VOLATILITY DYNAMICS IN CRYPTOCURRENCIES: COMPARING TRADITIONAL AND HYBRID DEEP LEARNING PREDICTIVE TECHNIQUES
Tam Phan Huy, Khuong Huu Pham, Linh Nguyen Thi Thuy, Nam Nguyen Nhat, Bach Le Danh
University of Economics and Law and Vietnam National University, Ho Chi Minh City, Vietnam
Abstract: This study investigates the predictive performance of traditional econometric and deep learning models in forecasting the return and volatility of Ethereum, a leading cryptocurrency known for its high price fluctuation and market sensitivity. Using historical daily price data from January 2018 to March 2025 obtained from Yahoo Finance, we construct a comprehensive set of features including log returns, multiple volatility estimators (Rolling, Parkinson, Garman-Klass, and Rogers-Satchell), and technical indicators such as moving averages, momentum oscillators, and volume-based metrics. Six predictive models—ARIMA, GARCH, VAR, Random Forest, LSTM, and GRU—are evaluated based on their ability to predict log returns and volatility using multiple performance metrics: Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and R-squared (R²), along with cross-validation to assess overfitting. The results reveal that machine learning and deep learning models significantly outperform traditional econometric methods. Random Forest yields the highest accuracy in volatility prediction, while GRU demonstrates the most stable and consistent performance across both return and volatility forecasts. Traditional models, particularly GARCH, show poor generalization and limited predictive capability under high market uncertainty. The findings highlight the importance of model selection and volatility measure choice in cryptocurrency forecasting. This study contributes to the growing literature on crypto analytics by providing a direct comparative evaluation of predictive models and offering practical insights into model suitability for high-frequency, volatile digital asset markets.
Keywords: Ethereum, volatility forecasting, deep learning, machine learning.
JEL codes: C22, C45, G17
VOLUME 9 ISSUE 11 2025: 60 – 75