應用機器學習方法預測台指期貨隔日漲跌之研究

A Study on Predicting the Trend of Taiwan Stock Index Futures Using Machine Learning method

黃馨儀(Hsin-Yi Huang) ； 黃華山(Hwa-Shan Huang) ； 林俊騰(Jun-Teng Lin)

《華人經濟研究》 18卷2期 (2020/09) Pp. 55-80

摘要

在台指期貨市場裡，有人將其視為避險工具，亦有人將它視為高槓桿的投資標的，而各式各樣的技術指標都有各自的愛好者。然而，投資行為的目的即為探究哪些技術指標能預測未來趨勢。因此，本研究使用23種技術指標，並建立三個模式，分別採用機器學習中的決策樹、類神經網路，以及結合決策樹與類神經網路兩者等三種方法，期望找出能預測隔日漲跌的重要變數與建立高預測隔能力的模型。本研究採台指期貨創辦以來的12年多的資料，共計3164筆，將分為訓練資料集2335筆與測試資料集829筆，並將台指期貨分類為平盤、大漲、大跌三類來建立模型。研究結果顯示，以決策樹和類神經網路結合的模型三最具有預測能力，準確度高達85.89%，且模型三在預測錯誤的部分，將大漲預測為大跌或大跌預測為大漲的次數為0，依成本角度而言，大幅度降低分類錯誤所要付出的成本。而就投資的觀點，模型三不但有不錯的預測能力，也可大大降低投資人在期貨市場中的投資風險，在近三年確實能有效預測台指期貨隔日漲跌。

關鍵字

機器學習 ； 決策樹 ； 類神經網路 ； 台指期貨

並列摘要

Some people regard the Taiwan index future as a hedge tool or an investment target with high leverage in the futures market. Every technical indicator has its own fanciers. However, the purpose of investment behaviors is to know which technical can forecast future trends. Therefore, this study use 23 kinds of technical indicators and establish 3 molds by using decision tree, back propagation neural networks and the combination of both two techniques. We expect to find out the important variables which could forecast the next day's price fluctuation and build up a model with high predicting ability. The study adopts the data of Taiwan Index Future for 12 years with total amount of 3,164 observations which will be separated into training set (2,335) and testing set (829). We also divided the sample into three groups including market tick, up and down to build model. The study result shows that the most powerful model is using combined type which the accuracy rate is 85.89%. On the investment point, model 3 can effectively predicate the up and down of next day in three years.

並列關鍵字

Machine Learning ； Decision Tree ； Neural Network ； Taiwan Index Future