證券交易決策支援系統-以類神經網路實作

Abstract

從目前類神經網路應用領域的研究來看，類神經網路運用於投資決策已是未來的一個發展趨勢。而本研究是以類神經網路來預測未來趨勢，並根據預測值定出一買賣決策依據。 本研究首先針對每一技術指標單獨訓練，以找出與推論輸出較相關之輸入變數，並對輸入變數加上移動平均、變動率、乖離率等處理，使其更能擷取出股價行為模式。 在實驗的過程中發現，當訓練過程收斂到某一程度以後，過度的訓練反而使得測試範例的誤差加大，因此必須找出一合適的控制方法以免過度訓練，使得測試誤差加大又浪費訓練時間，因此，為了避免訓練過度，本研究使用判定係數來控制訓練的過程。 為了隔離出上下震盪的股價變動，正確地預測出有效的反轉點，本研究採用趨勢指標的３日移動平均線，作為未來趨勢的確認，另外，本研究以事先分類的方法，事先依股償所在的走勢特性將每日資料分成持續上漲、持續下跌、由上漲轉下跌、由下跌轉上漲四類，根據不同走勢特性，資料被送到不同的類神經網路學習，經過實證，此方法對於增進收斂的速度、提高學習的精度有極大的改善。 最後，本研究根據上述買賣規則在1993年１月６曰月間模擬買賣操作 ，由買賣測試結果可知，總獲利百分比分別為50.51%、52.41%，優於長期持有，由此可見，本研究之模式是樂觀的。

The application of Neural Networks, from its recent findings of studies, may most likely become future trend in investment decision making. The following study is to forecast the stock future through Neural Networks and hence to determine the optimum point of buy-and-sell. To better set up the behavior models of stock pricing, this study was set forth with the identification of various input and output variables through numerous try outs of technical index perspectively. Each input variable was then screened with the analysis of moving average, rate of variance, and the rate of separation. There was a special finding in the search of technical indexes. Some confusing contradiction might result from over testing of certain variables and hence conclusion misht be distorted as well as variable time wasted. To improve this situation, coefficient analysis was then added in this research. To separate the normal fluctuation of stock prices from its unique pricing behavior and to forecast the turning points of stock prices more effectively, 3-day moving average was adopted in this study. Besides this, each stock was pre-categorized into four groups: going-Lip, going-down, first lip then down, first down then up, according to its own group. It has been confirmed that this method does not only improve the speed of elimination, but also increase the precision of testing. Using the aforementioned method, a simulation of buying and selling was put in force from Jan. 1993 to June 1993 and produced gross profitability of 50.51% and 52.41% respectively, which were higher than long-term holding. Supported by the result of this study, I am gladly convinced that the model used in this study is more than promising.