本研究運用資料探勘方法中之逐步迴歸、決策樹CART、決策樹QUEST、決策樹C5.0、類神經網路進行兩階段之建模分析。變數選取包括財務及非財務變數，研究對象為2006年~2015年發生財務報表舞弊之公司，以同年度、同產業且資產總額相當為標竿，以1:3進行配對。本研究共分為二個階段:第一階段分別以逐步迴歸、決策樹CART與類神經網路進行變數篩選，而在第二階段以決策樹CART、決策樹QUEST、決策樹C5.0及類神經網路來建立模型。實證結果顯示，模式一、模式二與模式三皆以決策樹C5.0為最佳結果，其中以決策樹CART搭配決策樹C5.0(模式二)準確率82.78為最高，應可有效地偵測財務報表舞弊發生，並降低審計人員之查核風險。
The Stepwise Regression, Decision Tree, and Neural Networks are used in this study, for a two-stage model. The period is from 2006 to 2015. In the first stage, Stepwise Regression, Decision Tree CART and Neural Networks are applied for screening variables. In the second stage, Decision Tree CART, Decision Tree QUEST, Decision tree C5.0 and Neural Networks are applied to establish the prediction model for financial statement fund. The results show that Decision Tree CART (the first stage) and Decision Tree C5.0 (the second stage) has the best performance 82.78%. This kind of method proves to be more effective in predicting the financial statement fund of companies.
