Abstract
Accurate cryptocurrency price forecasting represents an important yet challenging problem in financial time-series analysis due to the highly volatile, nonlinear, and noise-sensitive nature of digital asset markets. Although transformer-based architectures have recently demonstrated strong capabilities in temporal sequence modeling, their behavior under high-frequency cryptocurrency dynamics and the role of explainability-guided feature refinement remain insufficiently explored. To address this gap, this study presents a comprehensive transformer-based forecasting framework for hourly cryptocurrency price prediction and investigates the impact of explainability-guided feature optimization on forecasting performance, robustness, and interpretability. Five transformer architectures—Vanilla Transformer, Informer, Autoformer, Reformer, and Temporal Fusion Transformer (TFT)—are systematically evaluated across five major cryptocurrency assets: Bitcoin (BTC), Ethereum (ETH), Solana (SOL), Dogecoin (DOGE), and Ripple (XRP). The experimental framework employs Open, High, Low, Close, and Volume (OHLCV) data together with a broad set of engineered technical indicators and evaluates model performance using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Mean Squared Error (MSE), Mean Absolute Percentage Error (MAPE), and the coefficient of determination (R2). To improve interpretability and reduce feature redundancy, SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) are integrated directly into the forecasting pipeline. Based on the resulting explanations, asset-specific feature subsets are constructed, and all models are subsequently retrained using the refined feature representations. The results show that explainability-guided feature refinement provides compact, model-aware, and interpretable feature subsets with competitive forecasting performance; however, its effect on prediction accuracy is dependent on the cryptocurrency asset, transformer architecture, retained feature subset size, and market conditions. Additional robustness, sensitivity, alternative feature-selection, and statistical significance analyses indicate that the SHAP–LIME Top-15 subset should be interpreted as a conservative dimensionality-reduction strategy rather than a universally optimal feature-selection rule. The findings further reveal that transformer architectures incorporating sparse attention, decomposition mechanisms, or gating structures generally provide stronger performance than the Vanilla Transformer under highly volatile hourly market conditions. Overall, the proposed framework demonstrates that combining transformer-based forecasting with explainability-guided feature refinement can support interpretable and parsimonious high-frequency financial time-series modeling, while highlighting the importance of evaluating robustness, feature-selection sensitivity, and statistical variability alongside average forecasting errors.