CastFSR: Agentic LLM Framework for Context-Aware Time Series Forecasting

A novel agentic framework called CastFSR has been introduced to enhance context-aware time series forecasting by leveraging large language models. This framework operates through a three-stage process: Fast, Slow, and Reflect. Initially, the "fast thinking" stage profiles observations and selects lightweight forecasters to establish a data-driven forecast prior. Following this, the "slow deliberation" stage retrieves contextual evidence, dynamically determines relevant historical windows, and reasons about how these contexts influence future dynamics. Finally, the "reflection" stage iteratively refines the forecasts to ensure consistency across temporal, contextual, and domain constraints. CastFSR supports both training-free inference with off-the-shelf LLMs and efficient deployment via a two-stage fine-tuning and reinforcement learning strategy, demonstrating superior performance against current benchmarks.