Stock Price Prediction Based On TCN Hybrid Model

Stock investment is one of the most popular means of high-yield investment in contemporary society.However,high yield must be accompanied by high risk.It is the realistic demand of investors to be able to predict the future stock price trend,to avoid investment risk in time and master the optimal investment strategy.In order to better predict the stock price and provide reasonable investment suggestions for investors,it is necessary to put forward a method that can effectively predict the future development trend of the stock market.However,the financial field is a complex and dynamic field,and financial time series data is often accompanied by high noise,non-stationarity,nonlinearity and other characteristics,so that a single model may have poor prediction effect,model gradient disappearance explosion and other problems.In this regard,this article proposes a construction scheme for Time Series Convolutional Neural Network(TCN)hybrid model,aiming to further improve the accuracy of stock price prediction.The article selects the more representative data of the Shanghai and Shenzhen 300 Index constituent stocks as the research object,which is mainly divided into two parts in the actual model construction process.The first part is the stock price forecasting experiment based on TCN single model,which discusses the accuracy of TCN neural network for short-term stock price forecasting under different parameter combinations,and builds the TCN benchmark model with optimal forecasting effect.In addition,the model prediction accuracy of random forest,support vector machine,convolution neural network,long and short memory neural network and time series convolution neural network on the same data set is compared,and the advantages of TCN model are further determined.The other part is the stock price prediction experiment based on the TCN hybrid model,which aims to improve the algorithm and model structure based on the TCN model determined in the first part,and further mine the time series information of stocks,so as to achieve more efficient short-term stock price prediction.Specifically,we will gradually improve the accuracy of the model from the following three aspects:Firstly,considering the features of the model.The recursive feature elimination method based on random forest(RF＿RFE)was selected to complete the feature engineering.Six key feature variables were selected from 52 initial feature sets,which reduced the average relative error percentage(MAPE)of TCN model prediction from 1.802% to 1.768%,and the root mean square error(RMSE)from 1.031 to 0.996.Secondly,considering from the output level of the model.Introducing an attention mechanism between the TCN network layer and the fully connected layer in the model reduces the interference of unimportant features.The results show that the RMSE and MAPE of the model are reduced by 0.017 and 0.11%,respectively.Finally,considering from the input level of the model.The LSTM network is integrated into the model input layer to facilitate deeper mining of data temporal characteristics.A hybrid stock forecasting model(RF＿RFE-LSTM-TCNA)is proposed,which combines the random forest based recursive feature elimination method(RF＿RFE),short-term memory network(LSTM),temporal convolutional neural network(TCN)and attention mechanism(Attention).The results show that the RMSE and MAPE of the model are reduced by 0.079% and 0.036%respectively,and the prediction effect is significantly better than the individual TCN model.