The IRS (U.S. Internal Revenue Service) has been using computers to choose tax returns for audit since 1962. Early on, the selection was rule-based, but the IRS turned to statistical modeling in 1969, using the oldest predictive analytics model in the toolbox – discriminant analysis. Discriminant analysis, a linear classification technique, was first proposed by Ronald Fisher in 1936. Computer scientists think of discriminant analysis as quaint and old-fashioned, that is, if they think of it at all. However, it it has the merit of being computationally economical (hence fast), and works well with smaller datasets.
According to statistician Amir Aczel, the IRS was still using discriminant analysis in 1995. He published a book that claimed to reverse engineer the IRS discriminant function, and provide definitive guidelines about ratios (e.g. deductions to income) that would guarantee or avoid an audit.
The use of more advanced machine learning methods got a boost in 2011 with the creation of the Office of Compliance Analytics. Now, the selection of returns for audit relies primarily on predictive models whose “rules” are internally-generated and may not even be visible to the IRS analysts.
On the data side, the IRS now has access to much more data than was used in the simple days of discriminant analysis, which relied solely on data in returns and forms filed with the IRS. Now the IRS monitors Facebook, Twitter and Instagram for patterns that might alert them to tax fraud by individual taxpayers. Washington State professors Kimberly A. Houser and Debra Sanders report in their paper, The Use of Big Data Analytics by the IRS: Efficient Solutions or the End of Privacy as We Know It?, that the IRS now tracks over 1 million attributes, or predictor variables, on individual taxpayers, drawn from numerous sources, including, in addition to the social media sources noted above, individual email traffic.
One new area where the IRS has been particularly vigilant is crypto-currencies. Last year, the IRS reported on its successful proceedings against Coinbase, which is the largest domestic crypto-currency exchange. While fewer than 1000 taxpayers reported crypto-currency gains over the 2013-2015 period, the IRS thinks that millions should have done so. The agency added the Coinbase taxpayer data, which it seized, to its other big data resources.
The efficacy of big data analytics for predicting tax fraud can be seen in the following statistic: while its enforcement budget has continuously declined over time, last year the IRS reported that it caught more than 400% more tax fraud cases and recovered 1000% more in taxes than in the prior year.
REFERENCES:
- http://www.jetlaw.org/wp-content/uploads/2017/04/Houser-Sanders_Final.pdf
- http://cavqm.blogspot.com/2011/07/reverse-engineering-irs-dif-score.html
- https://www.smartdatacollective.com/can-predictive-analytics-prevent-tax-evasion/
- https://freemanlaw.com/the-irs-and-big-data-the-future-of-fighting-tax-fraud/