應用類神經網路以移動視窗法及單次訓練法預測臺灣加權股價指數變動趨勢

摘 要 本文應用基金經理人常用之技術指標作為類神經網路預測模型的輸入變數。基於 多層誤差倒傳遞類神經網路（Multi-Layers Backpropagation Artificial Neural Networks， MLBPN）建立類神經網路預測模型用於預測台灣加權股價指數漲跌幅。以2日、5日、 10日、15日及22日加權股價指數漲跌幅為模型輸出變數，分別以230日及1095日訓練組 數據對進行單次類神經網路訓練及移動視窗法類神經網路訓練，將訓練結果應用於預 測台灣加權股價指數變動趨勢進行比較。 實證結果顯示預測2日加權股價指數漲跌幅之方向，準確度最高且訓練完成之收歛 誤差最小，其次依序為5日、10日、15日、22日，而1,095日的訓練模型對於2日加權股 價指數漲跌幅有最佳的預測準確度，由實證得知以較多的訓練資訊訓練類神經網路預 測模型可以得到較高的準確度。
ABSTRACT This research applies technical indexes, which usually used by the fund manager, to be the input variables for a neural predictive model. The neural predictive model based on the Multi-Layers Backpropagation Artificial Neural Networks is used to predict the variations insector stock index of Taiwan. The output variables of the proposed neural predictive model are the variations of 2、5、10、15 and 22 days stock indexes. The data pairs of 230 and 1095 days are used to train the neural models by conventional training, and the data pairs of 230 days are also used to train the neural predictive model by moving window training. The simulation results reveal the prediction of 2 days variations of stock indexes is best accurate with least convergent error, and 5, 10, 15 and 22days prediction is second, third, forth, fifth respectively. The 2 days prediction with 1,095 days training data pairs has the best accuracy. It means more information for training will obtain better predictive accuracy.