| Summary: | 碩士 === 中華大學 === 資訊管理學系碩士班 === 99 === Over the past most multi-factor stock selection models used score approach, and subjectively set the score weight for each factor. This subjective approach not only can not optimize of performance of stock selection model, but also can not determine the best weights in accordance with the preferences of investors. This study employed mixture design of experiment and neural networks to construct stock investment decision-making system to overcome these shortcomings. Using six stock selection concepts, small price to book value ratio (PBR), large return on equity (ROE), large annual revenue growth rate during recent three months, large quarter return, large total market capitalization, and small systematic risk β, and a mixture design of experiment called Simplex Centroid Design, 63 experiment points were generated. The samples of this study contain all listed stocks on Taiwan stock market, and the study period is from January 1997 to June 2010 with a total of 13.5 years. The performance of each year was normalized to 0.2~0.8 to be as the dependent variable of the performance prediction model. The results showed that (1) in the annual rate of return prediction model, the large ROE concept representing growth (proportional) and the small PBR concept (proportional) representing the value were the most important predictors. In the standard deviation of annual rate of return prediction model, the small PBR concept (proportional) and the small β concept (inversely proportional) were the most important predictors, which means the smaller the past risk, the smaller the future risk. The risk of equity is persistent. (2) When learning cycle of neural networks reached to 300, except that the model of "average market capitalization of stocks in portfolio" may be further improved, the coefficient of determination during test period of the rest models had reached the highest and can not be further improved. (3) Shortening the moving time frame to one year can not improve the prediction ability. (4) The explanatory power of neural networks is superior to that of regression analysis. However, neural networks are useless for the monthly relative winning rate model and the monthly absolute winning rate model, which are difficult to build accurate models. (5) In the maximizing annual rate of return optimization model, the most important stock selection factors were the ROE and PBR. In the minimizing standard deviation of annual rate of return optimization model, they were the ROE and β. The empirical results show that the stock selection strategies generated by the optimization models can meet the demand of stock picking for different investors.
|
|---|
