| This article compares the predictive capability of heterogeneous autoregressive type(HAR-Type)models for forecasting the realized volatility of stocks in different industry.The industry classification is based on the Global Industry Classification System(GICS).In this paper,215 US-listed shares high-frequency data of 6816 trading days from January 2,1990 to February 28,2017 are used as empirical data.We employ a rolling window of 1000 observations to estimate the regression coefficients of HAR-Type model,random walk model and the first-order autoregressive model with macro-variables and then assess their out of sample performance in the remainder of the sample by looking at short-term(h=1),mid-term(h=5,10)and long-term(h=22)ahead forecasts.Using the mean squared forecast error(MSE)and the mean absolute forecast error(MAE)as loss function respectively,the optimal model of each stock can be determined by Giacomini-White Test.By statistical analysis of the best model classified by industries,we can draw four conclusions.First,there are some difference in realized volatility by different industries.The correlation between realized volatility of different individual stocks in the same industry is higher than that outside the industry.Second,the volatility of stocks has obvious jumping characteristics.The threshold bipower variation(TBPV)can effectively separate the continuous components and the jumping components from realized volatility.Third,the best models in different industries at the same forecast horizon are inconsistent.The optimal models in the same industries but at different forecast horizons are also inconsistent.When forecasting the realized volatility of individual stocks,we should select the optimal model according to the industry and the forecast horizon.Finally,when forecasting the short term volatility of individual stocks,adding the macro variables to HAR-Type models can improve the predictive stability.In different industries,we should add specified macro-variables which are significant,and the number of macro-variables should not be too many.At the end of this paper,we provide the optimal models for different industries under different forecast horizons. |
