Adaptive algorithmic technology could help optimize audit sampling to better identify tax underreporting and close revenue gaps, according to a recent study by academics and government officials.
The use of sequential decision-making, or machine learning, integrated with the National Research Program model used in randomized IRS audits is “fertile ground” for exploration, a team of Stanford University and IRS researchers wrote in a paper published in April.
So-called multi-armed bandit algorithms are geared toward a specified “reward”—in this case, identification of underreporting that would close the tax gap, or the difference between what taxpayers owe and what’s actually collected. Such algorithms aren’t prevalent in the public sector, but the paper’s authors said machine learning could improve the efficiency and quality of government services if carefully deployed.
Peter Henderson of Stanford University, the paper’s primary author, presented the research at a June 16 webinar co-hosted by the IRS and the Urban-Brookings Tax Policy Center. In a panel discussion, he said these technologies have been used by private companies. Netflix, for example, uses an algorithm that analyzes user engagement data to better predict what people are more likely to watch next.
Modern IRS audit procedures have an “optimization problem,” Henderson said. The agency can struggle to pin down where misreporting occurs while also “actively finding the largest amounts of misreporting.” The paper used returns from tax years 2006 through 2014 to conclude that a “unified optimize-and-estimate” program under a bandit-like framework could efficiently maximize revenue while maintaining accurate estimates of the tax gap.
Unlike other machine learning research, this study maintains that algorithms that can adapt to data should still be paired with the NRP random model. Former National Taxpayer Advocate Nina Olson submitted a question at the panel to Henderson asking why randomized audits need to be relied upon if machine learning can supposedly better select returns to audit that are more likely to reveal underreporting.
In response, Henderson said randomness alleviates a risk of bias that machine learning algorithms can only reduce but not fully eliminate. Without random sampling, a model can get stuck in “suboptimal feedback loops” in which it oversamples a particular area, possibly leading to significant problems, he explained.
Alan Plumley of the IRS Research, Applied Analytics, and Statistics Division said there are two competing interests at work in the National Research Program model.
On one hand, the its sample size has declined 43% over recent years as the IRS conducted fewer total audits due to budget and resource constraints. This undermines the quality of estimates based on the model, Plumley said, suggesting that the remedy would be more-random audits.
Conversely, random audits have become a larger percentage of overall audits the IRS has conducted in recent years. This is detrimental to the goals of nonrandom operational audits aimed at recapturing revenue, and implies that the model’s sample size should be decreased, Plumley said.
The Stanford/IRS research has the potential to lead to a better balancing of these competing revenue and measurement objectives, according to Plumley. However, his analysis of the research shows that unbiased compliance estimates at the line-item level are more important than those of the gross underreporting tax gap. In addition, the paper doesn’t address the cost of implementing machine learning into audit procedures.
“Total reward is budget-constrained, unlike one-armed bandits,” Plumley said. He also argued that audits widely vary in cost and that audit selection should be based on a revenue/cost ratio. It may be cheaper to increase the number of random audits than it would to “fix the operational lineup and data,” he added.
Henderson countered that while it’s “definitely true” that cost is an important factor, a budget-first mindset can have consequences. For example, the IRS might stop auditing parts of the income base because of cost, despite misreporting, he suggested.
The IRS must “be careful” and consider the full picture, Henderson said.
Get all the latest tax, accounting, audit, and corporate finance news with Checkpoint Edge. Sign up for a free 7-day trial today.