Can machine learning models better volatility forecasting? A combined method

Volatility forecasting for Bitcoin has garnered increasing attention due to heightened investment interest and the inherent risks associated with cryptocurrencies. Traditional forecasting models, such as the Generalised Autoregressive Conditional Heteroskedasticity (GARCH) family models, are widely employed. However, there is a need for careful consideration regarding their ability to capture extreme shocks and long-term volatile features. In this study, we fit several GARCH models, with the Exponential GARCH model demonstrating the best goodness of fit. We further utilise their volatility observations for an automated forecasting solution, using the Long Short-Term Memory (LSTM) neural network for predictions. Our results indicate a significant clear improvement in volatility forecasting regarding both the model’s in-sample and out-of-sample accuracy. Notably, the LSTM model optimises information intake through its short- and long-memory states. Overall, our novel LSTM neural network model is more robust in responding to market shocks and regime changes.