Analysis and Forecasting of Bitcoin Price Volatility: A Deep Learning Approach Using DNN, LSTM, Transformers, and the ARMA-GARCH Model

Abstract

This study evaluates the performance of deep learning models and a traditional econometric approach in forecasting cryptocurrency price movements using the BTC and BNB datasets. Four models are compared: recurrent neural network (RNN), long short-term memory (LSTM), transformer, and GARCH, under two loss functions: log-likelihood and mean squared error (MSE). The results show that BTC and BNB datasets exhibit similar stability, with only minor differences in predictive accuracy. Models trained with log-likelihood loss consistently outperform those using MSE, achieving lower MAE and RMSE, higher R2, and improved probabilistic modeling. Among all models, RNN-loglik delivers the best performance, accurately capturing both short-term volatility and long-term trends, while LSTM and transformer perform moderately well and GARCH underperforms significantly. These findings demonstrate the effectiveness of deep learning combined with probabilistic loss functions for cryptocurrency forecasting, supporting applications in algorithmic trading, portfolio optimization, and risk management.