Nonparametric Volatility Modelling,Forecasting And Evaluation:A High-Frequency Data Perspective

Modelling and forecasting the volatility of financial assets are the hot and hard issues in the academic and practical fields of modern finance, because they are closely linked to the test of asset pricing theory, the selection of optimal asset portfolio, the hedging strategy of derivatives and the measurement and management of financial risks. Since more than a decade, with the rapid development of computer technology and increasing acquisition of financial high-frequency data, the measurement and forecast of volatility based on intraday high-frequency data have received an extensive attention from scholars at home and abroad. Among the nonparametric high-frequency volatilities, the biggest concern from academic and practical fields is the realized volatility and realized range volatility. Therefore, the dissertation uses the heterogeneous autoregressive model for realized volatility (HAR-RV) and realized range volatility (HAR-RRV) on the basis of the heterogeneous market hypothesis, as the basic volatility models. First, the article compares them with traditional volatility models in terms of forecast ability, and emphatically studies the jump and signed jump variation's impact on the estimation and forecast of them. Furthermore, on the strength of jump and signed jump variation, the article further presents the newly-constructed signed jump variation including jump, and builds the new volatility model. Then the dissertation makes a comparison with new volatility models in the field of predictive ability. Based on the aforementioned research, this dissertation further uses the Markov regime-switching mechanism to expand the models with better forecast effect, and analyses whether they can create economic value or not on the basis of portfolio. This dissertation chooses to use the out-of-sample rolling forecast technique and precise and novel Model Confidence Set (MCS) method. The main empirical results are listed as follows.(1) Compared to several traditional volatility models mentioned in the chapter (such as ARFIMA-LnRV, GARCH model), the HAR-RV model has a better performance in describing and forecasting the volatility in the Chinese stock market. And compared to other volatility models mentioned in this dissertation, the HAR-RRV model and ARFIMA-RRV model both have better performances in forecasting the volatility in the Chinese stock market. But through the empirical results, this chapter finds that the HAR-RV model has a higher forecast accuracy compared to other volatility models. Furthermore, the article also uses the realized bi-power variation (BPV) to replace RV as the substitute variable of volatility in real market to evaluate the forecast performance of each volatility model, and the aforementioned conclusions are still correct.(2) The signed jump variation not only can enhance the fitting precision of each volatility model but also can improve the forecast accuracy of models; compared to signed jump variation, the positive and negative signed jump variation have better explanatory abilities to future volatility, and their impact on future volatility is asymmetric; HAR-RV-TJ (or HAR-RRV-CJ) model has the best forecast performance among all the models.(3) In comparison to the positive realized variation, the semi-variation based on negative return has a stronger link to future volatility. Furthermore, the newly-constructed signed jump variation in this chapter has a remarkably negative impact on future volatility, which can contribute to lowering future volatility; through the Wald test, we find that the positive and negative realized variation and signed jump variation have different impacts on future volatility, that is, they both have the asymmetry; compared to other high-frequency volatility models, the newly-presented model has a better prediction accuracy, that is, the forecast ability is stronger; in the case of noise, the article also chooses RK more robust than RV as the basic volatility, and conclusions are nearly the same.(4) The realized volatility model with the best performance generally in each chapter all have a better forecast performance, but for realized range volatility models, their forecast performances are not good enough; when the Markov volatility model is expanded to volatility models with a better forecast performance in each chapter, either realized volatility or realized range volatility models have a better forecast performance. Especially for the realized range volatility models, they are not eliminated by MCS test just like linear realized volatility models; taking the realized volatility and realized range volatility as the basic volatilities to evaluate each volatility model, the empirical results by MCS test show that MS-HAR-RV-PS9 model has the best forecast performance in the chapter and dissertation compared to other models; when evaluating the economic value of volatility models, if taking each volatility model as a strategy, the article finds that the strategy constructed by MS-HAR-RV-PS9 model newly presented in this chapter can achieve the biggest portfolio return, and then the practical value of this dissertation is indicated, which provides a new research perspective for investors to choose the portfolio strategy.