Využití velkých jazykových modelů k analýze dat z kapitálových trhů

Abstract

This thesis presents an experimental evaluation of large language models (LLMs) as tools for capital markets data analysis. The primary objective is to quantitatively assess whether current LLMs can match or surpass established state-of-the-art methods in specific financial document analysis tasks, thereby reducing the information disadvantage of retail investors relative to institutional market participants.

A modular experimental platform was designed and implemented in Python, utilizing the Streamlit framework, a PostgreSQL database, and the LangChain ecosystem. The platform automatically acquires regulatory filings from the U.S. Securities and Exchange Commission (SEC), market data, and structured financial data in XBRL format. Ten controlled experiments were designed on this data infrastructure, covering sentiment analysis, structured information extraction, semantic document comparison, financial event prediction, and information retrieval. Each experiment compares a traditional baseline method with an LLM-based approach, and results are evaluated using standard statistical tests at a significance level of $\alpha = 0.05$.

Four key findings emerge from the ten completed experiments. LLMs demonstrate statistically significant superiority in tasks requiring contextual understanding, such as detecting tone-fundamentals misalignment or classifying management stance toward discussed entities. Conversely, LLMs systematically fail in deterministic numerical and extraction tasks -- a control experiment revealed critical hallucination rates when computing financial ratios from unstructured text, and rule-based systems significantly outperformed LLMs in structured data extraction. Model capacity has a measurable but non-monotonic effect on output quality -- higher capacity improves performance on contextual tasks but may degrade it on tasks dominated by format consistency requirements. The thesis therefore recommends a complementary hybrid architecture in which LLMs provide qualitative analytical capabilities while deterministic systems process structured numerical data.