Hybrid machine learning and stochastic volatility models with blockchain data for high-frequency cryptocurrency trading

Abstract

High-frequency cryptocurrency markets, particularly for Bitcoin and Ethereum, are characterized by extreme volatility with daily price fluctuations often surpassing 10%. Traditional stochastic volatility models, such as the Heston model, prove inadequate in capturing the nonlinear dynamics and abrupt price movements driven by blockchain events. This study develops an innovative hybrid Heston-LSTM model enhanced with real-time blockchain metrics, including transaction counts and gas fees, effectively combining Heston's theoretical foundations with LSTM's ability to detect complex temporal patterns and on-chain market signals. Using high-frequency 1-min Bitcoin price data spanning January to March 2025 (129,600 observations), our model demonstrates significant improvements, achieving a 43% reduction in mean squared error compared to the standalone Heston model and a 20% enhancement over pure LSTM approaches. In high-frequency trading simulations, the framework generates 18.5% cumulative returns with a Sharpe ratio of 2.1 while maintaining a minimal maximum drawdown of 4.2%, substantially outperforming all benchmark models. The incorporation of SHAP analysis ensures model interpretability, addressing critical regulatory requirements for algorithmic transparency. The model's practical utility is particularly evident in the Asia–Pacific markets, where it successfully adapts to region-specific phenomena, such as South Korea's Kimchi premium, and complies with Japan's stringent exchange regulations. Robustness is further confirmed through extensive testing with historical data from 2023 to 2024, validating the model's performance across diverse market conditions. These results establish our hybrid framework as both a theoretically sound and practically valuable tool for cryptocurrency volatility forecasting and high-frequency trading.