Statistical Analysis Of Financial Data Based On Multiple Dimensionality Reduction Methods

This paper chooses financial data as experimental samples,and uses a variety of statistical methods and dimension reduction methods.Firstly,the financial data is preprocessed,reducing the data to different dimensions with four dimensionality reduction methods-principal component,projection pursuit,LLE and MDS.Secondly,four methods are compared to analyze the situation of the appropriate data.Then the original data and the processed data are classified by six algorithms-decision tree,bagging,random forest,boosting,SVM and neural network to predict the rising and falling trend of the financial data.Meanwhile,this paper uses five statistical methods to analyze the original data and the processed data,and to predict the exact value of the data.The five statistical methods include decision tree,random forest,SVM,PP as well as neural network.Finally,the paper compares the data before and after dimension reduction to draw the conclusion.This paper first introduces four kinds of dimensionality reduction methods,including principal component analysis,projection pursuit,LLE,and MDS.Then,seven statistical methods,decision tree,bagging,random forest,Boosting,SVM,neural network and projection pursuit are introduced.Under the analysis,the financial data is subject to certain research.The experiment of this paper is divided into two parts.The first part is the prediction of the closing price trend and the opening price trend of the stock,which is experiment 1 and experiment 2.The second part is the analysis of the futures data forecast and the stock index closing price forecast,which is experiment three and experiment four.Two groups of data are selected in the stock trend prediction experiment.The first group is 100 data of three stocks,Guangdong Expressuray B,Guizhou Maotai and Agricultural Bank.Among them,the first 80 data are used as training samples and the latter 20 are used as test samples.The second group selects 600 data of Shanghai Stock Index,Shenzhen Stock Composite Index,Small and Medium Stock Market Index and GEM Index.Among them,t he first 560 are training samples and the latter 40 are test samples.Respectively,they predict the opening and closing price trend of the stock.The experimental process is divided into two parts.Firstly,six algorithms are used to analyze the data directly,including decision tree,bagging,random forest,boosting SVM and neutral network.The calculated results are compared with the original results.The first group of experiments found that the SVM algorithm had better prediction performance.Then four dimension reduction methods-PCA,PP,LLE and MDS were used to analyze the results.The results show that compared with the direct data analysis,the prediction accuracy may be improved by adjusting the parameters after dimension reduction.In the second group of experiments,the data before and after dimension reduction were compared and analyzed respectively.It was found that the accuracy of boosting and neutral network was improved,and the effect of SVM was partly improved.After two groups of dimension reduction experiments,it was found that PCA and PP had better classification accuracy than LLE and MDS after dimension reduction.In the futures data forecasting and analysis experiment,four groups of data,namely gold index,gold main link,moving coal index and moving coal main link in the futures market are selected.The index number is 43 and each group has 400 data to forecast the opening price.Firstly,five regression methods are used to analyze the data directly,namely,decision tree,random forest,SVM,PPR and neural network.The result is the PPR algorithm has the best effect.Then,dimension reduction is processed.PCA,LLE and MDS are selected for repeated experiments.The final results show that the prediction progress of the combined dimension reduction by decision tree,SVM and neural network has been improved.Besides,part of the accuracy of random forest and PPR has been improved.In the experiment of stock index prediction and analysis,the closing prices of four stocks are selected.They're S&P 500,Russell 2000,Walmart and Disney 850 data are selected as training samples.The experiment uses time series model to predict and analyze the data.The same five regression methods are used for the data.Firstly,direct regression is used,and then PCA,PP,LLE and MDS are used to reduce the dimension of the data,and then the data are re-analyzed.Finally,the results are compared with direct regression.The results show that when there are some errors in the direct data analysis,the data can be reduced in dimension,the analysis results will be improved,and the improvement of high-dimensional data after dimension reduction is better than that of low-dimensional data after dimension reduction.