應用類神經網路於臺灣期貨指數極短線走勢行為知識發現

Abstract

臺灣交易市場為提供投資人更即時資訊並符合市場需求，臺灣證券交易所於民國100年元月17日將指數與成交及委託統計資訊揭露頻率由現行的1分鐘縮短為15秒，與亞洲地區交易所接軌。 本研究提出以倒傳遞類神經網路架構，搭配技術分析指標，學習臺灣加權指數期貨日內極短線的趨勢行為，嘗試找出市場的知識規則，建立極短線趨勢預測模型。在極短線下，交易時常處於盤整期或是微小跳動，本研究為了克服此問題，提出資料平滑化和過濾盤整期資料兩個資料處理的方法，嘗試增進倒傳遞類神經網路的學習效果。 由實驗結果得知，在極短線時資料蘊含的知識量不足，若輔以的額外的資料前處理方法，資料平滑化和過濾盤整期資料，可以得到較高的準確率和獲利能力。

In order to provide timeliness information and meet the market demand, Taiwan Stock Exchange (TWSE) has increased the frequency of statistical information disclosure on TSEC weighted index from originally 1 minute to 15 seconds on January 17, which is now synchronized with Asia exchanges. In this study, we propose a neural network architecture utilizing technical analysis indicators, attempting to find the knowledge rules from TAIEX Futures’ intraday trading data and construct a TAIEX Futures’ short-term time interval predicting model. In short-term time interval, the transaction data often changes in a small range; we propose two data pre-processing techniques to overcome this problem and improve neural network’s learning ability. The results of experiments indicates that although the transaction data contains insufficient information and knowledge in a short-term time interval, it is possible to achieve better accuracy and profitability through data smoothing and data filtering.