股票的可預測性研究在金融學與經濟學界中有著悠遠的歷史,有些是秩序性狀況是簡單且可以預測,傳統使用線性迴歸方法可以找出固定的漲跌方向;但市場中存在許多不確定原因影響著股價的波動,隨機性特質在系統本質上是無法預測的,隨著時間不斷的進步與類神經網路的興起,股價也多數採用時間序列分析方法進行預測,深度網路經常會遇到梯度爆炸或梯度消失的問題,導致整個模型運算失敗。 本研究以IBM SPSS軟體進行迴歸分析方法與深度學習方法訓練,以2017年1月3日至2019年12月30日之美國四大指數收盤價對台灣加權股價指數進行預測與討論,迴歸分析做t值與F檢定比較,再將預測結果進行殘差分析與變數重要性分析比較;研究發現資料集與預測值的型態會影響模型的選擇,具有關聯的多變數聚合投入模型不一定會取得好的預測結果,在選用變數需花更多時間去挑選與測試。
The study of the predictability of stocks has a long history in the field of finance and economics. Some of the orderly conditions are simple and predictable. The traditional use of linear regression methods can find a fixed direction of ups and downs; but there are many uncertainties in the market. The reason affects the fluctuation of stock prices. The randomness of the system is inherently unpredictable. With the continuous progress of time and the rise of neural networks, most stock prices are predicted by time series analysis methods. Deep networks often encounter problems. To the problem of gradient explosion or gradient disappearance, causing the entire model operation to fail. This research uses IBM SPSS software to train regression analysis methods and deep learning methods, and uses the closing prices of the four major US indices from January 3, 2017 to December 30, 2019 to predict and discuss the Taiwan Weighted Stock Price Index. The t value is compared with the F test, and then the prediction results are compared by residual analysis and variable importance analysis; it is found that the type of data set and predicted value will affect the choice of model, and the multivariate aggregated input model with correlation may not be obtained. Good prediction results require more time to select and test when selecting variables.