機器學習應用於新制 IFRS 下之每股盈餘預測與套利分析——考量三因子風險溢酬

Abstract

本研究以2018年後IFRS 9、15、16新制實施後之臺灣上市櫃公司財報與財務比率資料為基礎，運用四種機器學習與深度學習模型——隨機森林（Random Forest）、梯度提升（Gradient Boost）、極限梯度提升（XGBoost）與深度神經網路（DNN）——進行調整每股盈餘（Adjusted EPS）及調整稀釋每股盈餘（Adjusted Diluted EPS）之預測，並據以建構多空套利策略。本研究旨在探討 IFRS 新制下財報資訊與市場預期間的落差是否具備可轉化為投資報酬的潛力。 研究結果顯示，Random Forest 與 Gradient Boost 在多數情境下具較佳的預測準確性，而 XGBoost 雖預測誤差略高，卻在套利策略穩定性與排序能力方面展現潛力。DNN 模型因樣本數限制，易出現過度擬合問題，整體表現不如其他模型。進一步分析發現，預測準確性不必然轉化為顯著報酬，預測誤差本身需具備資訊含量方能成為套利來源。在風險調整報酬方面，XGBoost 在Fama-French三因子模型下展現最高的解釋力與正向 Alpha，特別是在等權重策略下表現較佳。 整體而言，本研究證實機器學習模型能有效擷取財報中的非線性關係與潛在訊號，並有助於因子構建與策略設計，並補足新制 IFRS 下每股盈餘預測與投資應用領域之文獻空缺，並提供財報使用者與投資實務參考依據。

This thesis analyzes financial statements and ratio data of Taiwan-listed firms after the implementation of IFRS 9, 15, and 16 in 2018. Four predictive models—Random Forest, Gradient Boosting, XGBoost, and Deep Neural Network (DNN)—are applied to forecast Adjusted EPS and Adjusted Diluted EPS, serving as the basis for long-short arbitrage strategies. The study explores whether discrepancies between accounting data and market expectations can generate investment profits. Results show that Random Forest and Gradient Boosting generally outperform in prediction accuracy, while XGBoost, though slightly less accurate, excels in strategy stability and ranking effectiveness. DNN suffers from overfitting due to limited data. Importantly, prediction accuracy alone does not ensure return potential; only errors containing informative signals support arbitrage. XGBoost delivers the highest risk-adjusted returns and alpha under the Fama-French three-factor model, particularly in equal-weighted portfolios. Overall, the study highlights the usefulness of machine learning in capturing complex patterns in financial data, contributing to EPS prediction and investment strategy design under the new IFRS framework.