Cryptocurrency price prediction is challenging due to high volatility, non-stationarity, and complex interactions among diverse data sources. Existing models often fail to effectively integrate on-chain metrics, sentiment data, and technical indicators, while lacking adaptive attention mechanisms and uncertainty estimation. To address these limitations, this paper proposes a Multi-Mode Adaptive Attention Network (M2A2-Net), which combines Temporal Fusion Transformers, dynamic cross-attention, and probabilistic GRUs to model multimodal temporal dependencies. The framework incorporates 47 heterogeneous features, including blockchain metrics (e.g., hash rate, active addresses), sentiment embeddings, macroeconomic indicators, and technical patterns. A regime-adaptive attention mechanism dynamically adjusts feature importance based on market conditions, while Monte Carlo dropout provides uncertainty estimation. The model is evaluated on six major cryptocurrencies across multiple market phases spanning from early 2022 to the end of 2025. Results demonstrate competitive performance, achieving a mean absolute percentage error of 1.87% for Bitcoin and 2.34% for Ethereum, showing improvements over baseline models such as Transformers, LSTM-XGBoost, and BiGRU. Additionally, the probabilistic framework ensures reliable prediction intervals with 94.3% coverage. SHAP analysis highlights that sentiment and on-chain features dominate during volatile periods, while technical indicators are more influential in stable markets, confirming the model’s effectiveness and interpretability.
Lu et al. (Fri,) studied this question.