Empirical Essays On Volatility Forecasting

Abstract

This thesis aims to examine and improve forecasting performance for both univariate volatility and multivariate covariance models. This thesis investigated the forecasting ability of volatility models and covariance models including univariate GARCH methods, HAR models, and multivariate DCC process on stock indices in twelve countries. Moreover, several hybrid models combined by the current GARCH genres and neural networks are investigated in three empirical exercises. The accuracy of forecasting by different models is addressed. There are four main contributions of this study. First, the comparison among the univariate normal GARCH genre, HAR model and hybrid models by neural networks reveals that the hybrid models are superior to others which gives an empirical result in a wide comparison. The policymakers can benefit from the results to formulate their policies to avoid risk. Second, with the application of DCC process, the new multivariate model built by neural networks are preferred rather than original DCC GARCH models when forecasting covariance which give some empirical results on multivariate covariance forecasting. The results are able to provide some suggestions for market managers on risk control, especially for the portfolios containing multivariate assets in different countries. Third, the trading volume is found to be useful for improving volatility forecasting in the hybrid process. Finally, the original neural networks are improved by a deep learning model which has more hidden layers than the previous neural networks. The forecasting ability of all the models are investigated and the hybrid model built with deep learning are still superior. This research provides valuable insights and a reliable framework for improving stock volatility predictions.