營利事業所得稅逃漏稅預測模式之比較研究

Abstract

由於營利事業所得稅是國家主要的稅源之一，且該稅之查核係採選案查核而非普查，在查審人力無法增加的限制下，如何從事該稅逃漏之有效查緝值得深入探究！本研究透過22項財務比率差異與比值，組合類神經網路、因素分析及羅吉斯迴歸等三種計量方法，提出五種預測模式，並進行預測效果之實證比較。我們以財政部台灣省北區國稅局轄區內的電子資訊產業及其八十六及八十七年稅務資料為例，利用複式抽樣法及重複測度分析，透過型一歸類錯誤率 (係反映未能偵測實際逃漏稅之風險)、型二歸類錯誤率 (係反映查緝實際未逃漏稅之風險) 及整體歸類錯誤率 (係反映型一及型二之綜合風險) 三項指標評估五種預測模式之效果。研究結果顯示，直接以 22 項財務變數納入類神經網路模式中之預測效果最佳，但卻無法獲悉具顯著預測能力之財務變數。除預測效果外，協助查審人員設定查緝方向亦是訴求重點，基於此一觀點，並根據研究結果，我們提供了一些實務上的應用建議。

Business income tax evasion has been a very popular and serious problem. A direct way to reduce tax evasion is to enhance auditing. Owing to limited resources, it seems hard to increase auditors. Therefore, how to effectively predict if a tax evasion will occur becomes important. The purpose of this study is to conduct comparisons of performance among different prediction approaches, and then find out the most effective one. Five prediction approaches by means of neural networks, factor analysis and logistic regression are used. They are based on 22 financial variables. The overall prediction error rate, Type I prediction error rate, and Type II prediction error rate serve as prediction performance criteria. Using Bootstrapping sampling and repeated measures analysis, the approach of neural network based on 22 financial variables was found to be the most effective. However, variables possessing explanatory power cannot be identified by the approach. Factor analysis on 22 financial variables, followed by stepwise logistic regression, is suggested to help with critical variable selection.