股價預測在股票中是個相當重要的問題,投資者在進入股票市場時都希望利用高準度的預測系統去得到較高獲利。近幾年,類神經網路在股價預測上有著很好的成果。歷年來學者們在類神經網路上做了不少改良並且獲得更好的預測系統,但還是存在著許多問題。其中之一就是類神經在預測前需要決定非常多的參數,像是輸入的種類、隱藏層的層數、隱藏層神經元的個數…等,不同的參數組合就會有不同的預測準確度。本篇論文在眾多的類神經網路中選擇最廣為使用的到傳遞類神經網路(Backpropagation Neural network, BPN),提出一個既簡單又具高準度的類神經預測系統,將與目標明天股價最為相關的前幾天的股價做為輸入值,並改良傳統的正規化公式,使其能夠動態的配合股價的走勢作調整,達到改良預測準度的目的。所提出的系統有很好的成果,但參數過多的問題依然存在,而且傳統類神經所使用的公式太過於複雜導致每個神經元所連結的權重沒法直接得知其意義,因此本論文更進一步地提出另一個簡單易懂的類神經網路。以輸入參數的重要性所設計的公式讓使用者能夠直觀的知道每條權重的意義並且減少所需設定的參數值。實驗的環境為台灣50、宏達電、深證成分股指數和標準普爾500,預測的測試標準為平均相對誤差絕對值(Mean Absolute Percentage Error, MAPE)和均方誤差(Mean Square Error, MSE)。兩個提出的系統都有很好的預測準確度。
Stock price predicting is an important concern for investors, who by using high accuracy prediction systems are able to make a great profit. In recent years, artificial neural networks (ANNs) have shown promising results in this area, and have been improved in many ways. However, there are still some issues with ANN that remain unanswered. Different combinations of setting parameters bring about different consequents, such as the constitution of input nodes, hidden nodes, and initial values of weight. Hence, we propose a simple but useful method, which only uses stock closing prices as inputs and experiments with different kinds of setting parameters. In addition, normalized function of backpropagation neural network (BPN) is enhanced with a novel idea. The proposed method has promising result. Nevertheless, because of the complicated formula, the meaning of each weight is not able to understand directly. Therefore another method is proposed. With innovated weight changed formula, the input resemble the predicting target is found. System is applied to Taiwan’s Top 50 Exchange Traded Fund, HTC Corporation, S&P 500, and Shenzhen Composite in order to forecast the next day closing price. For an evaluation of effectiveness and accuracy, mean absolute percentage error (MAPE) and mean square error (MSE) are used. Given the experimental results, the first proposed method shows excellent performance with the best set of parameters, and the innovative normalization method effectively improves accuracy. To make comparison, GA-WNN is chosen, and our first method shows great results. What is more, the second novel method demonstrates the meaning of every weight outstandingly.