不同財務資訊對每股盈餘預測和套利報酬之比較- 機器學習模型應用

Abstract

本論文使用財務報表及財務比率建立三種訓練集，並透過隨機森林、梯度提升、極限梯度提升及深度神經網路四種演算法預測調整每股盈餘及調整稀釋每股盈餘，利用預測之調整每股盈餘及調整稀釋每股盈餘計算套利因子並建立投資組合，藉此驗證使用財務比率是否較使用財務報表有較佳之預測和套利能力。 結果發現隨機森林和梯度提升兩種演算法在預測調整每股盈餘及調整稀釋每股盈餘之能力較佳，且相較於使用財務報表，使用財務比率有較佳之預測能力。三種訓練集透過四種演算法所預測之調整每股盈餘及調整稀釋每股盈餘皆能建立具套利空間之投資組合，財務比率僅在應變數為調整每股盈餘，且利用隨機森林和梯度提升演算法進行預測時比使用財務報表有更佳之表現，這顯示出每股盈餘及調整稀釋每股盈餘之能力和套利報酬並非呈現完全正相關。另外，透過深度神經網路所建立之投資組合較其他演算法有較高之套利報酬。

This thesis uses financial statements and financial ratios to establish three types of training sets and employs four algorithms—Random Forest, Gradient Boosting, Extreme Gradient Boosting, and Deep Neural Networks—to predict adjusted earnings per share (EPS) and adjusted diluted EPS. The predicted adjusted EPS and adjusted diluted EPS are then used to calculate arbitrage factors and create investment portfolios. This thesis aims to verify whether using financial ratios provides better predictive and arbitrage capabilities compared to using financial statements. The results indicate that the Random Forest and Gradient Boosting algorithms have superior predictive power for adjusted EPS and adjusted diluted EPS. Additionally, using financial ratios offers better predictive power than using financial statements. All three training sets with the four algorithms to predict adjusted EPS and adjusted diluted EPS can establish investment portfolios with arbitrage opportunities. Financial ratios demonstrate better performance than financial statements only when the dependent variable is adjusted EPS and predictions are made using the Random Forest and Gradient Boosting algorithms. This indicates that the capability to predict EPS and adjusted diluted EPS and the arbitrage returns are not perfectly positively correlated. Moreover, investment portfolios created through Deep Neural Networks exhibit higher arbitrage returns compared to those created by other algorithms.