Toward Reasoning-Centric Time-Series Analysis

TL;DR

The paper argues that time-series analysis should move beyond pattern recognition toward reasoning about causal structure and context, especially in dynamic real-world settings. It proposes a reasoning-centric paradigm that leverages LLMs to process multimodal inputs, generate explicit causal explanations, and act as cognitive agents guiding data selection, hypothesis formation, and evaluation. A three-level taxonomy of time-series reasoning (Structured, Context-aware, Open-world) with seven reasoning types anchors the framework, complemented by an instructive LLM-based pipeline and a toy demonstration showing adaptive causal filtering. The outlook discusses benefits and challenges of this approach, emphasizing interpretability, robustness to regime shifts, and directions such as domain-grounded instruction tuning and integration with symbolic causal tools to enable practical, transparent time-series understanding.

Abstract

Traditional time series analysis has long relied on pattern recognition, trained on static and well-established benchmarks. However, in real-world settings -- where policies shift, human behavior adapts, and unexpected events unfold -- effective analysis must go beyond surface-level trends to uncover the actual forces driving them. The recent rise of Large Language Models (LLMs) presents new opportunities for rethinking time series analysis by integrating multimodal inputs. However, as the use of LLMs becomes popular, we must remain cautious, asking why we use LLMs and how to exploit them effectively. Most existing LLM-based methods still employ their numerical regression ability and ignore their deeper reasoning potential. This paper argues for rethinking time series with LLMs as a reasoning task that prioritizes causal structure and explainability. This shift brings time series analysis closer to human-aligned understanding, enabling transparent and context-aware insights in complex real-world environments.

Figures (4)

• Figure 1: Seven types of reasoning essential for Time Series Analysis.
• Figure 2: An instructive pipeline for time series analysis with LLMs.
• Figure 3: Toy example illustrating how LLM-guided reasoning improves causal alignment in TSF. We compare attention distributions under three strategies: (1) Baseline, with no causal filtering; (2) LLM-guided, where LLM interprets textual prompts to generate dynamic attribution masks; and (3) Manual, using oracle masks based on known ground-truth structure. The gray boxes show the input context provided to each strategy: natural language prompt for LLM-guided, hard-coded rule for Manual, and no causal input for Baseline. The green box summarizes the true causal rules used in generating the synthetic time series -- $x_1$ influences $y$ in the early stage ($t{<}150$), and both $x_1$ and $x_2$ influence $y$ in the late stage ($t{\geq}150$). Results show that only the LLM-guided and Manual models adaptively shift attention from $x_1$ to $x_2$ as causal drivers evolve, while ignoring the spurious $x_3$.
• Figure 4: Interpretability in time series analysis (time series forecasting as an example). This figure illustrates how (1) pre-training data influence, (2) architectural behavior analysis, and (3) instance-level attribution provide insights into the relationship between inputs and their influence on the target output. The red curve represents key features impacting the blue curve, while the red text highlights important textual features relevant to the curve in the yellow-shaded area. The red network layer corresponds to the component responsible for analyzing red curve data (key features).