- 學位論文
類神經網路股價預測-以實驗設計為優化基礎
Predicting the stock price based on the hybrid DOE-based optimization of neural networks
摘要
近年來油電價不斷調漲進而帶動民生物價向上飆升，再加上全球通膨壓力高漲及低利率存款的多重壓力下，利用投資理財的獲利為自己加薪進而達到累積財富的目的，已成為生活的重要課題。在多元化的金融商品中，股票是一種公開且較容易取得資訊的投資工具，而投資理財最終目的是以最小成本，獲得最大收益。不論是藉由專業知識的判斷，或是透過各種股票分析工具的輔助來進行股價漲、跌預測，對於股價變動一直是個別投資人或法人機構所關心的焦點，也是學術研究的重心。 在學術領域，利用類神經網路(Artificial Neural Network, ANN)模型進行股價預測的相關研究一直在持續中。類神經網路在實證研究上具有預測正確率較高，且不受限於樣本為常態分配的假設，及具有處理非線性問題等優點，其網路預測功能的優越性取決於網路的參數設定、網路架構設計及問題的複雜度，若是能夠選取適合的參數及網路架構，則分析結果將會更加顯著。在這些研究中，對於類神經網路參數的設定，研究者多數採用試誤法(Trial-And-Error Methods) 來獲取網路最佳參數，這種方式既耗時又費力，找出的參數也未必是最佳，本研究分別以兩時期的財務指標為研究變數，採用倒傳遞類神經網路建構股價預測模型，應用實驗設計以系統化的實驗排程，達到節省時間與成本，並透過因子主效應與交互作用分析，來優化參數，實驗結果可以提升相關係數至0.93與0.87，證明本研究的方法確實能提高股價預測的準確率。
並列摘要
In recent year, people are is suffering from the raising of commodities prices, which is caused by oil price and electrical price rising at the same time. Under the pressure of inflation and low saving interest rate, how to make money and accumulate fortune by investment has become the main concern of people’s everyday life. Among the financial commodities, the stock is one investment tool which investors can get the public information more easily. However, the goal for investment is to make great profit by the lowest cost. Predicting price activities in stock market on the basis of either professional knowledge or stock analytical tools have been a great concern of individual and institutional investors around the world, because price variations result in gains and losses for investors. Stock return predictions are at the core of many research issues too. The stock forecasting model which based on NN models has been increasing. The goods of NN’s applications on empirical research studies are high accuracy rate of forecasting. Besides, the application of NN will not be constrained by the assumption of normality and it can deal with the non-linear distributions. The effect of a network’s functional approach depends on the network architecture, parameters, and problem complexity. If inappropriate network architecture and parameters are selected, the results may not be desirable. On the controversy, the results will be more significant if good network architecture and parameters are setting. Researchers set the parameters of ANNs intuitively or by trial-and-error processes to obtain the results. It is time and money consumption, besides the parameters setting won’t get the best result. In this study, adopting the financial data and the applying of ANNs was proposed. Besides the application of experimental design and through the process of main effect analysis and interaction analysis, the best parameters for the ANN model can be found. The research result shows that by the method applied in this research the correlation coefficient can improve to be 0.93 and 0.87 which is much better than the result by try-and-error. The research method applied in the research was approved to improve the accuracy rate of stock price forecasting.