The Research Of Realized Volatility Based On Multivariate-Hybrid Neural Network

Volatility prediction plays an important role in the optimal allocation of assets,the pricing of derivatives,investment decisions,asset evaluation and the correct measurement of market risk.Early prediction of volatility is based on low-frequency data,including historical volatility model,ARCH model,GARCH model,SV model and so on,With the availability of high frequency data,the concept of realized volatility was put forward.Some scholars proposed two new models based on the realized volatility calculated from high frequency data,namely the Autoregressive Integrated Moving Average model(ARIMA model)and the Heterogeneous Autoregressive Realized Volatility model(HAR-RV model).These two models get better prediction results.Realized volatility has gradually become an important means to study price volatility.Later,many scholars have studied realized volatility,but most of them are based on HAR-RV model and its variants.In this paper,Fully connected Neural Network(FNN)and Long-Short Term Memory neural network(LSTM)are used to study the realized volatility by comprehensively considering the factors affecting the realized volatility.In addition to the explanatory variables of the heterogeneous autoregressive model,RVtd,RVtw and RVtm,this paper also introduced Implied Volatility and two market variables as the inputs of the model,namely,trading volume and amplitude.First,the FNN-IV model is built,and predict implied volatility by the Fully connected Neural Network.Secondly,the six explanatory variables of realized volatility,implied volatility,volume and amplitude were pre-trained by autoencoder(AE)to extract the reduction and amplitude characteristics.Finally,the output of AE coding process,namely the low-latitude vector extracted from the original input multidimensional data sequence,is used as the input of LSTM for training and prediction.To sum up,the realized volatility model of multivariable-hybrid neural network(FNN-AE-LSTM-RV-MV model)was established in this paper to predict the realized volatility.Empirical analysis of the model(FNN-AE-LSTM-RV-mv model)established in this paper is carried out on Shanghai50ETF.At the same time,the LSTMRV model and HAR-RV model,are empirically-studied on Shanghai50ETF.The comparison and analysis of model prediction results show that the FNN-AELSTM-RV-mv model established in this paper has better prediction effect.The empirical analysis process mainly includes the following conclusions:The error analysis of the fully connected neural network implied volatility model is better than that of historical volatility;The AutoEncoder model(AE)has strong feature extraction ability,and the prediction effect of LSTM neural network is improved by pre-training with the AutoEncoder(AE);The nonlinear hybrid neural network model is superior to the linear regression model;Considering other influencing factors,that is,forecasts that include more market information are better than those that take only history realized volatility into account.