VAT tax gap prediction: a 2-steps Gradient Boosting approach
It provides a robust method for tax authorities to improve tax gap estimation, though it is incremental as it builds on existing bottom-up approaches with a novel correction for bias.
This paper tackles the problem of estimating the VAT tax gap for individual firms in Italy by addressing selection bias in audit data, using a 2-steps Gradient Boosting model that significantly boosts prediction performance compared to classical parametric approaches.
Tax evasion is the illegal evasion of taxes by individuals, corporations, and trusts. The revenue loss from tax avoidance can undermine the effectiveness and equity of the government policies. A standard measure of tax evasion is the tax gap, that can be estimated as the difference between the total amounts of tax theoretically collectable and the total amounts of tax actually collected in a given period. This paper presents an original contribution to bottom-up approach, based on results from fiscal audits, through the use of Machine Learning. The major disadvantage of bottom-up approaches is represented by selection bias when audited taxpayers are not randomly selected, as in the case of audits performed by the Italian Revenue Agency. Our proposal, based on a 2-steps Gradient Boosting model, produces a robust tax gap estimate and, embeds a solution to correct for the selection bias which do not require any assumptions on the underlying data distribution. The 2-steps Gradient Boosting approach is used to estimate the Italian Value-added tax (VAT) gap on individual firms on the basis of fiscal and administrative data income tax returns gathered from Tax Administration Data Base, for the fiscal year 2011. The proposed method significantly boost the performance in predicting with respect to the classical parametric approaches.