應用倒傳遞類神經網路研究台灣機構投資人的交易行為

Abstract

本研究運用倒傳遞類神經網路方法，針對外資及陸資、自營商、投信等機構投資人在證券、期貨與選擇權市場的交易行為進行深入研究。主要藉由類神經網路在處理非線性問題的優越能力，找出機構投資人的交易行為與市場走向的內隱知識。目的在於比較機構投資人在證券、期貨、選擇權市場的交易行為，對於台灣指數期貨市場的未來走勢有較好的預測能力；並進一步比較，距期貨到期日的時間因子，對於倒傳遞類神經網路的預測能力是否有影響。 本研究結果顯示，實驗一，在以機構投資人證券買賣金額、臺灣證券交易所股價指數期貨交易資料、選擇權市場交易資料做為倒傳遞類神經網路的輸入值，並以台灣股票指數期貨的成交值日資料做為目標值，所訓練出的倒傳遞類神經網路中，只有實驗組-期貨可以在「交易策略總損益」一項績效評估指標打敗對照組-隨機漫步，且優於其他實驗組-證券與選擇權的表現；實驗二，實驗組-期貨保留時間因子的倒傳遞類神經網路，在訓練類神經網路時，若扣除距期貨到期日時間因子，將導致「交易策略總損益」與「交易策略最大連續損失」兩項績效評估指標表現明顯變差。 因此，利用機構投資人在期貨市場的交易行為所訓練出的倒傳遞類神經網路，在「交易策略總損益」一項指標有不錯的表現；並當倒傳遞類神經網路，在排除距到期日時間因子做為輸入值後，會讓「交易策略總損益」與「交易策略最大連續損失」兩項指標表現變差。故本研究所推出之模型，可以在台股期貨市場有較好的表現，該結論可做為學術及一般獨立投資人之參考。

This study applies Back-Propagation Neural Network for investigating the trading behavior of Institutional Investors (Foreign & China Investor, Security Dealer, Investment Trust) in Security, Futures & Option Markets. By using the outperform ability dealing with non-linear problems of Neural Network, we can find the relationship between trading behavior of Institutional Investors and the Market Trend. One wonders which predictive power is better, and we also analyze whether the time factor of due date is critical for the predictive power. The empirical results indicate that only the Neural Network trained with trading behaviors of institutional investors in TAIFEX Future Market, not Trading Value of Foreign & Other Investors or TAIFEX option, outperforms that of Random Trading Model in the Evaluation Indexes of predictive winning ratio and profit of trading model. In addition to the first experiment, if we deduct the “how many days before the settlement day” time factor from our training data, we find that there are significant bad performances in the Evaluator Index of “trading profit” & “trading minimum continuous loss.” Therefore, we can show that containing the time factor will make a better performance of the trading profit with a minimum continuous loss. Moreover, the trading behavior of Institutional Investors has effective ability to the Future Market, and the time factor before Future Settlement also affects the prediction of the Neural Network. Additionally, the trading behavior of Institutional Investors is worthy of reference for Individual Investors.