每股盈餘預測及套利投資策略探討-機器學習模型應用

Abstract

本論文使用公司過去發佈之資產負債表、綜合損益表和現金流量表之會計科目資訊，搭配三種資料特徵處理方式，並分別使用隨機森林迴歸器、極限梯度提升迴歸器和長短期記憶模型，以預測下一年度之調整每股盈餘和調整稀釋每股盈餘。另外，本論文接著使用模型所預測下一年度的調整每股盈餘和調整稀釋每股盈餘，以該數值相對於本年度調整和調整稀釋每股盈餘之漲跌為基礎，建立不同投資組合以實證套利之可能。 經過模型和實證結果，即使在樣本期間具新會計公報發行或疫情之影響，本論文得到以下三個結論：第一，以決策樹為基礎的機器學習模型和深度學習模型皆有預測下一年度調整每股盈餘和調整稀釋每股盈餘之能力，其中又以深度學習之長短期記憶模型表現更為精確。第二，三種資料特徵處理方式中，以所使用的第三種資料特徵處理方式—以流通在外股數為基礎做資料標準化—較為準確。最後，本論文發現在假設市場完美、沒有交易成本下，在每一個年度預測下一年度調整每股盈餘和調整稀釋每股盈餘所形成之投資組合中，以當期年度的真實值為基準下，使用長短期記憶模型幾乎皆具有套利之空間，且使用預測較準之長短期記憶模型搭配第三種資料特徵處理方式，更為套利結果前五名。

This study utilizes accounting information from the historically published balance sheets, income statements, and cash flow statements of companies. Three different data preprocessing methods are applied, and three models—Random Forest Regressor, Extreme Gradient Boosting Regressor, and Long Short-Term Memory (LSTM) model—are used to predict the adjusted earnings per share or adjusted diluted earnings per share for the next fiscal year. Additionally, based on the predicted values, the study establishes different investment portfolios to explore potential empirical arbitrage opportunities. Based on the model and empirical results, even with the impact of the adoption of IFRS 9, IFRS 15, and IFRS 16 and the pandemic during the sampling period, this paper draws the following three conclusions: Firstly, both tree-based machine learning and deep learning have the ability to predict the adjusted earnings per share and adjusted diluted earnings per share for the next fiscal year. Among them, deep learning, specifically the LSTM model, performs more accurately. Secondly, among the three data feature selection methods, the third method, which involves normalizing the data based on the number of outstanding shares, is more accurate. Lastly, this paper finds that assuming a perfect market with no transaction costs, in each year's forecast of the next fiscal year's adjusted earnings per share and adjusted diluted earnings per share, there is consistently an arbitrage opportunity when using the LSTM model along with the third data feature selection method. This combination ranks among the top five in terms of arbitrage results.