A Research On Forecast Of Chinese Stock Price And Volatility-based On Deep Learning Method

Artificial intelligence has achieved a series of remarkable achievements in recent years.It has brought people a great shock and rich imagination.At present,artificial intelligence,especially deep learning technology,has had a wide-ranging impact on all aspects of people's lives.As the application of artificial intelligence,big data,and blockchain technology in the financial sector continues to deepen,the integration of finance and technology has brought disruptive changes and innovations in many financial scenarios.For the financial industry and research field,intelligence is an important direction for the development of the financial industry.Robo-advisors,robo-investment research and intelligent investment will all be manifestations of intelligent finance,and the key to these forms is how to achieve "intelligence." The successful application of deep learning in other fields provides a good reference for the development of smart finance.Therefore,with the advent of the artificial intelligence era and the increasing demand for financial data analysis,deep learning has become the frontier of applications in the financial field.One of the frontiers of application that has received widespread attention is the prediction of stock prices and volatility.The stock market is a complex nonlinear dynamic system,and the prediction of stock prices and volatility has always been the focus of research in the financial field.The stock market is affected by many factors such as economic situation,political environment,national policies,investor psychology,and other external markets.Its internal change laws are extremely complicated.Traditional time series models cannot effectively solve such complex nonlinear problems.The existing analysis methods can not achieve satisfactory results in predicting such high-dimensional,low signal-to-noise ratio,nonlinear and non-stationary problems.Most traditional measurement methods require parametric modeling,which cannot effectively describe extremely complex dynamic systems,and relies on assumptions that are not satisfied in the actual securities market;while traditional machine learning methods such as support vector machines,decision trees,and Random forest,etc.,its nonlinear characteristics do not yet have the ability to accurately model such complex data,and its predictive performance is highly related to the processing methods of artificial feature design,which affects the robustness of the model and the actual application effect.With the gradual development of neural network technology and the increase in the amount and availability of financial market data,deep learning has become the forefront of the application of stock price prediction.Deep learning extracts deep features through a hierarchical structure,strengthens important factors,and filters noise,which is of great significance for improving the accuracy of prediction.Based on this,this paper uses deep learning methods such as recurrent neural networks,convolutional neural networks and composite neural networks to predict stock prices and volatility.According to the traditional efficient market hypothesis,it is more difficult to predict the rise and fall of stock prices based only on stock price data and technical indicators for a stock market that satisfies weakly efficient stock markets.However,after research in this article,it is found that the application of deep learning methods can effectively predict the rise and fall of stock prices to a certain extent.On the one hand,the theory of behavioral finance provides a reliable theoretical basis for the work of this article.At present,there are still a lot of market anomalies and irrational behaviors in the Chinese stock market.The effectiveness of the Chinese stock market is still questionable.The weak-form effective market has not been reached in the short-term time interval;on the other hand,the adaptive market hypothesis believes that under the constantly changing market environment,predictability comes from the time-varying nature of the risk premium required by investors.A certain degree of predictability in the financial market does not mean that the market is inefficient,but is the driving force necessary for the market to maintain its effectiveness.Under such circumstances,applying deep learning technology to predict the rise and fall of stock prices is a possible and meaningful job.This article is mainly divided into three parts.The first part mainly uses the improved model of recurrent neural network,namely LSTM model and GRU model to predict the rise and fall of the Shanghai and Shenzhen 300 Index.The traditional recurrent neural network cannot effectively solve the long-term memory problem of time series due to the problem of gradient disappearance.Therefore,this paper uses the long and short-term memory model(LSTM)and the gated recurrent unit model(GRU)developed on the basis of the prediction.This part analyzes and compares the influence of different neural network structures and time steps on the prediction results.Experiments have found that the GRU model has a certain improvement in the daily K-line and 5-minute K-line prediction compared to the LSTM model.Next,the multi-time scale fusion forecasting model is used to extract the price features on two time scales at the same time,so as to capture the medium and long-term trend and short-term trend of stock prices at the same time,so as to achieve higher prediction accuracy.In addition,this part also applied the latest attention mechanism method for prediction,and found that the attention mechanism can effectively improve the processing ability of the LSTM model on financial time series.The second part mainly uses a variety of convolutional neural networks and a composite model composed of convolutional neural networks and recurrent neural networks to predict the rise and fall of the Shanghai and Shenzhen 300 Index.This part first uses the most widely used two-dimensional convolutional neural network in the convolutional neural network to predict the rise and fall of the Shanghai and Shenzhen 300 index,and uses single-channel and multi-channel network structures respectively.The experiment found that compared with non-image recognition tasks Commonly used single-channel convolutional network structure,application of multi-channel convolutional network structure has better prediction effect.Next,a one-dimensional convolutional neural network is used for prediction,and it is found that the prediction accuracy is higher than that of a two-dimensional convolutional neural network.The experiments in this article show that the one-dimensional convolutional neural network is more suitable for the characteristics of financial time series.Furthermore,through the use of composite models such as convolutional recurrent neural networks and recurrent convolutional neural networks,combined with the multi-time-scale fusion model proposed in this article,the prediction effect has been significantly improved.The third part mainly uses the LSTM model and the CNN model to predict the realized volatility of the Shanghai and Shenzhen 300 index and the 50 ETF index in one day,one week,and one month in the future,and compares the results with the current mainstream HAR family volatility prediction model.Through the SPA test,it is found that the prediction performance of the deep learning model is generally better than other models.The prediction performance of the LSTM model is higher than that of the CNN model,and the longer the prediction period,the more obvious the prediction performance advantage of the deep learning model.In the task of forecasting volatility in the next week and one month,the LSTM model is significantly better than all other models under all four loss functions.This result shows that the deep learning method can effectively extract the dynamic characteristics of the volatility and thus has a better prediction ability.Compared with the current mainstream HAR family high-frequency data model,it can achieve better prediction results.The main research conclusions obtained in this paper are as follows: First,BP neural network is difficult to effectively predict stock price data.In this paper,various structural combinations of different network depths,number of neurons,input data cycles and input data types have been tried,but the prediction accuracy is not ideal.The experiments in this article show that the traditional BP neural network has limited ability to fit financial time series data.Second,the recurrent neural network can significantly improve the prediction accuracy.Compared with the LSTM model,the GRU model has a better effect in the financial time series prediction task of this article,and can reach a prediction accuracy of58.45%.And the GRU model can effectively train longer time steps than the LSTM model,and the training speed is faster.Third,the attention mechanism technology that has emerged in the past two years can improve the effect of neural networks on financial time series prediction,and can increase the prediction accuracy to 59.52%.Fourth,there is no significant difference between the predictability of stocks in daily K-line frequency and5-minute K-line frequency,but 5-minute K-line prediction is less sensitive to prediction methods and model structure,and is relatively easier to train.This is because the data volume of the 5-minute sample is larger.Different deep learning models and network structures are applicable to prediction tasks at different time frequencies.Fifth,although most non-image recognition tasks use a single-color channel model structure when applying convolutional neural networks,this article finds that for stock price prediction,multiple color channels are used,that is,different indicator data occupy different color channels.The forecast effect will be better.In addition,the applicability of one-dimensional convolutional neural networks to financial time series is higher than that of ordinary two-dimensional convolutional neural networks,and the prediction accuracy can be significantly improved.Sixth,the recurrent convolutional neural network and the convolutional recurrent neural network that combine the recurrent neural network and the convolutional neural network can combine their respective advantages to achieve better prediction accuracy than using a deep learning method alone.Seventh,the cross-time scale prediction model proposed in this paper can effectively predict the rise and fall of stock prices,especially the multi-time scale fusion prediction model proposed in this paper can significantly improve the prediction accuracy,and the training speed is faster,and the prediction accuracy can be Reached 60.86%.Eighth,the use of deep learning methods such as LSTM model and CNN model can predict the future volatility of the stock market more effectively than the current mainstream HAR family volatility model.