| Summary: | 碩士 === 國立中正大學 === 會計所 === 94 === The Financial Data Center of the Ministry of Finance selected audit cases of the profit-seeking enterprise income tax by computers and hand for further investigation by the National Tax Administrations. Due to the scarcity of human resources to fully audit all the cases, different National Tax Administrations units would audit the assigned cases in different ways. It might result in unfairness and become difficult to encourage the profit-seeking enterprises to honestly file their tax reports. Therefore, we proposed an objective and systematic model to assist tax auditors to spot the suspects of tax evasion. It helped avoid wasting time on invaluable cases and ensure fairness.
The research used the profit-seeking enterprise income tax data of the National Tax Administration of Southern Taiwan Province between 1995 and 2002. The data were preprocessed before being applied to IBM Intelligent Miner for Data V8.1 to create the decision tree and neural net algorithms that identified tax evasive declarers. The effectiveness of the models was evaluated by whether the savings on the costs of checking entire population and imposing the evasion tax exceeded the amounts of tax losses from the tax evasion declarers who were not identified by the models.
The findings are as follows.
1.The neural net outperformed decision tree in classifying tax evasion. The amounts of tax losses from the tax evasion declarers who were not identified by the models were lower than the savings on the costs of checking entire population and imposing the evasion tax. Therefore, both the decision tree models and neural net algorithms effectively detected tax evasion.
2.The five most important independent variables in discriminating tax evasion suspects identified by the decision tree models and neural net algorithms were gross profit margin, net profit margin, business net profit margin, salary expenditure rate, and commission expenditure rate.
|
|---|
