Large Language Model-Based Time-Series Data Summarization Using a Novel Multi-Tiered Retrieval Pipeline

Financial time-series data summarization is a critical task for analysts and decision-makers; yet current approaches rely on labor-intensive manual interpretation or bespoke reporting scripts. In this research, a novel multi-tier retrieval pipeline that leverages large language models (LLMs) to automatically generate concise and accurate summaries of financial data is proposed. By integrating traditional templated methods with retrieval-augmented prompts (using domain-specific examples to guide the LLM) the challenge of balancing linguistic fluency with numerical precision is addressed. The pipeline is evaluated on a knowledge base constructed from daily OHLCV data of 20 large-cap companies, with additional testing on a mid-cap ticker not present in the original dataset (COF). Over 15 independent runs were conducted; these experimental results demonstrate that while direct summarization captures general trends, this templated and refined multi-tier retrieval pipeline ensures factual correctness and achieves the best overall performance in terms of ROUGE, BERTScore, and numeric fidelity. This pipeline has the potential to transform time-series analysis by reducing manual workload and improving the reliability of automated financial data summarization.