Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Augur: Modeling Covariate Causal Associations in Time Series via Large Language Models

Zhiqing Cui, Binwu Wang, Qingxiang Liu, Yeqiang Wang, Zhengyang Zhou, Yuxuan Liang, Yang Wang

TL;DR

Augur tackles the challenge of integrating causally meaningful covariates into time series forecasting by making the LLM the central reasoning engine. It introduces a two-stage teacher–student pipeline where a powerful teacher infers a directed causal graph and a grounded narrative, which are distilled into a lightweight student that performs fast forecasting with interpretable prompts. The approach yields competitive predictive accuracy and robust zero-shot generalization while providing transparent causal explanations that can guide feature selection and model priors. Extensive experiments on real-world datasets demonstrate strong performance across 26 baselines and highlight the value of causal narrative synthesis for interpretability and robustness in out-of-distribution settings.

Abstract

Large language models (LLM) have emerged as a promising avenue for time series forecasting, offering the potential to integrate multimodal data. However, existing LLM-based approaches face notable limitations-such as marginalized role in model architectures, reliance on coarse statistical text prompts, and lack of interpretability. In this work, we introduce Augur, a fully LLM driven time series forecasting framework that exploits LLM causal reasoning to discover and use directed causal associations among covariates. Augur uses a two stage teacher student architecture where a powerful teacher LLM infers a directed causal graph from time series using heuristic search together with pairwise causality testing. A lightweight student agent then refines the graph and fine tune on high confidence causal associations that are encoded as rich textual prompts to perform forecasting. This design improves predictive accuracy while yielding transparent, traceable reasoning about variable interactions. Extensive experiments on real-world datasets with 26 baselines demonstrate that Augur achieves competitive performance and robust zero-shot generalization.

Augur: Modeling Covariate Causal Associations in Time Series via Large Language Models

TL;DR

Augur tackles the challenge of integrating causally meaningful covariates into time series forecasting by making the LLM the central reasoning engine. It introduces a two-stage teacher–student pipeline where a powerful teacher infers a directed causal graph and a grounded narrative, which are distilled into a lightweight student that performs fast forecasting with interpretable prompts. The approach yields competitive predictive accuracy and robust zero-shot generalization while providing transparent causal explanations that can guide feature selection and model priors. Extensive experiments on real-world datasets demonstrate strong performance across 26 baselines and highlight the value of causal narrative synthesis for interpretability and robustness in out-of-distribution settings.

Abstract

Large language models (LLM) have emerged as a promising avenue for time series forecasting, offering the potential to integrate multimodal data. However, existing LLM-based approaches face notable limitations-such as marginalized role in model architectures, reliance on coarse statistical text prompts, and lack of interpretability. In this work, we introduce Augur, a fully LLM driven time series forecasting framework that exploits LLM causal reasoning to discover and use directed causal associations among covariates. Augur uses a two stage teacher student architecture where a powerful teacher LLM infers a directed causal graph from time series using heuristic search together with pairwise causality testing. A lightweight student agent then refines the graph and fine tune on high confidence causal associations that are encoded as rich textual prompts to perform forecasting. This design improves predictive accuracy while yielding transparent, traceable reasoning about variable interactions. Extensive experiments on real-world datasets with 26 baselines demonstrate that Augur achieves competitive performance and robust zero-shot generalization.

Paper Structure

This paper contains 68 sections, 1 theorem, 17 equations, 9 figures, 12 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related work
  3. Time Series Forecasting
  4. LLM for Time Series Forecasting
  5. Problem Formulation
  6. Time Series.
  7. Causal Explanation.
  8. Problem Definition
  9. Method
  10. Causal Explanation Generation via Teacher Model
  11. Heuristic Search Space Reduction.
  12. Pairwise Causal Judgment.
  13. Iterative Causal Graph Refinement.
  14. Grounded Narrative Synthesis.
  15. Distillation and Training of Student Agent
  16. ...and 53 more sections

Key Result

Theorem 1

Figures (9)

  • Figure 1: Motivation and problem illustration. Breathing new life into time series tasks with our Augur.
  • Figure 2: Overview of the Augur framework. Including causal explanation generation and student agent distillation for efficient downstream time series tasks.
  • Figure 3: Impact of narrative granularity on zero-shot classification performance across four datasets.
  • Figure 4: Impact of the SFT data scaling up.
  • Figure 5: Spearman correlation matrix of the variables in the power dataset.
  • ...and 4 more figures

Theorems & Definitions (2)

  • Theorem 1
  • proof