engleski

Realized variance forecasting with high-frequency data

Different types of investors, weather they are small, individual investors or big, institutional investors, are all interested in volatility forecasting in order to get higher profits or less risky positions. Most commonly used models for forecasting volatility of daily stock returns are the generalized autoregressive conditional heteroscedasticit (GARCH) models. Since financial data show very strong persistence due to autocorrelation of the squared and absolute returns and return probability density functions are leptokurtic with shapes depending on the time scale ; neither ARCH nor GARCH models are able to reproduce all of them. To escape this problem, researchers used proxies instead of the unobservable variance for their out-of-sample forecasts. However, when the intra-daily financial data became available a new measure that could be used instead of the usual squared return proxy was proposed. With the assumption that there are jumps in the intraday price process, in this paper the extended HAR-RV- model is discussed. However, these models cannot account for nonlinear behaviour of realized variance. Therefore, the objective of this paper is to develop a parsimonious neural networks (NN) model which can capture the main stylized facts of realized volatility. The goal is to develop a neural network with appropriate connection in the context of nonlinear HAR-RV and HAR-RV-J model, i.e. HAR- RV-NN and HAR-RV- J-NN models respectively. Out of-sample forecasts will be compared to determine their predictive accuracy with application to DAX stock market index. The contribution of this paper can be seen in determining the appropriate NN that is comparable to HAR-RV and HAR-RV-J models and its application in forecasting realized variance. This way the investors are able to apply the chosen model and obtain the most accurate realized volatility forecasts for their potential investments.

Realized variance ; forecasting ; neural network ; high-frequency data ; nonlinearity

nije evidentirano