- 學位論文
VPIN 於股價預測應用分析: 以決策樹與隨機森林模型為例
An analysis for stock price prediction by VPIN based on decision tree and random forest models
摘要
本研究針對兩家資訊揭露評鑑高與一家低的股票，使用機器學習中 的決策樹(Gini) 、決策樹(Entropy)和隨機森林三個模型，驗證同步交 易量知訊交易機率(VPIN)技術指標對於下一個 VPIN 的時間點的股價， 預測漲與跌是否有幫助。透過一些研究中常用的技術指標和市場指標當 作控制變數，共 27 種特徵加入模型做預測，得出的結果並與分別加入 第 28 種特徵 PIN 或 VPIN 做比較。我們發現加入 PIN 或 VPIN 能明顯的改善 Type I error，故相較於 Recall 和 F1-Score 而言，Precision 績 效指標顯著改善。整體樣本期間若加入 PIN 或 VPIN，評鑑高與低的公司於隨機森林模型下的 Precision 皆呈現顯著改善;惟在金融風暴期間， 評鑑高的公司僅在加入 VPIN 後，採用決策樹(Gini)和決策樹(Entropy) 模型表現優於隨機森林模型，即兩個決策樹模型其 Precision 和 Recall 能顯著改善。總體上，研究結果顯示 PIN 或 VPIN 兩項指標皆能改善預 測效果，而 VPIN 略優於 PIN。
並列摘要
This paper examines the high and low stocks of the information disclosure assessment, using the three models of the decision tree (Gini), decision tree (Entropy) and random forest by machine learning to verify the synchronous transaction volume (VPIN). Through some technical and market indicators as control variables, a total of 27 characteristics were added to the model for prediction, and the results were compared with the addition of the 28th feature PIN and VPIN, respectively. We found that adding PIN or VPIN can significantly improve Type I error, therefore, we focus on precision performance. If VPIN is added during the overall sample period, both the high and low companies under the random forest model will show significant improvement. However, in the financial turmoil period, the companies with high information disclosure assessment adopt the decision tree (Gini) and decision tree (Entropy) model, which could improve the precision and recall. Overall, the research showed that both PIN and VPIN improved predictive outcomes, while VPIN was slightly better than PIN.