Greece’s Independent Authority for Public Revenue (AADE) identified the network of ‘ghost companies’ in Thessaloniki using a risk analysis system based on the Jaccard similarity index.
The authority uses the index in a model designed to detect links between businesses and dismantle networks of sham companies, known as ‘ghost companies’, and the frontmen behind them. The algorithm maps hidden links between different tax identification numbers as companies are set up and shut down in succession to avoid paying taxes and social security contributions.
How the algorithm works in practice
The algorithm maps recurring links between businesses, individuals, infrastructure and transactions across multiple tax identification numbers.
Its aim is to identify businesses that are registered as separate legal entities but share characteristics that suggest they may belong to the same network.
AADE’s risk analysis system treats each business as a set of data and compares its characteristics with those of other businesses. A Jaccard score closer to 1 indicates greater overlap, which may point to companies belonging to the same network.
The system looks at the following factors:
- Individuals: The algorithm checks how many partners, shareholders, managers or legal representatives appear across different companies.
- Infrastructure and contact details: It identifies shared registered offices, telephone numbers, email addresses, accountants or bank accounts.
- Commercial transactions: It measures overlap in customers and suppliers using data from the myDATA electronic bookkeeping platform and VAT returns. If a new company serves exactly the same customers as one that has just closed down leaving debts behind, the similarity score rises sharply.
- Network mapping: The tool led to the dismantling of a major network involving 380 businesses and 205 individuals that had accumulated €43 million in debts to the state.
How the index was used in the Thessaloniki case
AADE sources explained that, in this analysis, the index compares two individuals at a time. If each appears in 15 companies and they share 10, there are 20 distinct companies in total (15 + 15 − 10), with each shared company counted once.
The Jaccard index is calculated by dividing the number of shared companies by the total number of distinct companies. In this example, 10 divided by 20 gives 0.5, or 50%.
A Jaccard score of 50% means both individuals appear in half of the 20 companies.
Pairs with high similarity scores form ‘links’ that can be joined wherever they share an individual. These links form ‘chains’, revealing networks of business relationships and possible co-operation. The authority identifies the network, and investigators determine whether illegal activity has taken place within it.
In the Thessaloniki case, AADE’s DEOS enforcement unit used the Jaccard index to identify hundreds of links between tax identification numbers. Each pair shared at least two companies. Joining those links produced chains and groups:
- Above 5%: 719 links across 369 tax identification numbers, forming 29 groups. The largest contained 262 tax identification numbers.
- Above 10%: 504 links across 330 tax identification numbers, forming 60 groups. The largest contained 34 tax identification numbers.
- Above 20%: 358 links across 270 tax identification numbers, forming 77 groups. The largest contained 13 tax identification numbers.
In practice, if A is linked to B and B is linked to C, they form a chain even if A and C have no direct connection. Raising the threshold keeps only the links with higher similarity scores, breaking larger chains into smaller groups.
This allows auditors to focus on specific links and the companies the individuals share. Further investigation is needed to establish whether any group was involved in illegal activity.
What has changed in tax audits
The introduction of the Jaccard index has changed how the tax authority conducts audits:
- Identifying suspicious tax identification numbers: AADE teams can quickly spot suspicious patterns.
- Automated selection for targeted audits: Algorithms automatically select high-risk cases, directing auditors to the businesses and individuals at the centre of suspected networks.
- Linking databases: The index is used alongside data from Greece’s General Commercial Registry, banks, land registry records and electronic payments to build a single digital risk profile.
AADE has also released a video showing how the investigation unfolded. The authority uses the same method in similar cases. The squares outlined in red at the end of the video represent cases classified as high risk.
Ask me anything
Explore related questions
