Bridging the Last Mile of Time Series Forecasting with LLM Agents

Researchers present an LLM-agent framework that enhances time series forecasting by incorporating business context and expert judgment into statistical predictions. The system bridges the gap between raw forecasts and decision-ready outputs through structured reasoning, contextual evidence retrieval, and auditable revision mechanisms.

The paper addresses a critical but overlooked stage in forecasting workflows: the translation of statistical predictions into actionable business decisions. While foundation models have demonstrated strong zero-shot performance on numerical extrapolation tasks, real-world forecasting rarely stops at raw model output. Organizations typically adjust forecasts based on domain knowledge, upcoming events, seasonal patterns, and expert intuition—processes that remain largely unformalized in academic literature.

This research formulates what it calls the "last-mile forecasting" problem and proposes an LLM-agent solution that operates atop existing forecasting models. The system maintains a unified workspace, retrieves contextual evidence through tool invocation, and converts reasoning processes into explicit forecast revisions subject to safety constraints. Key features include map-reduce-style decomposition for long-horizon predictions and a memory bank enabling post-hoc reflection and learning.

For practitioners in financial forecasting, supply chain management, and demand planning, this approach addresses a significant operational gap. Current workflows often rely on manual adjustments or ad-hoc expert input, introducing inconsistency and reducing auditability. An LLM-agent framework that systematically incorporates business context while maintaining transparency could improve forecast quality and organizational accountability.

The emphasis on controllability and auditability is particularly valuable for regulated industries where forecast decisions must be documented and justified. As organizations deploy more AI systems in high-stakes forecasting applications, having transparent, traceable reasoning pathways becomes increasingly important. The framework's modular design also suggests it could integrate with diverse forecasting backends, maximizing applicability across domains.

• →LLM agents can systematically incorporate business context and expert knowledge into statistical forecasts, addressing a practical gap overlooked in forecasting research.
• →The framework maintains auditability and controllability through explicit reasoning trajectories and structural safety constraints.
• →Long-horizon forecasting becomes more tractable through map-reduce-style decomposition and iterative refinement with memory mechanisms.
• →This approach bridges the last-mile problem where raw predictions require revision before becoming decision-ready outputs.
• →The system's modularity enables integration with multiple forecasting backbones, broadening applicability across industries and use cases.