Prediction Of Stock Average Price Based On Improved Time Convolution Network

With the advent of economic globalization,this new type of investment in stocks has received a lot of attention from investors,and how to accurately predict the stock market movements has become a matter of great concern to investors.Stock forecasting is a special kind of time series forecasting problem,and its own nonlinearity and uncertainty lead to many factors to be considered when making stock forecasts.Temporal convolutional network shows good prediction results in stock prediction.In order to better solve the challenges in stock prediction,enable investors to gain more profit in the stock market,and further enhance the stock prediction ability of TCN,this thesis proposes two improved methods for the temporal convolution module and the connection method in TCN.The data set constructed by the top 10 stocks of Shenzhen A-share was used to conduct the ablation experiment on the weekly line ma＿5,monthly line ma＿20 and seasonal line average price prediction of stocks.The experimental results show that the two improved time convolution networks have improved the prediction accuracy of the three moving averages,and have strong practical value.The main content of this article:First,a stock mean price prediction model Res-TCN based on multiscale convolution and residual connectivity is proposed.The multi-scale convolutional Inception structure is added to the Temporal block of the temporal convolutional module of the standard TCN.The multiple convolutional kernel structures in the multiscale convolutional make the features extracted by the model more complete and enhance the extraction of features by the model.Immediately after expanding the number of layers of the network,the TCN blocks of each layer are connected together using residual connections,which help prevent gradient degradation during network training.The prediction accuracy of the Res-TCN network designed in this thesis is70.58%,72.14%,and 74.53% for weekly ma＿5,monthly ma＿20,and quarterly ma＿120,respectively,which improve 5.89%,0.62%,and 1.98%,respectively,relative to the TCN,and obtain a good prediction result.Secondly,in order to make the model more efficient in prediction and better in feature extraction,a network model Dens-TCN-GAM based on global attention and dense connectivity is proposed.The global attention module GAM is added to each layer of Temporal block to form a TCN-GAM block,and the TCN-GAM block of each layer is convolved with 3×3 as the conv block in the Dense block to maximize the ability of the network to utilize the features.After the experiments using Adam optimization algorithm,the prediction accuracy of the improved model Dens-TCNGAM for the three mean lines is 75.48%,77.33%,and 79.05%,which is 4.90%,5.19%,and 4.52% higher than that of the Res-TCN network,respectively,and the prediction efficiency and accuracy are improved.Finally,a multi-influence factor-based stock selection system is designed to enable investors to take fewer steps.Multiple impact factors affecting stocks are used as input features of the dataset,while expanding the types of stocks in the dataset and selecting different types of stocks to be added to the stock pool.After the output of the prediction model,the 10 stocks with the best weekly,monthly and quarterly predictions are selected to form the stock pool,and investors can choose according to their needs.