A novel framework for cryptocurrency price forecasting: integrating dual attention mechanisms, genetic algorithm feature selection, and Hybrid Adam-PSO optimization

This paper proposes a robust framework for cryptocurrency price forecasting by integrating technical indicators, genetic algorithm based feature selection, and hybrid deep learning models. Technical indicators capture historical patterns, while a genetic algorithm optimizes features for Bitcoin, Ethereum, and Litecoin. Four hybrid models are introduced: Convolutional Neural Network Long Short Term Memory Gated Recurrent Unit with Attention, Convolutional Neural Network Gated Recurrent Unit Long Short Term Memory with Attention, Convolutional Neural Network Bi Long Short Term Memory Bi Gated Recurrent Unit, and Convolutional Neural Network Bi Gated Recurrent Unit Bi Long Short Term Memory with Dual Attention. These architectures combine Convolutional Neural Networks for feature extraction with Long Short Term Memory and Gated Recurrent Unit networks for temporal modeling, while attention mechanisms emphasize critical time steps influencing prediction. A dual attention mechanism further improves prediction by applying both spatial and temporal attention. Hyperparameters are optimized using a hybrid Adam Particle Swarm Optimization strategy that balances local and global search. Evaluation metrics include Mean Squared Error, Mean Absolute Error, and R squared. Results show the CNN BiLSTM BiGRU model performs best for Ethereum with an MSE of 0.0002 and for Litecoin with 95.72 percent accuracy, while the CNN LSTM GRU with Attention model achieves the highest accuracy for Bitcoin at 99.44 percent. Model interpretability is supported using LIME, providing insights for traders. The proposed framework demonstrates statistically significant improvements with p values less than 0.05 and shows potential for future enhancement through integration of sentiment and macroeconomic indicators.