Percorrer por autor "Janesch, Daniel Malvin"
A mostrar 1 - 1 de 1
Resultados por página
Opções de ordenação
- Forecasting next-day market prices of stocks and bitcoin using social media sentiment analysisPublication . Janesch, Daniel Malvin; Fernandes, Pedro AfonsoThis thesis investigates whether sentiment scores derived from social media can improve the prediction of next-day market prices for financial assets, with a focus on Tesla ($TSLA) and Bitcoin ($BTC). Using freely available X (Twitter) datasets from Kaggle, three pre-trained Natural Language Processing (NLP) tools—VADER, TextBlob, and FinTwitBERT—were appplied to individually score the sentiment of each post. After extensive cleaning, filtering, and sampling, a total of 47,800 tweets were used for analysis—33,200 for $TSLA and 14,600 for $BTC. Sentiment scores were aggregated into daily averages and merged with financial data from Bloomberg. These features were then used as explanatory variables in three forecasting models: ARIMA/ARIMAX, XGBoost, and Long Short-Term Memory (LSTM) neural networks. The results show that for $TSLA, the introduction of sentiment scores—especially with FinTwitBERT—substantially improved forecasting performance. The strongest model, an LSTM with sentiment and financial variables, achieved a test R² of 0.76 compared to a baseline model with R² of 0.49. In contrast, all models performed poorly on $BTC, likely due to its higher volatility, optimistically biased sentiment, and large data gaps (about 50% of tweet days missing due to structural issues). These findings suggest that sentiment analysis can enhance short-term price forecasting for traditional stocks, while its value for highly volatile assets like Bitcoin remains unclear. Additionally, the results highlight the superior predictive performance of transformer-based sentiment models like FinTwitBERT over lexicon-based alternatives. This study lays the groundwork for future research in real-time or intraday prediction using more complex models and diverse social media sources.