Representing Time Series as Structured Programs for LLM Reasoning

The paper introduces the Time-Series-to-Structured-Program (T2SP) representation, a deterministic method that reformulates time series into structured symbolic programs to enhance reasoning capabilities of large language models (LLMs). By decomposing time series into trends, periods, and events, T2SP aligns with the textual modalities of LLMs, improving performance on reasoning tasks like editing, captioning, and question answering while reducing computational overhead. This approach allows practitioners to leverage existing LLMs for time-series analysis without the need for fine-tuning, addressing the modality mismatch that typically hampers performance.