Study On Stock Index Prediction Model Based On Multiscale Decomposition And Deep Learning

With the development of domestic financial markets,the stock market occupies an increasingly important status in our economic system,and more and more people have begun to participate in stock investment.The violent fluctuations in stock prices will bring economic losses to investors.In extreme cases,it will even destroy the original market balance and affect the development of social economy.Therefore,the prediction study of the stock price trend has great application value.The stock price series has non-linear,high noise,non-stationary features,and traditional statistical models are difficult to fully capture nonlinear relationships in the data,and the data needs to be hypothesized before use.Compared with traditional methods,deep learning has powerful capabilities of extracting nonlinear feature,which can dig more important and deeper information from complex stock data.Based on the deep learning theory,this thesis explores the feasibility of deep learning in stock index prediction.The main research work is as follows:(1)A hybrid stock index prediction model of long short – term memory(LSTM)neural network based on principal component analysis(PCA)was constructed.Based on LSTM,PCA technology was added to reduce the dimension of input data.The memory storage unit and cyclic structure in LSTM were very suitable for analyzing time series,and could learn the dynamic change information in stock index data.PCA was used to reduce the feature dimension of the high-dimensional input in this experiment,removing noise and redundant information,and improved the feature extraction ability of the model.(2)The technology of complementary ensemble empirical mode decomposition(ceemd)was introduced to preprocess and decompose the stock index series,and the pca-lstm stock index prediction model based on ceemd was constructed.Because the daily closing price series of Shanghai stock index is nonlinear,noisy and non-stationary,the sub series with different time scale fluctuation characteristics can be obtained by adaptively decomposing Shanghai stock index with ceemd technology;Then,the run length determination method was used to reconstruct the high-frequency component,low-frequency component and residual term;Finally,the three subsequences were predicted respectively,so as to reduce the difficulty of model prediction.(3)Based on the ceemd-pca-lstm model,the attention mechanism based on time and space was integrated,proposing an improved model.Made the model give different attention weights in the two dimensions of time and space according to the influence of the input series on the final result.Thus,the network paid attention to more important values,increasing the impact of important information on the prediction results,and further improved the ability of LSTM model to learn the long-term dependence between data.Four groups of comparison models were set up,including LSTM prediction model,pcalstm prediction model,ceemd-lstm prediction model and ceemd-pca-lstm prediction model.We compared the prediction model in this thesis with the comparison models,and evaluated the prediction performance of the model.The experimental results show that the model designed in this thesis is able to achieve higher prediction accuracy and stronger prediction performance than the comparison models.Compared with the ceemd-pca-lstm model which performs best in comparison models,MSE error is reduced by 66.16%;MAE is reduced by 46.52%;and Rsquared is increased by 4.75%.This fully reflects the excellent performance of our model for the prediction of the Shanghai stock index,and proves that the idea of stock index prediction proposed in this thesis is feasible,that is,the deep learning modeling method based on decomposition-dimension reduction-attention mechanism.By using our model,we can predict the stock index of the next trading day,which is also of certain significance to guide stock trading.