The comparison of cluster based Soft – Computing methods on the trend forecasting with the weighted stock price index

• Main Authors: Shih-Ban Yu, 尤仕邦
• Other Authors: Chen-Hao Liu、Pai-Hsien Pang
• Format: Others
• Language: zh-TW
• Published: 2009
• Online Access: http://ndltd.ncl.edu.tw/handle/28777945982161194188

碩士 === 開南大學 === 財務金融學系 === 97 === This study utilized Back-Propagation Neural Network (BPNN)、Support Vector Machine(SVM) and Genetic Algorithm Fuzzy Decision Tree (GAFDT) as three primary prediction models to forecast the stock trading signals. And the data clustered or not will be compared in both models. The object of this research was the Weighted Price Index of the Taiwan Stock Exchange. In 25 technical indices, we first used stepwise regression analysis to sieve out relative important factors, and k-means cluster analysis was adopted for data clustering. Then five-full alternation was applied in the experimental design process. Model from one to three are utilized to predict by BPNN, SVM and GAFDT, respectively. GAFDT began with transferring continuous data to discrete data by fuzzy sets so as to increase the comprehensibility of decision tree, and then utilized Genetic Algorithms to optimize the parameters like fuzzy term numbers. Finally, we compare the different intervals between Weighted Price Index of the Taiwan Stock Exchange and S&P 500 by GAFDT. Empirical findings revealed that the highest accurate rate was GAFDT (82.93%), followed by SVM (76.12%) and BPNN (75.53%). The more clusters GAFDT and SVM swarm off, the higher the accurate rate is. The accurate rate of falling of stock was higher than correction and rising in Weighted Price Index of the Taiwan Stock Exchange. The accurate rate of rising of stock is higher than correction and falling in S&P 500. Consequently, the prediction will be affected by training and testing in different stock markets and tendencies.