Cryptocurrency price volatility prediction based on GARCH-LSTM: A multi-currency empirical study

Abstract

This paper proposes a hybrid prediction framework based on GARCH-LSTM to address the high volatility, nonlinearity, and structural complexity of cryptocurrency price volatility. The framework first extracts conditional volatility features using GARCH-class models (GARCH, EGARCH, GJR-GARCH), then combines historical returns and technical indicators with LSTM for nonlinear prediction. A dynamic residual correction mechanism is introduced to eliminate autocorrelation in prediction errors. Empirical analysis using daily price data from Bitcoin, Ethereum, Binance Coin, and Dogecoin demonstrates that EGARCH performs best in single-currency volatility characterization. While the original GARCH-LSTM model can stably capture trends, it exhibits significant residual autocorrelation. After AR (1) correction, RMSE, MAE, and MAPE all show substantial improvements, enhancing prediction accuracy and statistical validity. Cross- currency comparisons reveal strong generalization across major cryptocurrencies, though optimization potential remains for high-volatile currencies. This framework provides valuable insights into digital asset risk management and unified multi-currency prediction.