Analyse der kontextabhängigen Wirksamkeit von Social-Media-Sentiments in KI-basierten Aktienkursprognosen

Abstract

This master’s thesis investigates the impact of social media sentiment data on the predictive accuracy of stock price models. The starting point is the assumption that traditional pricebased approaches only partially capture the dynamics of modern financial markets, which are increasingly driven by sentiments, collective expectations, and media narratives. The aim of the study was therefore to examine whether, and under which conditions, sentiment data can complement established forecasting approaches. Following the Design Science Research methodology, a forecasting framework was developed that integrates Long Short-Term Memory (LSTM) networks with social media data. The empirical evaluation focused on eight S&P 500 constituents and two non-index stocks (GameStop, AMC) over the period 2021–2025. GameStop and AMC were included because of their exceptional relevance within the retail investor community and their role in the public sentiment and meme stock discourse, making them a particularly suitable test case for sentiment-driven forecasting models. In addition to a baseline model using only price data, sentiment-augmented variants were tested and compared across different forecasting horizons and market phases. Furthermore, macroeconomic indicators, trading volumes, and stock-specific factors were included in the analysis. The results reveal that sentiment data does not provide a universal improvement. Enhancements were observed selectively, particularly in short-term forecasting windows and for stocks with high media attention. In sideways markets, however, the added value of sentiment was weak or even negative. Macroeconomic variables and trading volume consistently emerged as the most robust factors across both time horizons, whereas sentiment measures proved most relevant in highly volatile and speculative contexts. Overall, the thesis contributes to the discussion on the role of alternative data sources in financial forecasting. It demonstrates that sentiment data should be applied selectively and that its effectiveness is highly context-dependent. Both research and practice may benefit from advancing hybrid modeling approaches that integrate fundamental financial variables with sentiment indicators to improve the understanding of dynamic market behavior.

Date of Award	2025
Original language	German (Austria)
Awarding Institution	Johannes Kepler University Linz
Supervisor	Martin Stabauer (Supervisor)