財報盈餘預測及套利可能探討—序列深度學習模型應用

Abstract

每股盈餘是許多投資人評估公司盈利能力的關鍵指標，公司每股盈餘的表現是影響投資人決策的重要因素之一，若能夠對公司每股盈餘進行準確的預測，將成為可靠的投資依據。本論文使用公司過去資產負債表、綜合損益表及現金流量表中的會計科目，透過建立 GRU、LSTM、雙向LSTM、Transformer 及 TST 等 5 個深度學習模型以及隨機森林和 XGBoost 等 2 個機器學習模型，對公司下一年度之調整每股盈餘及調整稀釋每股盈餘進行預測，再根據預測結果同時部署多頭及空頭部位，形成零成本投資組合，以實證人工智慧技術的套利可能性。 本論文建立之深度學習模型及機器學習模型在預測能力上皆優於基準值，其中以遞迴神經網路模型的預測能力表現最佳，深度學習模型整體優於機器學習模型。此外，本論文對於自變數使用了兩種特徵維度縮減方法，其中自編碼器方法能夠顯著的提升本研究所建立之深度學習模型的預測能力。然而，雖然機器學習方法之預測能力不如深度學習方法，但若以市值加權投資策略觀察套利結果，XGBoost是報酬率最高的模型，其他模型在適當的投資比例下亦存在套利空間；而以平均加權投資策略觀察套利結果，則所有模型均存在套利空間。

The earnings per share (EPS) stands as a crucial measure for many investors when evaluating a company's profitability, the performance of a company's EPS is one of the significant factors influencing investor decisions. An accurate prediction of a company's EPS can serve as a reliable basis for investment. This study utilizes accounting information from the company's historical published balance sheets, comprehensive income statements, and cash flow statement. Five deep learning models (GRU, LSTM, Bidirectional LSTM, Transformer, and TST) as well as two machine learning models (Random Forest and XGBoost) are applied to predict the adjusted EPS and adjusted diluted EPS for the upcoming fiscal year. These predictions are then used to form a zero-cost investment portfolio, incorporating both long and short positions, to demonstrate the arbitrage potential of artificial intelligence technology. This study demonstrates the capacity of both deep learning and machine learning models to predict the adjusted earnings per share and adjusted diluted earnings per share for the upcoming fiscal year. All models exhibit superior performance compared to the baseline; the Recurrent Neural Network model exhibits the strongest predictive capability. Additionally, two methods for dimensionality reduction were applied in this study, the autoencoder method significantly enhanced the performance of the deep learning models. However, while the predictive ability of machine learning methods trails behind that of deep learning methods, when evaluated through a market-capital weighted investment strategy, the XGBoost model demonstrates the highest return rate. At appropriate investment ratios, other models also demonstrate potential arbitrage opportunities; On the other hand, under an equal-weighted investment strategy, all models exhibit potential arbitrage opportunities.