| 摘要: | 企業普遍使用財務報表向外界傳遞公司內部信息，不良企業得藉由竄改財務報表，來掩飾其不佳的財務狀況以獲得更多投資。類似重大舞弊在台灣有國票、博達，國外較出名的安隆、世界通訊…等。近年來，許多學者用資料探勘的方法對財務舞弊做偵測，像是Kotsiantis (2006) 實證分析採用決策樹 C4.5 及類神經網路、Song, Hu (2014) 實驗分析使用的是 Logistic 迴歸、類神經網路、支援向量機及 C5.0 決策樹。這些研究都有相當不錯的準確率，但相關文獻還需要更加完整。本研究對象為台灣上市、上櫃之全部產業，並依據投資人保護中心公有發生財務報導不實公司作為研究樣本，資料選取來自台灣經濟新報(Taiwan Economic Journal, TEJ)，以資產總額相當為標竿使用舞弊與非舞弊比例1:3配對，選取期間為2002年至2017年，總計15年(財團法人證券投資人及期貨交易人保護中心，2018)。第一階段採用類神經網路，第二階段之建立模型應用了決策樹C5.0、決策樹CHAID以及支援向量機(SVM)來建構預測模型，配對後模型準確率為87.194％。
Enterprises often use financial statements to communicate internal company information to the outside world. Bad companies can tamper with their financial statements to cover up their poor financial situation to obtain more investment. Similar to major fraud in Taiwan, there is a national ticket, Boda, the more famous Anlong abroad, the world communication...etc. In recent years, many scholars have used data exploration methods to detect financial fraud. For example, Kotsiantis (2006) empirical analysis uses decision tree C4.5 and neural network. Song, Hu (2014) experimental analysis uses logistic regression. Neural networks, support vector machines, and C5.0 decision trees all have good accuracy, but the literature needs to be more complete. This research object is listed in Taiwan, the whole industry of the upper cabinet, and based on the financial report of the investor protection center, the company has taken the financial report as a research sample. The data is selected from the Taiwan Economic Journal (TEJ), with the total assets as the standard. Use fraud and non-fraud ratio 1:3 pairing, the period of selection is from 2002 to 2017, a total of 15 years (2010 Securities Investment Investors and Futures Traders Protection Center, 201 8). The first stage adopts the neural network of the first stage, and the second stage establishes the model using decision tree C5.0, decision tree CHAID and support vector machine (SVM) to construct the prediction model. The accuracy of the model after pairing is 87.194%. |
