Title: Connecting the Final Gap in Time Series Prediction via LLM Agents

Abstract:

The field of time series forecasting has experienced significant progress, particularly driven by foundation models that demonstrate robust zero-shot capabilities in numerical extrapolation. Nevertheless, in practical applications, a statistically sound baseline is seldom the final output employed in decision-making processes. Prior to becoming actionable, forecasts typically require adjustment to incorporate loosely structured business nuances, including holiday impacts, marketing initiatives, external occurrences, historical precedents, and insights from domain experts. This critical phase has received limited attention in current forecasting research.

This study defines this overlooked stage as the last-mile forecasting challenge and introduces an LLM-agent framework built upon a forecasting backbone. The proposed architecture manages a consolidated forecast workspace, utilizes tools to gather contextual evidence, and transforms reasoning processes into concrete forecast revisions while adhering to structural safety protocols. Furthermore, the system facilitates long-horizon predictions via map-reduce-style decomposition and enhances accuracy through post-hoc reflection supported by a memory bank. Designed for both controllability and auditability, our approach is validated through real-world case studies, illustrating how LLM agents effectively close the divide between statistical models and business-ready forecasting solutions.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC