A new hybrid approach combining transformer-based language models and graph neural networks for cryptocurrency forecasting

Abstract

Due to the highly speculative and volatile nature of digital asset markets, financial price prediction has recently attracted much attention. This study presents a new model based on Graph Neural Networks and enhanced with sentiment analysis to predict the future prices of Bitcoin and Ethereum. It provides a more comprehensive forecasting framework than traditional machine and deep learning models by accurately modeling price fluctuations, market trends, and the complex relationship between sentiment and the market. When tested on real-world cryptocurrency datasets, the proposed model achieves performance metrics such as Root Mean Square Error of 0.0132 and Coefficient of Determination of 0.9941 for Bitcoin, Root Mean Square Error of 0.0159, and Coefficient of Determination of 0.9997 for Ethereum. The Mean Absolute Errors of 0.0097 and 0.0119, and the Mean Absolute Percentage Errors of 0.0027 and 0.0029, respectively, further confirm the models' strong prediction strength and generalization across different market conditions. These results demonstrate how well the model generalizes across periods and market conditions and how accurate its forecasts are. This study demonstrates the advantage of integrating sentiment data with a hybrid Long Short-Term Memory-Graph Neural Network architecture for cryptocurrency price forecasting by outperforming individual Long Short-Term Memory, Gated Recurrent Unit, Graph Neural Network, and other deep learning models across Root Mean Square Error, Mean Absolute Error, Mean Absolute Percentage Error, and Coefficient of Determination metrics.