none

• Main Authors: Hon-Seng He, 何鴻聖
• Other Authors: Chien-Fu Chen
• Format: Others
• Language: zh-TW
• Published: 2004
• Online Access: http://ndltd.ncl.edu.tw/handle/15654132410379216267

碩士 === 國立東華大學 === 國際經濟研究所 === 92 === This essay examines the performance of nonlinear stock-selection strategies based on Kohonen (1982, 1995) SOM neural network.The sample period covers January 1995 to December 2003. Our samples include all electronic companies listed in the Taiwan Stock Exchange. The stock-selection strategies based on SOM can be used to catch the nonlinearity of stock returns due to the ability of nonlinear clustering.Our findings show that the perfomance of SOM-based stock-selection strategies is obviously superior to the traditional method based on sorting returns. Our major empirical results are as follows. First, the stock-selection performance of SOM model is obviously superior to the tradition model in all testing periods, including one month, three months, and six months, and one year. Secondly, under the consideration of return and risk meanwhile, in the one, three and six months of the testing period, the portfolio of SOM neural network has the advantages of higher rate of return and lower risk over the traditional stock-selection model.Thirdly, comparing with the Sharpe ratio of the two models, we can find that SOM model has higher Sharpe ratio in the one, three, six, and twelve months of test period. As a result, even though we take investing risk into consideration, SOM is still better than traditional sort method. Finally, the winner-loser protfolios which SOM model constructs has many financial properties, including size e ect and book-to-market ratio e ect. In addition, in SOM portfolio, the growth rate of sales and future return are in an opposite relation.