FAE: A Multi-Agent System for Automated and Explainable Multivariate Time Series Forecasting Pipelines

ACL ARR 2026 January Submission2728 Authors

Keywords: Agents, AutoML, Times Series, Forecaster, Explainer

Abstract: This paper introduces FAE (Forecaster Agent and Explainer), a multi-agent system that automates and explains multivariate time series forecasting pipelines integrated with an AutoML tool called AutoDCE-TS. The system integrates three cooperative agents to execute pipelines, handle errors, summarize intermediate steps, and generate user-oriented explanations using Large Language Models. FAE autonomously performs data preprocessing, model configuration, forecasting, and result interpretation, while supporting interactive explanations through Retrieval-Augmented Generation. Experimental results show that AutoDCE-TS configured by FAE achieves predictive performance comparable to manual configurations across multiple datasets and horizons. A case study with different LLMs highlights trade-offs between reasoning depth, efficiency, and communication quality. The results demonstrate the potential of agent-based LLM systems to automate forecasting workflows while improving transparency and interpretability.

Paper Type: Long

Research Area: AI/LLM Agents

Research Area Keywords: LLM/AI agents, Interpretability, retrieval-augmented generation, applications, reasoning

Contribution Types: Model analysis & interpretability, Data analysis

Languages Studied: English

Submission Number: 2728