The Stock Pricing Model Based On Data Mining Technology

Stock pricing model is used to predict the returns of single stock and portfolios.At present, traditional stock pricing model is statistics model or econometrics model,and there are some defects in both of them: Firstly, all of them are single model, bywhich the prediction can be showed in one time period. So the characters of bigvolatility and frequent change of data are ignored, and the multi-scale calculation ofstock returns cannot be got. Secondly, all of them come from the mature stock market,and fixed two to three index are as the model factors, but the returns are affected bymore than one hundred factors in A stock market. Thirdly, the explain efficacy ofsome portfolios are not well when they are used to analyze A stock market, whichmeans the adjusted variance of value is non-significant. Finally, the factors of modelsare greatly complicated to choose and compute, and typically,70sheets for size of223*256in Excel Workbook have been used to fulfill the computing of factors in theFama-French three factors model.Three kinds of Data Mining Technology are combined in the thesis. And thetraditional stock pricing model, which mainly means the Fama-French three factorsmodel, is improved and expanded.At first, the multi-scale Fama-French three factors model on basis of Daubechieswavelet is constructed, and the experiment result shows that the effect of three factorsto stock returns in different time period are investigated at the same time. And thedefect of traditional mode, which is a single model, is solved.At second, the cross-sectional factors are mined by the clustering technique. Thehierarchical segmentation clustering algorithm is given in the thesis, and thehierarchical clustering BIRCH algorithm and partitioning clustering CLARAalgorithm are combined. So the defects that BIRCH algorithm is inefficient inaspheric and outlier data and CLARA algorithm can be hard to decide the categoricalmeasure just by users themselves are remedied. At the same time, the large-scale dataset can be clustered and analyzed by improved clustering algorithm.At third, the N factors prediction model on basis of improved clusteringalgorithm and BP neural network is structured. It is realized by organizing neuralnetwork through four factors from the clustering analysis, so the returns are predictedin six portfolios in different time periods, and they are compared with the real value. The results are showed that the adjusted variance is significant in B/H portfolio,which is not significant in formal models.Last but not least, the prediction model on basis of improved clustering algorithmand BP neural network is used to choose stocks. The experiment results show that thepredicted values match well with the actual values. By which the complexity ofcalculation of factors in normal models are reduced, and the time to make stockpricing analysis is saved for investors.