A Comparative Study of GARCH and Deep Learning Models in Predicting Bitcoin Daily Returns

Abstract

Bitcoin has rapidly emerged as a focal point for the media, investors, and researchers due to its prominent role as an investment alternative to traditional currencies. However, its marked price volatility presents notable risks, particularly for organizations with significant Bitcoin holdings. To manage these risks effectively and enhance trading insights, accurate forecasting of Bitcoin's price fluctuations is crucial. This study presents a comparative analysis of Generalized Autoregressive Conditional Heteroskedasticity (GARCH) and Deep Learning (DL) models in predicting daily Bitcoin returns, aiming to identify the most effective approach for this highly volatile asset. Three variations of GARCH models, Exponential GARCH (EGARCH), Threshold GARCH (TGARCH), and Glosten-Jagannathan-Runkle GARCH (GJR-GARCH) were utilized, each evaluated under three distributional assumptions: Normal, student t, and Skewed student t. Additionally, four DL models Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), Multilayer Perceptron (MLP), and Gated Recurrent Unit (GRU) were implemented to assess the efficacy of neural networks in volatility and return prediction. Model performance was measured using Root Mean Square Error (RMSE) and Root Mean Square Percentage Error (RMSPE), The results show that EGARCH with t-distribution achieved the lowest RMSE (2.720574) among the GARCH models, while MLP had the best overall performance among deep learning models with an RMSE of 2.731826 and the lowest RMSPE (3.431030) across all models. These findings indicate that both GARCH and DL models offer valuable insights, with EGARCH and MLP excelling in different performance metrics, suggesting complementary benefits in predicting Bitcoin returns.

International Journal of Statistical Sciences, Vol. 25(2), November, 2025, pp 173-178