Dynamics in Realized Volatility Forecasting: Evaluating GARCH Models and Deep Learning Algorithms Across Parameter Variations

The modeling and forecasting of return volatility for the top three cryptocurrencies, which are identified by the highest trading volumes, is the main focus of the study. Eleven different GARCH-type models were analyzed using a comprehensive methodology in six different distributions, and deep learning algorithms were used to rigorously assess each model's forecasting performance. Additionally, the study investigates the impact of selecting dynamic parameters for the forecasting performance of these models. This study investigates if there are any appreciable differences in forecast outcomes between the two different realized variance calculations and variations in training size. Further investigation focuses on how the use of expanding and rolling windows affects the optimal window type for forecasting. Finally, the importance of choosing different error measurements is emphasized in the framework of comparing forecasting performances. Our results indicate that in GARCH-type models, 5-minute realized variance shows the best forecasting performance, while in deep learning models, median realized variance (MedRV) has the best performance. Moreover, it has been determined that an increase in the training/test ratio and the selection of the rolling window approach both play important roles in achieving better forecast accuracy. Finally, our results show that deep learning models outperform GARCH-type models in volatility forecasts.