應用類神經網路於營業稅逃漏稅預測模式之建構

Abstract

營業稅是國家稅收的主要來源之一，目前營業稅之查核方式，採用選案查核，而非普查，造成營業人常利用各種方法逃漏稅。以民國八十九年為例，營業稅收入為2212億元，而違章金額高逹112.7億元，占所有稅目違章金額之冠，可見營業稅逃漏情形嚴重。本研究利用營業稅資料，結合資料探勘之類神經網路技術，探討影響預測逃漏稅之重要因素，以建構營業稅逃漏稅預測模式。本研究之資料來源為民國89年第2、3期及民國90年第1期之製造業及服務業營業稅申報資料；經分析過濾後，選擇服務業及製造業各6000筆資料作為樣本，共12000筆資料，進行實證。研究結果顯示，以應納稅額、銷售額、進項稅額、銷項稅額、及費用率等資料建構之預測模式，分類正確率達90.39%。其中較重要之影響因素為「本期應納稅額」與「銷售額總計」的比例、「未扣抵前之稅額」與「銷售額總計」的比例、及「未扣抵前之稅額」與「可供扣抵之稅額」的比率。

The business tax is one of the main tax revenue of our country. The current method of tax audition is by selection not by universality. Therefore, there are many business used different ways to evade paying tax. The total amount of business tax is about 221.2 billion NT dollars during 2000, whereas the amount of tax evasion is about 11.27 billion NT dollars. It is obvious that the tax evasion behavior is serious. This research using business tax data associated with Neural Network technique to build a tax evasion prediction model. The samples of business tax data we used are 3 periods of service and manufacture industry from 2000 till 2001. We selected 6000 records from each industry. The total samples are 12,000 records. The results showed that the optimal accuracy rate of classification is about 90.39%. The significant factors are the ratio of ＂tax-duty/total sales＂, ＂tax before deducted/ total sales＂, and ＂tax before deducted/ tax-deductible＂.