Role of high-frequency data, distribution assumption and trading volume in volatility forecasting in China stock market

Volatility forecasting has become a crucial process in risk management over recent decades. With the second largest stock market by market capitalization in 2019, China has gained increasing attention from recent research. This study aims at providing better volatility forecasts by investigating the role of high-frequency data, distribution assumption and trading volume in volatility forecasting based on the China stock market. The behavior of high-frequency data in financial markets highly relates to market efficiency and information flow. The heterogeneous market hypothesis (HMH) is in response to the behavior of non-homogeneous market participants. In contrast to Efficient Market Hypothesis (EMH), HMH states that investors interpret information flow differently. Particularly, on a short-term basis, such as minute to minute, speculative behavior dominates the markets. In this regard, the study investigates the role of intraday data in volatility forecasting by using Generalized Autoregressive Conditional Heteroskedasticity (GARCH) model. Besides, regarding the non-normal distribution of financial time series, a variety of distribution assumptions are incorporated in application. Furthermore, to examine the role of trading volume in volatility forecasting and test the validity of two conflicting hypotheses: the Mixture of Distribution Hypothesis (MDH) and the Sequential Information Arrival Hypothesis (SIH), trading volume is regarded as both long-run and short-run predictors by this research. The considered methods contain the GARCH family model, the Heterogeneous Autoregressive (HAR) family model, the Smooth Transition Exponential Smoothing (STES), the Autoregressive Fractionally Integrated Moving Average (ARFIMA), and the GARCH-MIDAS model. In particular, in GARCH application, both intraday returns and daily returns are used and estimated under normal and non-normal distribution ii assumptions. The contributions of this study are that: (1) it provides clear evidence to support that the superiority of traditional time series models in volatility forecasting remains by taking advantage of high-frequency data; (2) it incorporates different distribution assumptions in GARCH models to capture the stylized facts of highfrequency data; (3) it makes the first attempt to evaluate the performance of STES in volatility forecasting by using RV as the proxy of actual volatility; (4) it provides a more consistent comparison to evaluate the forecasting ability of a mixed data sampling approach; (5) it extends the literature on the forecasting performance of trading volume to the GARCH-MIDAS approach. The empirical results show that: (1) data frequency in GARCH application substantially influence the accuracy of volatility forecasting, as the higher the frequency is of the return series, the better are the forecasts provided; (2) non-normal distributions are more capable at reproducing the stylized facts of both intraday and daily return series than normal distribution; (3) GARCH estimated by 5-min returns not only outperforms other GARCH alternatives, but also considerably beats RV-based models and STES at volatility forecasting; (4) no clear evidence appears that SIH holds in the China stock market; (5) GARCH-MIDAS is not able to beat the traditional GARCH method when both are estimated by the same predictors sampled at different frequencies.

Text SPE 2022 30 UPM IR.pdf Download (1MB)

Actions (login required)

View Item