Identifying business misreporting in VAT using network analysis

Abstract

Efficient detection of incorrectly filed tax returns is one of the main tasks of tax agencies. Value added tax (VAT) legislation requires buyers and sellers to communicate any exchanges that exceed a certain amount. Both statements should coincide, but sometimes the seller/buyer and its counterpart declare different amounts. This paper presents a method to detect those businesses that are more prone to misreport in their VAT declaration. Using the information of such declarations for a region in Spain during year 2002, we generated a transaction network formed by the tax declarations of buyers and sellers. Four types of error were assigned to each business in the network, defined from the mismatch between the amount declared by the firm in question and its counterpart. We applied a random forest algorithm to detect which firm-related and which network-related characteristics influence each error type. The results show the importance of relational factors among businesses in determining the probability of presenting VAT declaration errors. This information can be used to promote more efficient inspections.