The Study On The Realized Minimum Variance Hedge Ratio Of CSI 300

This paper constructs the Realized Minimum Variance Hedge Ratio(RMVHR)from 5-min high-frequency data of the CSI 300 stock index and its corresponding futures contract.Several models are built to estimate and forecast the RMVHR series,including AutoRegressive Moving Average(ARMA),Regime Switching(RS),AutoRegressive Fractional Integrated Moving Average(ARFIMA),AutoRegressive Moving Average-Generalized AutoRegressive Conditional Heteroskedasticity(ARMA-GARCH),Hetero-geneous AutoRegressive(HAR),and MIxed DAta Sample(MIDAS)which we introduce in such context for the first time.And for comparison,we introduce Ordinary Least Square(OLS),Error Correction Model(ECM),Dynamic Conditional Correlation(DCC),Dynamic Conditional Correlation with Realized Variance(DCC-RV)whose left-hand term is return and Vector Hetero-geneous AutoRegressive(VHAR)whose left-hand term is realized varance/covariance.We use hedge effectiveness(HE),sharp ratio(SR),value at rist(VaR)and expected shortfall(ES)as economic measures,compare the out-of-sample forecast results of all the nine models mentioned.The results show that models directly target at the RMVHR series performs better than the daily models,in the sense that all the best models under different economic measures is one of the direct models,especially the MIDAS model.And we also find out that the HAR model outperforms the VHAR model in three of the four economic measures.Notice these two models share the same basic form and information set.This observation also agrees with our finding that this measure(RMVHR)could improve the forecast performance from the economic perspective.Then,an intra-comparison is performed between all the models directly describe the dynamic of the RMVHR series.We use MSE,QLIKE,pMSE,logMSE,MAE,pMAE,logMAE as the loss function.Results show that HAR model scores a better forecast under error measurements,and the MIDAS model who also describes the long-term memory generates good forecast results as well.A Model Confidence Set(MCS)test then is introduced to offer more information about the forecast performance of all the RMVHR models.We can see more clearly that the three long-memory models,i.e.,HAR,MIDAS,and ARFIMA,perform better than the other three,who only contain short-memory process,especially the RS model.And for the purpose of robust testing,we divide the whole out-of-sample forecast period into three subsamples according to the magnitude of volatility and return level.We compute the economic and error measures of the forecast results of all the models in the three subsamples.The finding supports our conclusion before.Except for only a few exceptions(the DCC-RV model gets the highest sharp ratio in the low-volatility period and the mid-return period,and the OLS model gets the highest sharp ratio in the high-return period),models directly describe the dynamic of the RMVHR series generate better forecast results than models with return or realized variance on their left-hand side.And the MIDAS model performs ideally under economic measures as well as the error measurements,which recognizes our effort of introducing the MIDAS model in the study of RMVHR series.Another finding is that RS model generates good results under certain circumstance,however,results from the MCS test contradict this finding.We believe that the reason why RS model performs fairly well when volatility level goes up is that the RS model successful describes the different dynamic of the RMVHR series under different volatility level so that the out-of-sample forecast performance of the RS model does pretty well under certain circumstance while does not do as well in the whole out-of-sample forecast period.Notice that the abnormal high volatility period caused by the extreme fluctuation of China’s stock market in 2015 is included in our whole out-of-sample period,which suggests that certain level of deviation should be tolerated in this phase.In the robust testing section,we also find out that the HAR model who estimates directly the RMVHR series generally outperforms the VHAR model who estimates jointly the realized variance and covariance series.This result also suggests that this measure(RMVHR)does have an advantage formulating a hedging strategy.