Unsupervised Identification of Corporate Tax-Risk Profiles Using Multilayer Network and Financial Characteristics
This study proposes an unsupervised framework that integrates multilayer network characteristics with financial deviations to identify heterogeneous corporate tax-risk profiles without relying on predefined criteria or labelled data.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Tax authorities around the world face a persistent challenge: companies do not operate in isolation. They form complex webs of ownership, management, and family ties, often spreading across borders to create structures that are difficult to untangle. When these groups of companies shift money, profits, or assets in unusual ways, it can signal aggressive tax planning or avoidance. Traditionally, tax officials have relied on expert judgment and fixed rules to spot these red flags, looking at individual company accounts or applying standard checklists. However, as the number of interconnected businesses grows and their financial data becomes more varied, these traditional methods struggle to keep up. They often miss the bigger picture of how a whole group of companies behaves together, rather than just how one company looks on its own. To solve this, researchers are turning to a different approach: using computers to find hidden patterns in massive amounts of data without needing to know in advance what a "risky" company looks like. This method, known as unsupervised learning, allows the data to reveal its own strange or unusual shapes, helping officials identify groups that stand out from the crowd based on how they are built and how their money moves.
A team of researchers from the Kaunas University of Technology in Lithuania has developed a new way to map these corporate groups to better understand their tax risks. Instead of looking at a single company, they treated entire groups of related businesses as a single unit, analyzing both their financial numbers and their structural connections. The researchers started with a vast dataset provided by the Lithuanian State Tax Inspectorate, which contained information on over 113,000 legal entities and more than 261,000 people. These entities were linked through three specific types of relationships: who owns shares in whom, who manages or controls whom, and family ties. The team built a digital model of these connections, keeping each type of relationship separate to preserve the unique information it held, much like keeping different layers of a map distinct rather than smearing them all into one. From this complex web, they extracted clear, readable descriptions of how each group was organized, counting things like how many companies were involved, how many foreign entities were part of the group, and whether the ownership loops back on itself in circles.
To understand the financial side, the researchers looked at how money was distributed within each group. They calculated how much each company contributed to the group's total assets, revenue, profits, and debts. In a typical group, these contributions might be somewhat balanced, but in some groups, one company might hold a massive share of the profit while another holds almost all the debt. The researchers measured these imbalances to create a profile of financial behavior for each group. They then combined these structural descriptions with the financial profiles and fed them into a specialized computer algorithm called the Reduced Clustering Method. This algorithm is designed to find natural groupings in data without being told what to look for. It scans the information and says, "These groups look similar to each other, and different from those over there," effectively sorting thousands of corporate groups into distinct categories based on their unique mix of structure and finance.
The study focused on 722 corporate groups that had enough complete financial data to be analyzed, representing a total of 7,050 legal entities. When the researchers let the computer sort these groups based only on their financial numbers, it found eight distinct profiles. Most groups fell into a large, common category where financial roles were relatively balanced. However, a few smaller groups stood out with extreme characteristics. One tiny group, consisting of only three companies, showed massive deviations in its operating results, while another small group of seven companies had unusually large swings in profits. These were not flagged as guilty of anything, but they were identified as statistically unusual, suggesting that something about their financial arrangement was rare compared to the norm.
When the researchers analyzed the groups based only on their structural connections, the computer found twelve different profiles. Some groups were defined by having many foreign owners, while others were characterized by complex chains of control or circular ownership structures where Company A owns Company B, which owns Company C, which in turn owns Company A. These circular structures can sometimes make it hard to see who is truly in charge. The analysis showed that groups of similar size could have very different internal organizations; some were simple and direct, while others were intricate networks involving many layers of ownership and foreign participation.
The most revealing part of the study came when the researchers combined both the financial and structural information into a single analysis. This approach created seventeen distinct profiles, offering a much more detailed view than looking at either factor alone. The combined analysis showed that financial oddities and structural complexity do not always happen together. Some groups were financially strange but structurally simple, while others had complex webs of ownership but very standard financial patterns. Most importantly, the study identified specific groups that possessed both unusual financial behaviors and complex structural features. For instance, the researchers found groups that had significant foreign participation, circular ownership structures, and extreme imbalances in how profits or assets were distributed among their members. These specific combinations are what tax officials are most interested in, as they represent the kind of hidden complexity that often accompanies aggressive tax planning.
The researchers emphasize that this method does not prove that any specific group is evading taxes. The algorithm simply identifies groups that look different from the average, acting as a tool to help human experts decide which cases deserve a closer look. By sorting the vast sea of corporate data into understandable profiles, the system allows tax authorities to move away from guessing and toward a more targeted investigation strategy. The study demonstrates that by looking at how companies are connected and how their money moves together, it is possible to uncover hidden patterns that would remain invisible if the companies were examined in isolation. While the current results are based on data from Lithuania and are limited to groups with complete financial records, the approach offers a promising path for modernizing tax risk assessment. It suggests that the future of tax compliance may lie not in stricter rules, but in smarter ways of seeing the connections that bind the corporate world together.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.