本研究應用資料探勘方法中的決策樹QUEST、CHAID、C5.0及類神經網路(neural network)，作兩階段的篩選建模分析，並採用財務及非財務變數，敝作為審計人員偵測財務報表舞弊之工具。研究對象為2007年~2013年之32家發生財務報表舞弊公司及96家非財務報表舞弊公司以資產總額1:3作樣本配對，本研究共分為二個階段:第一階段分別以決策樹QUEST、決策樹CHAID及類神經網路(neural network)作變數篩選。而在第二階段以決策樹C5.0及類神經網路(neural network)做建立模型。本篇研究結果顯示，第一階段應用資料探勘之三種方法作變數篩選再導入第二階段建立模型，在第二階段建模中皆以決策樹C5.0的分類表現為最佳，可以有效地偵測財務報表舞弊發生，藉此降低審計人員之查核風險。
This study applies decision tree (QUEST, CHAID, and C5.0) and neural network method in data mining to a two-stage screening and modeling analysis, and uses financial and non-financial variables, as a tool for auditors to detect financial statements fraud. This study took total assets of 32 companies who have involved in financial statements fraud, as well as 96 companies who have not during 2007-2013 as paired samples (1:3). The study is divided into two stages: the first stage uses QUEST-based decision tree, CHAID-based decision tree and neural network respectively for variables screening. In the second stage the study uses C5.0-based decision tree and neural network for modeling. The results show that regardless of which one of the three methods in data mining is used for variables screening before introducing their results into the second stage, the C5.0-based decision tree method has a best classification performance in the second stage modeling. It can effectively detect the occurrence of financial statement frauds, thereby reducing audit risks for auditors.
