Research On Stock Price Index Prediction Based On Multi-scale Deep Learning Mode

With the advent of the era of big data,the application of deep learning in financial forecasting has made a breakthrough.As one of the research subjects in the field of financial forecasting,stock price index is more difficult to predict than other common financial time series due to its high non-stationary,unstructured and nonlinear characteristics,which are affected by complex factors such as investor sentiment,major public events,policy changes and international economic trends.It is of great significance for regulators and investors to construct a set of reasonable stock price index model.In this paper,ICEEMDAN-RF-SSA-LSTM price prediction model is proposed based on the difficulties existing in the improvement of the accuracy of financial time series price prediction.In the construction of the model,ICEEMDAN decomposition model is used to decompose the originally complex stock prices into time series under multiple scales to resolve the non-stationary and chaotic characteristics of financial timing.Through the more effective IMF restructuring strategy proposed in this paper,the IMF sequence decomposed by ICEEMDAN is restructured.The stock index factor library was constructed,and RF model was used to match the influential factors with high importance of each component at different scales after the decomposition of stock index in the factor library.In combination with LSTM’s excellent performance in long memory in time series prediction and SSA’s adjustment of LSTM super parameters,the feature recognition,learning and fitting of each component were carried out.To achieve accurate prediction of each segment component;Finally,the prediction results of each component are reconstructed by the integration method to form the final prediction results.On this basis,this paper takes Shanghai Composite Index and CSI 300 index as examples,constructs 4 single models and 8 combination models for comparison test,and further verifies the effectiveness and stability of ICEEMDAN-RF-SSA-LSTM through DM test.The contribution of this paper is: 1.From the perspective of price prediction model construction,a new ICEEMDAN-RF-SSA-LSTM stock price index prediction model is proposed by integrating complementary empirical mode decomposition model in signal decomposition technology,random forest model in machine learning and LSTM model in neural network.This model can solve the non-stationary and nonlinear problems of financial timing.2.In this paper,the stock price index factor library was constructed,and RF model was used to match the influential factors of high importance of stock prices at different scales after decomposition in the factor library,which successfully made up for the regret that other factors were not introduced into the "factory-integration-prediction" model in the past papers.3.LSTM model is used to predict the decomposition sequence.By combining ICEEMDAN and LSTM model,signal processing and artificial intelligence technology are successfully applied to the research in the field of finance,and multi-disciplinary research results are integrated to build a more accurate and perfect stock price prediction model.