An Investigation of The Contribution of Attention-Based Hybrid Deep Learning Models to Prediction Errors in Cryptocurrency Markets

This study comparatively analyzes deep learning architectures for predicting the daily closing prices of Ethereum, Solana, and BNB cryptocurrencies. Given the high volatility, non-linear price movements, and noise-intensive data structures of cryptocurrency markets, accurate price modeling holds significant importance for both academic research and investment decisions. The models examined include LSTM, GRU, CNN, CNN-LSTM, CNN-GRU, Attention-LSTM, and Attention-GRU. The dataset covers the period from April 10, 2020, to February 12, 2026, at daily frequency, split into 70% training and 30% testing. Hyperparameter optimization was conducted across varying window lengths, epochs, batch sizes, neuron counts, and dropout rates. Model performance was evaluated using MAE, MSE, MAPE, and R² metrics. Results indicate that GRU-based models consistently yield lower error values and more stable performance on cryptocurrency time series. The hybrid CNN-GRU architecture produces competitive results on certain series; however, attention mechanism-integrated models generally increase error values. These findings suggest that model complexity does not guarantee superior prediction accuracy in financial time series, and that compact architectures may offer stronger generalization in highly volatile data environments. The study also addresses a methodological gap by benchmarking multiple deep learning models under identical datasets and experimental conditions.