LTAT bakalaureusetööd – Bachelor's theses

Sirvi

Nüüd näidatakse 1 - 1 1

Sentiment Analysis of Forward-Looking Statements from Annual Reports Using Large Language Models
(Tartu Ülikool, 2025) Post, Hardo; Milani, Fredrik Payman, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
This thesis presents a system for extracting and analyzing forward-looking statements from the annual reports of OMX Nasdaq Stockholm main list companies. The system uses large language models (LLMs), particularly Google Gemini 2.5 Pro, to identify forward-looking statements, classify them by type and theme, and assign sentiment labels. These statements are aggregated into concise company and sector summaries, enabling sentiment-based rankings and natural language querying via a chatbot. A prototype was developed that combines report scraping, statement extraction, vector storage, and a web interface. Validation was conducted both manually and using a second LLM, confirming relevance and metadata accuracy. While the results were promising, challenges such as occasional misclassification, report retrieval issues, and the absence of a gold-standard dataset for testing, were noted. Still, the project demonstrates the viability of using LLMs for financial text analysis and highlights future development directions, including continuous data collection and analysis, improved model evaluation, and expanded chatbot functionality.