<![CDATA[The cryptocurrency market has grown into a multi-trillion-dollar domain with extreme volatility. This paper addresses the forecasting of crypto price movements and volatility by integrating market metrics with sentiment analysis. We identify a gap in existing studies, which often ignore multi-source sentiment and thus miss early warning signs of volatility. We propose a Sentiment-Aware Transformer model inspired by the Temporal Fusion Transformer (TFT). The model ingests daily price, volume, and market cap features from CoinMarketCap alongside aggregated sentiment scores from Twitter, Reddit, and financial news (extracted via FinBERT). We train and evaluate the model on 5 years of data for 10 major cryptocurrencies (2020–2024), comparing performance against LSTM and GRU baselines with identical inputs. The proposed Transformer achieves 83.2% volatility prediction accuracy with an F1-score of 0.81, exceeding the LSTM (79% accuracy) and GRU (80%) baselines. It also shows the lowest RMSE in price forecasting and a higher return correlation (0.72) with actual prices, indicating improved trend alignment. These gains are statistically significant (p<0.01). We also discuss how attention weights offer interpretability, as the model focuses on sentiment spikes during impending volatility.]]>
Copyrights © 2025
