CS33: LLMs and ML in Dynamic Risk Control

Organizer: Wolfgang Karl Härdle (Humboldt-Universität zu Berlin)

LLM-Driven Stock Movement Prediction

Daniel Traian Pele

Stock price forecasting remains inherently difficult due to the volatility of financial markets and the impact of external factors such as investor sentiment and macroeconomic events. Traditional statistical models often fall short in capturing the nonlinear dynamics and unstructured signals embedded in textual data. This study explores the use of Large Language Models (LLMs), specifically the LLaMA family, to enhance predictive performance by integrating historical price data with financial news. We compare LLaMA 3.3 with LLaMA 3.1 and a benchmark ARIMA model, evaluating their ability to capture both time series patterns and sentiment-driven signals. Results show that LLaMA 3.3 significantly outperforms the alternatives, confirming the added value of incorporating news sentiment into the forecasting process. The findings underscore the potential of LLM-based approaches to improve market predictions by fusing structured and unstructured data, with implications for traders, analysts, and financial institutions. Future research will focus on domain-specific fine-tuning and enhancing model interpretability.