Research On Hybrid Stock Index Forecasting Model Based On Deep Learning

With the continuous improvement of the domestic market economic system,the stock market plays an increasingly important role in China's economic system,and the number of people participating in stock investment has gradually increased.The fluctuation of stock price affects all aspects of social and economic life,so it is of great value in practical application to effectively predict the stock trend.Traditional time series forecasting models generally can only find the linear relationship between the data,while the stock data is essentially high-noisy,dynamic and non-linear.Compared with traditional methods,the deep learning theory,with its strong self-learning ability and non-linear approximation ability,can well solve the problems such as numerous data and complex non-linear relationship in stock forecasting.Based on the deep learning theory,this paper explores the feasibility of applying artificial neural networks to stock index forecasting.The main research work is as follows:(1)The closing price of stock index is affected by many factors,so the data set must be able to contain as comprehensive indicators as possible.In addition to selecting the most common basic indicators such as "four prices and one volume"(opening price,closing price,highest price,lowest price and trading volume),this paper also calculates relevant technical indicators based on the daily data of stock index and establishes the index system of stock index forecasting.(2)A hybrid stock index forecasting model combining Convolutional Neural Network(CNN)and Gated Recurrent Unit(GRU)neural network is constructed.In view of the high-dimensional and high-noisy characteristics of stock index data,the excellent feature extraction ability of CNN is used to fully explore the internal relationship between stock index data,thereby reducing the scale and complexity of the original data.Stock index data is a kind of time series data.Through the time series memory ability of GRU neural network,the internal dynamic change law of stock index data can be further learned,and the non-linear relationship between input and output can be established.(3)When the input time series is long,GRU neural network is easy to lose the sequence information and difficult to model the structural information between the data.In order to solve this problem,this paper introduces attention mechanism into GRU neural network,and proposes a CNN-GRU hybrid stock index forecasting model based on attention mechanism.Attention mechanism gives different weights to the hidden state of GRU neural network to ensure that the features containing important information will not disappear with the increase of step size,strengthens the influence of important information,and makes it easier for GRU to learn the long-term dependence relationship between stock index data.(4)Taking the CSI 300 Index as an example,the influence of model parameters on the forecasting results of stock index is analyzed,and the determination process of the number of convolutional layers,the number of convolution kernels in convolution layers,the number of GRU layers,the number of neurons in GRU layers and optimization learning methods is discussed respectively.In order to evaluate the performance of the models constructed in this paper,CNN-GRU model,CNN-GRU-Attention model,CNN model,Long-Short-Term Memory(LSTM)model and GRU model are compared to predict the CSI 300 Index.CNN-GRU model and CNN-GRU-Attention model are further applied to the forecasting of SSE Composite Index and SZSE Component Index.The experimental results show that,compared with CNN model,LSTM model and GRU model,the models designed in this paper have stronger forecasting ability and higher forecasting accuracy.At the same time,the models in this paper have also achieved good results in the forecasting of SSE Composite Index and SZSE Component Index.This fully reflects the powerful feature extraction ability and good non-linear time series data processing ability of the models.This paper provides a new idea for the research of stock index forecasting,and it has a certain practical significance in assisting high-frequency data trading.