許多的研究探討企業是否在財務報表忠實表達其經濟實質，盈餘管理便是其中一個議題關於管理階層操縱盈餘來符合某些目的。在近幾年的研究中，資料探勘技術已經廣泛的被應用在商業活動的相關領域中。本研究也嘗試利用資料探勘的方式偵測盈餘管理並建立三階段模型，首先，利用逐步迴歸以及隨機森林(random forest)演算法將變數進行篩選；其次，利用三種主要的研究方法諸如多層感知器(MLP)、支援向量機(SVM)以及最小平方法支援向量機建立模型並評估分類績效；最後在第三階段的部分。利用決策樹找出影響盈餘管理的相關的規則。本研究之樣本來自台灣新報資料庫，範圍為台灣2008至2012年台灣公開上市電子產業，用以建立模型的變數共有19個如公司績效、來自營運活動之現金流量以及其他財務及非財務指標。實證結果指出，LS-SVM較SVM和MLP在探測盈餘管理方面擁有較好的績效。而在附加測試部分，本研究將分類最好的兩個類別選用去建立C5.0決策樹模型並從中找出影響盈餘管理的相關規則。
Many studies focus on whether companies faithfully present their economics sub-stance in the financial report. Earning management is one of those topics that manage-ment manipulates earnings to serve certain purpose.
In recent researches, data mining has been applied to business activities exten-sively. In the present study, the three steps model has been built to detect earnings management levels. First, applied step wise regression and Random Forest algorithm to the sample selection. Second, structure data mining models by using MLP、SVM and LS-SVM techniques. At the last step, the Decision Tree has been used to find out some rules related to earning management.
The sources are from Taiwan Economic Journal database, ranging from 2008 to 2012 and containing electronics companies listed on Taiwan Stock Exchange. 19 varia-bles including firms performance, cash flow from operating, and other financial or non-financial index are used in the model.
The result indicates that LS-SVM has better performance than MLP and SVM to detect earning management. Moreover, the best two classifiers have been chosen to build the C5.0 Decision Tree model, and then discovered the rules behind the earning management in the third step model.
