Application Research Of Stock Index Forecasting Based On Convolutional Neural Network

How to effectively apply artificial intelligence to the prediction of stock market is one of the challenges we are facing.If an effective stock index forecasting method can be applied,it will be helpful for investors to make more informed investment decisions.This paper mainly focuses on the ability of convolutional neural networks to on predicting stock index trends.This paper uses the CSI 300 Index as the main research object to predict its four types of ups and do wns in the next minute.The trading data of stock index from January 2nd,2019 to July 15th,2019 were selected in the experiment.Based on the technical indicators of expanding stocks by four prices,one volume and one transaction amount,multiple feature selection methods were applied to determine the optimal feature subset.A 10×10 two-dimensional image is constructed by sliding per minute,where the x-axis is composed of 10 features in each time dimension,and the y-axis is composed of 10 time dimensions.Based on a simple stacked model of one-dimensional convolution and two-dimensional convolution,the model structure and parameters were optimized from multiple aspects to determine the optimal model.The accuracy rate of the model in the four-class rise and fall problems of the CSI 300 Index is 47.8%,which is 91.2%higher than the accuracy rate of random guessing at 25%.In addition,this paper conducted two sets of comparative experiments.One is to use logistic regression,support vector machine and ensemble learning.The other is to build models with different structures such as shallow ResNet and DenseNet.In the stock index prediction method characterized by technical analysis indexes,the experimental results show that the simple convolutional neural network is superior to traditional machine learning and complex convolutional architectures in the stock index prediction method characterized by technical indicators.Moreover,our convolution model’s prediction of small ups and downs is more feasible than big ups and downs,and the recall rate is relatively balanced.In this paper,the model is applied to 40 stock indexes in Shanghai and Shenzhen stock markets.The prediction results show that our model has a good generalization ability.The accuracy rate of 8 stock indexes is higher than 50%.At the same time,this paper also conducted experimental exploration based on two aspects of image construction and convolution architecture.Among them,the mixed input model can increase the accuracy of three common indexes by 1-2%.This research is enlightening,which is conducive to the future development of the stock forecasting model.