LLMs Generate Explainable Insights for Time Series Forecasts

Time series forecasts are crucial for decision-making in many fields, but their utility is often limited without clear explanations. Manually generating these explanations is costly and time-consuming, while automated methods using large language models (LLMs) frequently suffer from hallucination when dealing with temporal data. This research proposes a novel, domain-agnostic framework designed to produce grounded natural language explanations for time series forecasts. The framework comprises three main components: first, it extracts structured explanatory factors from historical, human-written explanations; second, it generates explanations conditioned on verifiable evidence; and third, it includes a scalable evaluation process for readability, logical consistency, and persuasiveness. By explicitly restricting generation to evidence-based facts, the framework significantly reduces the risk of unsupported claims. The system was tested on financial forecasting (NASDAQ-100) and freight pricing, demonstrating that its generated explanations closely matched human-written ones in terms of quality metrics like readability, consistency, and persuasiveness, all without requiring domain-specific fine-tuning.