Stacking hybrid GARCH models for forecasting Bitcoin volatility

Machine learning techniques have been used frequently for volatility forecasting. However, previous studies have built these hybrid models in a form of a first-order GARCH(1,1) process by following general use for GARCH models. But the way of estimating parameters for GARCH and machine learning models differs considerably. Hence, we have investigated the effect of different model orders of the GARCH process on the volatility forecasts of Bitcoin obtained by the four most used machine learning models. Furthermore, we have proposed a stacking ensemble methodology based on GARCH hybrid models to improve the results further. The proposed stacking ensemble methodology utilizes the techniques of feature selection and feature extraction to reduce the dimension of the predictors before meta-learning. The results show that using higher model orders increases the accuracy of volatility forecasts for hybrid GARCH models. Also, the proposed stacking ensemble with LASSO produces forecasts superior to almost all hybrid models and better than the ordinary stacking ensemble.