steps (e.g., DAGs or SOPs). It prioritizes interpretability and stability but may lack flexibility for unseen scenarios.
• Agentic Reinforcement Learning (AgenticRL): Treats forecasting as an interactive decision process where an agent learns policies through trial-and-error and reward feedback. It enables autonomous evolution and discovery of novel strategies but suffers from training instability and sample inefficiency.
• Hybrid Agentic Workflow (AgentFlow): Integrates the stability of structured workflows with the adaptability of reinforcement learning. Learning is applied selectively to specific decision points (e.g., context selection) within a fixed workflow, balancing reliability with flexibility.

3. Key Contributions

• Conceptual Reframing: The paper formally defines ATSF, shifting the focus from "predictive model capacity" to "process organization." It argues that forecasting performance depends on how activities are structured, coordinated, and adapted.
• Theoretical Framework: It defines the five core components (Perception, Planning, Action, Reflection, Memory) and explains how they interact to create a multi-turn, iterative decision-making loop.
• Implementation Roadmap: It categorizes existing and future approaches into three distinct paradigms (Workflow, AgenticRL, Hybrid), providing a clear taxonomy for researchers.
• Critical Analysis: The paper identifies 7 key challenges that must be addressed to realize ATSF:
  1. Memory design for effective experience transfer.
  2. Standardization of toolkits and interfaces.
  3. Multi-agent coordination and credit assignment.
  4. Reliability and uncertainty management in iterative reasoning.
  5. Efficiency and scalability (computational overhead).
  6. Safety, privacy, and deployment constraints.
  7. Accountability in human-agent collaborative decision-making.

4. Results and Evidence

Note: As this is a Position Paper, it does not present a single new model with benchmark results (e.g., MSE/MAE scores) to prove superiority over SOTA. Instead, its "results" are theoretical and argumentative:

• Evidence of Necessity: The authors use logical deduction and analysis of current limitations (Table 1) to demonstrate that model-centric approaches cannot solve adaptive, long-term, or decision-oriented problems.
• Feasibility: The paper cites recent advances in Large Language Models (LLMs) and agentic systems (e.g., Toolformer, Reflexion) to argue that the technical capabilities (reasoning, planning, tool use) required for ATSF now exist.
• Representative Works: The paper references emerging works (e.g., TimeCopilot, Timeseriesscientist, Cast-R1) that already embody elements of ATSF, serving as proof-of-concept that this paradigm is viable.

5. Significance and Future Impact

• Paradigm Shift: ATSF moves the field from "better models" to "better systems." It suggests that the next leap in forecasting accuracy will come from system design, tool orchestration, and iterative reasoning rather than just deeper neural networks.
• Alignment with Human Expertise: By explicitly modeling planning, reflection, and memory, ATSF aligns automated forecasting with how human experts actually work (iterative, hypothesis-driven, and experience-based), leading to more interpretable and trustworthy systems.
• Decision-Centric Focus: It bridges the gap between raw prediction and downstream decision-making, allowing forecasts to be revised based on operational constraints and uncertainty, which is critical for high-stakes domains like healthcare, finance, and energy.
• Research Agenda: The paper establishes a unified framework for future research, encouraging the community to develop principled agentic workflows, adaptive mechanisms, and evaluation metrics that capture the process of forecasting, not just the final output.

In summary, the paper argues that the future of time series forecasting lies not in refining the prediction engine alone, but in building agentic systems that perceive, plan, act, reflect, and remember, thereby transforming forecasting from a static calculation into a dynamic, evolving decision-support process.