In a recent academic study, machine auditors were found to be highly effective in detecting tax evasion as well as recouping delinquent taxes.
The study, recently published by the
In order to build their model, researchers used files from 2007, 2008 and 2009, for which they observed the complete fiscal cycle — that is, the following five years in which they might be audited. By doing so, they received a representative sample of the composition of the files that are audited over time. The model had two goals: detecting evasion and recovering the detected amount of evaded tax. Specifically, for each file the two main variables of interest were tax evasion, defined as the difference between the tax amount assessed during an audit and the tax paid; and a proxy for the actual tax evasion recovered by the IRA equal to tax evasion when the taxpayer pays back within due time, and zero when the taxpayer is insolvent.
Key to using this model was discarding the 10% worst-performing audits from the data set (those that cost more money to perform than was recovered) and replacing them with files associated with later, more successful audits.
They found that, at a 10% discard rate, selecting the replacements using machine learning from the sample pool yields an improvement of 38% of tax evasion detection, and a 29% improvement of tax revenue collected. An even larger pool, it was found, could have even bigger effects: 83% improved tax evasion detection and 65% more revenue recovered.
"This evidence indicates that there are considerable gains from replacement even when keeping the composition of the audited files by key characteristics broadly unchanged," said the paper.
