Predicting token prices in the blockchain market using neural networks.

Area of Science:

• * Computational finance and artificial intelligence applied to financial markets.
• * Development of advanced machine learning models for time-series forecasting.

Background:

• * Cryptocurrency price prediction faces challenges from high volatility, non-stationarity, and complex data interactions.
• * Existing models struggle to integrate on-chain metrics, sentiment data, and technical indicators effectively.
• * Lack of adaptive attention mechanisms and uncertainty estimation hinders model performance.

Purpose of the Study:

• * To propose a novel Multi-Mode Adaptive Attention Network (M2A2-Net) for enhanced cryptocurrency price prediction.
• * To effectively integrate multimodal temporal dependencies using heterogeneous features.
• * To provide uncertainty estimation and improve model interpretability.

Main Methods:

• * Development of M2A2-Net combining Temporal Fusion Transformers, dynamic cross-attention, and probabilistic GRUs.
• * Integration of 47 heterogeneous features: on-chain metrics, sentiment embeddings, macroeconomic indicators, and technical patterns.
• * Implementation of a regime-adaptive attention mechanism and Monte Carlo dropout for uncertainty estimation.

Main Results:

• * M2A2-Net achieved competitive performance with a mean absolute percentage error of 1.87% for Bitcoin and 2.34% for Ethereum.
• * Demonstrated significant improvements over baseline models like Transformers, LSTM-XGBoost, and BiGRU.
• * Probabilistic framework ensured reliable prediction intervals with 94.3% coverage.

Conclusions:

• * M2A2-Net effectively models multimodal temporal dependencies for cryptocurrency price prediction.
• * Regime-adaptive attention and uncertainty estimation enhance model robustness and reliability.
• * SHAP analysis confirmed the dynamic influence of different features based on market conditions, highlighting model interpretability.