本文旨在開發一種預警模型，以提前發現盈餘管理水平。由於盈餘管理，會造成不透明性，從而降低了利益相關者評估公司實際業績的能力，也破壞了資本市場的運作。研究人員已經廣泛研究了通過操縱公司的會計應計利潤進行的盈餘管理，但是，這些研究很少關注盈餘操縱的程度，而是探索了特定因素與盈餘管理之間的相關性。盈餘管理的計算水平與Dechow等人一致。 （1995）誰執行了改良的瓊斯模型。文本數據是從年度報告中的管理討論和分析（MD＆A）中提取的。由於年度報告中披露的信息包括許多專業術語和特定註釋，因此可讀性水平的日益下降已經對公司年度報告的傳遞功能產生了負面影響。有目的的是，這項研究能夠激勵公司使用更清晰，更易混淆的語言來提高年度報告的清晰度和可理解性。數據收集自2017年至2018年的《台灣經濟日報》數據庫（TEJ），主要針對集成電路（IC）製造商。然後將分析的數據輸入到人工智能（AI）技術中，以構建用於盈利管理預測的模型。通過提出一種非並行超平面支持向量機（NHSVM），它將傳統SVM的大QPP問題分解為兩個小QPP問題。為了增強研究結果，本研究還以提出的模型為基準，並將其與其他兩個模型（決策樹（DT）和神經網絡（NN））進行比較。結果表明，在所有評估標準下，該模型均優於其他兩個模型。為了防止結果僅因巧合而被遮蓋，我們進行了統計檢驗。經實際案例檢驗的模型是盈利管理預測的有希望的替代方法。
This paper is conducted to develop an early warning model to detect the level of earnings management in advance. Due to earnings management could create opacity which reduces the stakeholders’ ability to assess the real performance of the firm and also break down a functioning of capital markets. Earnings management through the manipulation of firm’s accounting accruals has been extensively investigated by researchers, however, these studies barely focus on a degree of earnings manipulating but rather explored the correlation between a specific factors and earnings management. The level of earnings management calculation is in accordance with Dechow et al. (1995) who performed modified Jones model. The textual data is extracted from management discussion and analysis (MD&A) in annual reports. The increasing deterioration in levels of readability has negatively affected the transmission function of firm annual reports due to the information disclosed in annual reports includes many professional terms and specific notes. Purposefully, this study is able to stimulate corporates to improve the clarity and understandability of their annual report by using language that is clearer and less convoluted. The data are collected from Taiwan Economic Journal databank (TEJ) ranging from 2017 to 2018 by targeting Integrated Circuit (IC) manufacture. The analyzed data are then fed into artificial intelligence (AI) technique to construct the model for earning management forecasting. By proposing, a non-parallel hyperplane support vector machine (NHSVM) that decomposes the conventional SVM’s big QPP problem into two small QPP problems. To robust the research findings, this study also takes the proposed model as a benchmark and compares it with the other two models which are decision tree (DT), and neural network (NN). The result shows that the proposed model outperforms the other two models under all assessing criteria. To prevent the results just happed by coincidence, the statistical test is conducted. The model, examined by real cases, is a promising alternative for earning management forecasting.
