Sentiment-driven cryptocurrency forecasting: analyzing LSTM, GRU, Bi-LSTM, and temporal attention model (TAM)
Phumudzo Lloyd Seabe, Claude Rodrigue Bambe Moutsinga, Edson Pindza
Abstract
Abstract Predicting cryptocurrency prices is challenging due to market volatility and external influences like social media sentiment. This study integrates Twitter sentiment analysis with deep learning models (LSTM, GRU, Bi-LSTM, and Temporal Attention Model) to enhance Bitcoin price forecasting. Sentiment features were extracted using VADER and RoBERTa, with findings showing that RoBERTa-based models significantly outperform VADER. Bi-LSTM (RoBERTa) achieved the lowest MAPE of 2.01%, demonstrating the effectiveness of deep contextual embeddings. SHAP analysis identified Sentiment Momentum, RoBERTa Compound Score, and VADER Negativity Score as key predictors of price movements. These results highlight the value of sentiment-driven forecasting and provide insights for traders, investors, and researchers.