Cryptocurrency Volatility Forecasting Using Transformer-Based Deep Learning Models and On-Chain Metrics

Abstract

Cryptocurrencies have emerged as highly dynamic digital assets, characterized by extreme price volatility and driven by both speculative behavior and network-level activities. Traditional volatility forecasting methods, including GARCH and LSTM-based models, often fall short in capturing the complex, nonlinear, and temporal dependencies inherent in crypto markets. This paper proposes a novel deep learning framework that leverages Transformer-based architectures originally designed for natural language processing to forecast short-term cryptocurrency volatility with enhanced accuracy and temporal sensitivity. By incorporating a comprehensive set of on-chain metrics (such as transaction volume, wallet activity, miner behavior, and token circulation), our model captures both market sentiment and blockchain-level dynamics. The proposed Transformer model is benchmarked against LSTM and GRU networks using Bitcoin and Ethereum datasets spanning multiple market cycles. Experimental results show that the Transformer-based model outperforms recurrent architectures in predicting both realized and implied volatility, particularly during high-turbulence periods. These findings suggest that attention mechanisms, combined with on-chain data, provide a powerful tool for managing risk and making informed decisions in the rapidly evolving digital asset ecosystem.