Augmenting Fundamental Analysis with Large Language Models: A RAG-Based System for Generating Investor Briefs

Battle Count: 85

The paper focuses on fundamental analysis and investor brief generation rather than quantitative trading strategies. It does not involve price prediction, algorithmic trading, portfolio optimization, or signal generation for automated execution. However, the information synthesis capabilities (news ranking, macroeconomic context, company-specific KPIs) could serve as inputs to a quantitative trading pipeline. The system is explicitly designed as an advisory tool, not a decision system.

The system requires: (1) data preprocessing pipelines for heterogeneous sources (EDGAR filings, macroeconomic data, news), (2) RAG architecture with vector database (LanceDB), (3) prompt engineering with heuristic templates and Kitchin cycle knowledge, (4) K-means clustering for news analysis, (5) GPT-4o API integration, (6) WordCloud visualization, and (7) a multi-phase processing pipeline for long documents. While individual components are well-documented, the integration and tuning of the full pipeline requires significant engineering effort. No code is provided.

The paper describes the methodology in detail but does not provide code, specific prompts, or a GitHub repository. The system relies on proprietary GPT-4o API access. Data sources (EDGAR, macroeconomic indicators) are publicly available, but the specific preprocessing pipeline, prompt engineering, and RAG configuration are not fully disclosed. The user study with 9 participants is not reproducible without access to the same individuals.