Research On Time Series Prediction Model Based On Deep Neural Network

Exploring the operation rules of the financial market and accurately predicting the trend of the financial market will help investors to make correct investment strategies with low risks and high returns.Therefore,the research on the prediction and analysis of financial time series has always been a hot topic of research.As a kind of nonlinear,non-stationary and high noise time series,financial time series is more difficult to study than traditional financial time series.The common machine learning methods for studying traditional time series are not highly applicable to financial time series.Only the deep learning neural network model can better fit the financial time series.Common neural network models include recurrent neural network(RNN),short-term and short-term memory network(LSTM),convolutional neural network(CNN),and the newly proposed time series convolutional network(TCN).Among them,convolution neural network(CNN)is mostly used for image analysis,while time domain convolution network can further solve the problem of time series prediction by applying cyclic algorithm on the basis of convolution network.Based on LSTM model and TCN model,this thesis selects the data of Shanghai Stock Exchange Index and two different stock price indexes.First,the data of the Shanghai Composite Index will be cleaned,normalized and standardized to eliminate some dimensional effects.First,introduce the basic theoretical knowledge of RNN network and LSTM network,GRU gating cycle unit,CNN network and TCN network,and establish LSTM model and TCN model respectively.Two sets of experimental results were obtained by data experiments on the two models.Then,two models,GRU-TCN and GRU-LSTM,were established respectively,and data experiments were conducted to obtain two groups of experimental results different from the first experiment.Three evaluation indicators are selected to compare the prediction results of the four groups of models.It is found that GRU has a good effect on improving the prediction accuracy of TCN model compared with the single TCN model,but it has no obvious effect on improving the prediction accuracy of LSTM model.Then,the validity and feasibility of GRU-TCN model are verified by comparative experiments.It is also extended to two different stock price index forecasts to further confirm the feasibility of GRU-TCN model in stock index prediction.