Integrating Econometric and Deep Learning Approaches for Ethereum Volatility Prediction: A GARCH–LSTM Framework

This study investigates the dynamic behaviour of Ethereum’s price volatility by integrating econometric and deep learning approaches. Using daily price data from July 2015 to December 2023, the research models Ethereum’s market fluctuations through both the Generalized Autoregressive Conditional Heteroskedasticity (GARCH) model and the Long Short-Term Memory (LSTM) neural network. The dataset was transformed into daily log returns, and a 30-day rolling volatility measure was used to capture short-term market variation. The empirical results reveal that Ethereum exhibits significant volatility persistence and clustering, consistent with the stylized facts of financial time series. The analysis identified several major volatility spikes during the 2017 speculative boom, the 2018 correction, and the 2021 bull run, where annualized volatility levels exceeded 1.5. In contrast, calmer phases, such as 2019 and late 2022, averaged around 0.4. Regime classification showed that 32 percent of the period was characterized by low volatility, 40 percent by moderate volatility, and 28 percent by high volatility. The findings indicate that volatility in Ethereum’s market is nonlinear, heteroskedastic, and strongly path-dependent, suggesting that traditional linear models alone are insufficient for accurate prediction. The integration of GARCH for short-term volatility modelling and LSTM for capturing nonlinear and long-term dependencies provides a more comprehensive framework for understanding and forecasting digital asset risk behaviour.