Can Machine Learning Improve Correspondence Audit Case Selection? Considerations for Algorithm Selection, Validation, and Experimentation

• Activity Number: 132 - SLDS CSpeed 1
• Type: Contributed
• Date/Time: Monday, August 9, 2021 : 1:30 PM to 3:20 PM
• Sponsor: Section on Statistical Learning and Data Science
• Author(s): Lucia Lykke* and Ben Howard and David Pinski and Alan Plumley
• Companies: The MITRE Corporation and The MITRE Corporation and The MITRE Corporation and Internal Revenue Services
• Keywords: Tax enforcement; compliance; machine learning

Abstract

Using machine learning (ML) techniques, the Internal Revenue Service can improve operational processes by exploiting large volumes of taxpayer data at its disposal. We applied ML to correspondence audit case selection with the objectives of improving revenue and reducing taxpayer burden. We demonstrated how ML methods that use many features of taxpayer reporting may improve audit results through focusing on two model outcomes: revenue and no-changes (audits resulting in no adjustment). The study has two parts. First, we shared results from preliminary proof of concept ML experiments for correspondence audit selection for 2013-2016 returns, conducted in cooperation with the audit program for two types of audits. Results showed increased revenue for one type of audit compared to baseline ranking methods and decreased no-changes for another audit type. Second, we refined the exploratory models based on insights from the experiments. We compared several ML techniques that handle regression and classification differently. We showed that strictly minimizing the audit no-change rate may come at the cost of forgoing returns resulting in higher revenue, and vice versa.