機器學習模型應用於單季每股盈餘預測及投資套利策略探討

Abstract

本論文欲探討財務報表季節性效應是否能使機器學習模型有更好的預測結果，以及縮短每期套利期限並增加一年之套利次數是否更能夠捕捉到股票價格之變動，並提高整體報酬率。本論文使用隨機森林迴歸器、極限梯度提升迴歸器、自適應增強迴歸器對過去八季以及前三年同季之三大財務報表的會計科目金額進行機器學習，並預測下季度的調整基本每股盈餘以及調整稀釋每股盈餘。 實證結果顯示，以去年同季度的每股盈餘與前三季平均成長變化相加的絕對誤差函數最低，顯示近期季度對公司每股盈餘變化的影響更大。在機器學習模型中，隨機森林迴歸器是唯一具有低於基準標竿的絕對誤差函數的模型。在套利方面，以隨機森林迴歸器預測調整基本每股盈餘並與去年同季度的真實值進行比較後以加權市值法買進與放空前後10%的零成本套利策略有最佳結果，每年皆能獲得至少14%的報酬率，最高能獲得65%，平均每年報酬率可達到30.324%，六年總計181%報酬率。最後，本論文的結論認為機器學習無法捕捉到財務報表季節性效應的影響，但是投資者更專注於季度之間的成長關係，導致縮短每期套利期限並增加一年之套利次數確實能夠提高投資者在股票交易市場中的套利能力。

This study aims to investigate whether the seasonal effects in quarterly financial statements can lead to better predictive results for machine learning models, and whether shortening the arbitrage period and increasing the frequency of arbitrage over a year can better capture changes in stock prices and improve overall returns in Taiwan Stock Market. The study utilizes Random Forest Regressor, XGBoost Regressor, and AdaBoost Regressor to learn from quarterly financial statements over the past periods, and predicts the adjusted basic earnings per share and adjusted diluted earnings per share for the next quarter. Empirical results show that among the three machine learning models, random forest regressor is the only model with an absolute error function lower than the three benchmarks. In terms of arbitrage, employing the random forest regressor to predict adjusted earnings per share and comparing it with the real value of the same quarter of the previous year, and then using a zero-cost arbitrage strategy of buying the top 10% highest growing companies and selling the top 10% highest shrinking companies, yields the best results. This strategy achieves an average annual return of 30.324% over six years. Finally, the study concludes that machine learning cannot capture the seasonal effects in financial statements, but shortening the arbitrage period and increasing the frequency of arbitrage over a year can indeed enhance investors' arbitrage capabilities in the stock market.