Deep Learning for Cryptocurrency Markets: Integrating Bidirectional LSTM and Graph Attention Networks to Predict Ethereum Prices

Abstract

Predicting cryptocurrency prices with precision is crucial for strategic financial planning, enabling stakeholders to mitigate risks in the highly unpredictable nature of digital assets. This paper presents an innovative framework combining Bidirectional Long Short-Term Memory (Bi-LSTM) and Graph Attention Networks (GATs) to improve forecasting accuracy for Ethereum. The Bi-LSTM analyzes time-based trends in historical price and trading volume over a 90-day horizon, whereas GATs examine correlations between critical market features, including 20-day and 30-day moving averages, through attention-focused techniques. When applied to Ethereum's historical price data, the model achieves an MSE of 0.0021, RMSE of 0.046, and MAE of 0.032, exceeding traditional LSTM-based approaches. These outcomes highlight the advantages of fusing sequential neural architectures with graph-structured relational modeling to refine predictive accuracy. By unifying time-series and graph-structured data analysis, this study contributes to the advancement of financial analytics powered by deep learning, equipping traders and researchers with an actionable tool to refine trading tactics within Ethereum's dynamic ecosystem.