Research Of The Realized Volatility Prediction Of CSI 300 Index Based On Independently RNN

Recurrent Neural Network is an important part of deep learning,compared with the traditional neural network,the cycle has the advantage of neural network can be processing with a sequence of time sequence data,but the ability to deal with longer time series is weaker,in order to solve this problem,many studies studied its long-time-memory ability has obtained certain achievement.As an important part of theoretical research and practical decision-making in the financial field,realized volatility is a method to calculate volatility based on intraday high-frequency data,which is superior to the volatility results obtained from low-frequency data and has important research value.As an important indicator of China's financial market,and it is also very important to study the volatility of CSI300 index.Based on the idea of Independently Recurrent Neural Network,the activation function and product form of Long-short Time Memory network are changed,and the model of Independent Long-Short Time Memory network is obtained.At the same time,this article will replace the activation function ReLu of IndRNN and IndLSTM by Elu function and softplus function to get new models and predict the realized volatility of CSI 300 index,analyzing the long memory ability of new models,and the model for the predictive accuracy of realized volatility has carried on the empirical analysis and comprehensive evaluation.This paper did an empirical analysis of the high frequency realized volatility obtained from the 5-minute high frequency data of CSI 300 index from 2010 to 2019.From the empirical results,we can find that the IndLSTM can process sequential data for a longer time than IndRNN,and has a long memory.At the same time,after the replacement of activation function,both IndRNN and IndLSTM can maintain good training performance and improve the accuracy of realized volatility prediction.By comparison,the IndRNN and IndLSTM with softplus function as the activation function are more accurate than the models with Elu and ReLu in predicting the realized volatility of CSI 300 index.