使用機器學習與深度學習模型，比較平減與產業對每股盈餘金額預測與套利結果的影響

Abstract

本論文係探討平減與產業對於訓練機器學習模型和深度學習模型之預測成效與最終套利結果之影響。除了借助資產負債表、綜合損益表與現金流量表，並針對近年所發布之國際財務報導準則，調整上述相關之會計科目和金額外，亦納入平減與產業之考量。接續分別使用隨機森林迴歸器、極限梯度提升迴歸器以及深度神經網路模型，以預測隔年度之調整後每股盈餘與調整後稀釋每股盈餘。最後透過預測結果，建立套利組合並進行後續為時一年之套利實證。 經過本論文之研究後，得出以下五個結論：第一，隨機森林迴歸器為預測能力最佳的模型，而深度神經網路模型則相反；然而於套利結果中，則是後者的表現最佳，而前者最差。第二，於全產業與電子業下，針對調整後每股盈餘與調整後稀釋每股盈餘進行平減，有較大之機會提升預測的能力；然而於非電子業下，則建議不要進行平減，因為有較大的機率使預測能力下降。至於平減對於套利結果之影響，經研究發現無論是預測調整後每股盈餘或是調整後稀釋每股盈餘，未平減之套利結果普遍較佳。第三，研究未發現存在特定產業有較佳之預測結果。但如果使用隨機森林迴歸器時，經研究發現非電子業的預測結果優於全產業與電子業；如果使用極限梯度提升迴歸器時，則僅於未平減之條件下，非電子業的預測結果才優於另外兩種產業。至於產業對於套利之影響，經研究發現到不同產業搭配不同的套利條件，會有不一樣的套利結果，因此得出不存在特定產業具有較佳的套利結果之結論。第四，將資料拆分成電子業與非電子業進行預測較未拆分之全產業所得之預測效果差。最後，發現到透過深度神經網路模型於調整後稀釋每股盈餘未平減之非電子業下，使用第一種套利因子並且按市值權重的方式做多與做空前後10%之公司，將可以賺取最高的報酬率。

This thesis investigates the impact of smoothing and industry on the predictive performance and final arbitrage outcomes of machine learning and deep learning models. In addition to utilizing balance sheets, income statements, and cash flow statements, this study adjusts the relevant accounting items and amounts pertaining to the recently published International Financial Reporting Standards. It also incorporates considerations of smoothing and industry. Subsequently, the study employs various predictive models, including the random forest regressor, extreme gradient boosting regressor, and deep neural network model, to forecast adjusted earnings per share (EPS) and adjusted diluted EPS for the upcoming year. Finally, based on these predictions, arbitrage portfolios are constructed to engage in arbitrage activities over the next year. Following the research conducted in this thesis, five key conclusions are drawn: First, the random forest regressor emerges as the model with the highest predictive capability, while the deep neural network model exhibits the opposite performance; however, in terms of arbitrage outcomes, the latter demonstrates superior performance, whereas the former shows the poorest results. Second, within both the overall industry and specifically the electronics sector, smoothing adjusted EPS and adjusted diluted EPS tends to enhance the predictive accuracy. Conversely, it is recommended to avoid smoothing in the non-electronics sector, as it tends to reduce predictive performance. Regarding the impact of smoothing on arbitrage results, it is found that non-smoothed predictions generally yield better arbitrage outcomes, whether forecasting adjusted EPS or adjusted diluted EPS. Third, the research does not identify any specific industry that consistently achieves better predictive results. However, when employing the random forest regressor, predictive results for the non-electronics industry are found to surpass those of the overall and electronics industries. In contrast, when employing the extreme gradient boosting regressor under non-smoothed conditions, the predictive performance of the non-electronics industry surpasses that of both the overall industry and the electronics sector. As for the influence of industry on arbitrage outcomes, different industries yield varied results depending on the arbitrage conditions employed, leading to the conclusion that no single industry consistently offers better arbitrage opportunities. Fourth, segmenting the data into electronics and non-electronics industries for prediction yields poorer results compared to models that consider the entire industry without segmentation. Lastly, it is discovered that using the deep neural network model to forecast non-smoothed adjusted diluted EPS in the non-electronics industry, while employing the first arbitrage factor and implementing a market value-weighted strategy to go long and short on the top and bottom 10% of companies, results in the highest return rates.