台灣5G相關產業股票報酬的分析與預測：基於機器學習方法

本研究以台灣5G相關產業公司2015/07至2018/12資料為基礎，建構線性之固定效果面板迴歸模型與非線性之極限梯度提升樹模型（XGBoost）以探討總體經濟面、個別公司面與籌碼面對股票報酬之影響，並比較線性與非線性模型結論之差異。研究結果顯示，不論在線性或非線性模型，籌碼面對股票報酬影響皆最大，本研究推測原因為5G產業較具未來前景，儘管基本面不好，投資人仍願意給予較高的估值並進場購買，導致籌碼面影響最大；而總體經濟面與個別公司面在線性與非線性模型中對股票報酬有不同的影響，在非線性模型中，總體經濟面重要因子平均增益大於個別公司面，對股票報酬影響較大。此外，本研究加入技術分析面建構極限梯度提升樹預測模型，對未來一個月股票之漲跌進行預測，其模型預測能力於樣本內獲得87.14%的準確度，而樣本外獲得56.52%的準確度。

關鍵字

股價報酬；機器學習；預測模型；面板迴歸；極限梯度提升

並列摘要

This study builds a linear fixed-effect panel regression model and nonlinear extreme gradient boosting tree model on Taiwan’s 5G related companies, data period is from 2015/07 to 2018/12, study the impact of economic factor, company factor and chip factor on stock return. The results show that whether in linear or nonlinear model, chip factor has the greatest impact on stock return. This study speculates that the reason chip factor has the greatest impact on stock return is that 5G industry is a promising industry, although weak fundamentals, investors are still willing to give a high valuation and enter the market. While economic factor and company factor have different impact on stock return in linear and nonlinear model, in nonlinear model, economic factor has more impact on stock return than company factor because of greater average gain. Additionally, this study constructs an extreme gradient boosting tree model to predict whether stock rise or fall in the next month by join the technical analysis factor, the model achieved 87.14% accuracy from in-sample forecast and 56.52% from out-of-sample forecast.