The following essay examines the Panama Papers from a sociological perspective, using a method based on analytic induction through graph mining. By way of an introduction, it provides a general overview of the chosen approach. A series of case studies concerning tax evasion practices as detailed in the Panama Papers help to elucidate the kind of query processing that might accompany the practical implementation of the proposed method, before the final section gives a brief summary and touches upon a methodological issue that has yet to be resolved.
Identifying relevant patterns in a large graph of open data is a complex,
exploratory process – a statement that finds ample support in the following
essay, which examines the Panama Papers from a sociological perspective,
employing a method based on analytic induction through graph mining.
From a user’s perspective, data visualization is a crucial step on the path to
data‐driven discoveries
In an eofrt to better understand how the process of analytic induction
through graph mining might function in practice, this essay explores open
data knowledge graphs (KGs) with the help of DERIVO’s SemSpect tool
In the following sections of this paper, I will provide a general overview of analytic induction (Section 2); discuss the kind of query processing that was used to examine a series of case studies on tax evasion practices, as revealed in the Panama Papers (Section 3); present a selection of data-driven discoveries, which will illustrate the benefits of this kind of query processing (Section 4); and present an outlook on further research based on analytic induction (Section 5). 2
Analytic induction
Building a singular complexion of properties Looking for interesting relations
Deep understanding of the data while building hypotheses Interesting patterns do not have lots of occurrences but are frequent enough to capture general features
The method of analytic induction highlights a number of complex
challenges associated with graph mining and semantic analysis (Table 1). First
and foremost, it underscores the importance of refining and developing the
categories that are initially used to define a particular social phenomenon;
computationally speaking, this corresponds to ontology refinement and
iterative semantic processing. Even more significant, perhaps, is the fact that the
primary goal of analytic induction is to identify cases
Using the method outlined in the previous section, we examined the open data package commonly referred to as the Panama Papers with a view towards gathering any relevant information that could help to shed light on the tax evasion techniques that were being used.
Typically, the tax optimization practices of large companies are overseen
by external treasury management providers and include methods such as
over-invoicing; transferring expenses to subsidiaries located in tax havens;
’forgetting’ sub-subsidiaries in the accounting consolidation; organizing
white sales; and laundering money on a large scale through the purchase of
raw materials, and on a small scale through cheques or prepaid card systems.
=>
=>
=>
All of these practices function as “black holes of power”
In the absence of open transaction data, the information provided by the
Panama Papers is predominantly topological. Suzanne
Historically, law firms and corporate service providers such as Mossack Fonseca, who were heavily implicated in the Panama Papers scandal, have used certain strategies to help companies play with the spatiality of jurisdictions in their relations with supervisory institutions. What these strategies have in common is that they are not codified, but rather based on tacit knowledge. We have therefore examined graph data from the Panama Papers with a view towards patterns that are characteristic of such implicit practices.
The online databases in question match the definition of what is
commonly called Open Data: “A piece of data is open if anyone is free to use,
reuse, and redistribute it – subject only, at most, to the requirement to
attribute and/or share-alike.” The progressive aggregation of such databases
means that they become increasingly linked
The Evasion Professionals database is an aggregate of the Ofshore Leaks database. The architecture of the database consists of oficers (operators identified as set-up operators), intermediaries (intermediaries working on the files), entities (natural or legal persons on whose behalf the file is processed), and addresses (addresses linked to the three types of predefined entities); companies are classified according to operational criteria (active, inactive, dissolved, relocated, redomiciled, etc.). The content available online does not, however, give access to all of the extracted data: in order to protect themselves from cumbersome and costly legal proceedings, the International Consortium of Investigative Journalists (ICIJ) and the participating hackers have left out information concerning the amounts and dates of transactions.
The absence of dates means that fine temporal processing is not available,
whether it be genealogical, archaeological, or sequential. As a result, the
attempt to represent the complex social temporalities involved in the various
transactions recorded in the data produces a synchronic fiction. In fact,
computation without dates makes it almost impossible to carry out a pragmatic
analysis of the operations of record capture, circulation, and manipulation,
or any form of matching of editing sequences forming the ‘careers’ of
clients
In the case of the Panama Papers, the ordering of the data corresponds neither to the logic of entirely raw and unstructured data, nor to semantically consistent object classes or ontologies.2 It does, however, appear semistructured for the purposes of statistical analysis, whether it be simple descriptive statistics or network statistics, and the ICIJ website provides a graph-based network analysis software that appears to be intended for precisely that purpose: the examples made available to users online show networks of co-afiliation located at the same tax address, and the various afiliations of operators identified as fraudsters.
2In computing, ontologies are structured sets of terms and concepts that are used to define the meaning of a given information field, either via the metadata of a namespace or the elements defined by a stabilized knowledge domain.
In combination, the openness of the available data, its sheer volume, and the kind of investigative tools that are provided on the ICIJ’s website exert considerable influence on how the mechanisms of tax evasion can be examined. 3.2
Sociologists investigating the phenomenon of tax evasion face a number of methodological challenges. When conducting exploratory research of open data, two types of problems immediately become apparent: the first concerns the analytical limitations of the tools that are made available to the researcher, as in the case of the ones provided on the ICIJ website; the second has to do with the tools used for alternative research strategies.
The ICIJ gives interested users the option of downloading network processing software with which they can generate graphs comparable to those created by information services companies to visualize the afiliations of company directors or other persons of interest. However, the ICIJ’s choice to focus on what is commonly referred to as self-centered networks or afiliation networks has some rather unfortunate consequences. Generally speaking, working with self-centered networks involves a simple query based on filters and dictionaries of proper names (celebrities, companies, etc.), which makes it possible to quickly retrieve information concerning the participation of a given person or company. In the case of Mossack Fonseca, the immediate targets of ICIJ’s eofrts were the British and Icelandic prime ministers, David Cameron and Sigmundur Davíð Gunnlaugsson, both of whom resigned in the wake of the Panama Papers scandal.3
By conjuring up visible networks, the ICIJ succeeded in arousing media interest in what are ultimately invisible practices. However, by choosing to focus on the relationship between a particular personality and the intermediaries of an optimization firm like Mossack Fonseca, the ICIJ participated in a logic of shaming the latter’s clients, as opposed to conducting an indepth analysis of the firm’s practices. Self-centered or afiliation networks are simply not conducive to exposing the creative bookkeeping techniques of such firms, which often duplicate, conceal, or anonymize company names – in the final analysis, assessing the phenomenon of tax evasion based on dyadic relationships, such as those between oficers and intermediaries, or entities and oficers and/or intermediaries, is an inadequate response to the challenges at hand.
3The news that Cameron held shares in an ofshore company together with his father broke in the middle of the Brexit campaign, while Gunnlaugsson’s involvement in ofshore dealings came to light shortly after he had signed a military agreement with the US. 3.3
One way to address the problems outlined above is to draw on software
solutions from the domain of visual analytics, which can be defined as a science
of analytical reasoning
By proceeding in an inductive way from basic data to the detailed analysis of complex files, our systematic exploration of the database enabled us to identify three recurring tax evasion techniques in the Panama Papers. 4.1
As discussed in the previous section, the specific nature of the data contained in the Panama Papers demands a suitable approach to processing it, and great care must be taken in regard to the inferences that are being drawn.
After conducting multiple tests, it became clear that the only reliable method for engaging with the data at hand – incomplete as it is – was to reconstruct the chains of tax packages starting with their associated addresses. The registered companies are letterbox companies that have been set up for the very purpose of bypassing legal restrictions, and they typically carry more or less creative names. But if one manages to move past this ruse, it becomes possible to reconstruct chains of letterboxes, to follow circuits of evasion, and to expose the accounting heterotopias that are made possible by the geographical location of the financial vehicle corporations (FVCs) involved in the scheme.
With this goal in mind, the query logic of searching for co-afiliations of companies at specific addresses appears far more promising than the analysis of egocentric networks. Indeed, the mass of data uploaded by the ICIJ is rendered practically useless if it is not reconfigured under an ontological database model which allows the user to define object classes and inferences according to comparable semantic properties or certain logical properties.
The need for such an ontological database model is easily demonstrated. For example, the ICIJ site provides a graphical representation of the relationships that can be reconstructed based on the identification of the various companies registered to the same hotel suite in the Seychelles (Figure 1).
Here we see how an unmodeled database can be used to trace the relationships between the hotel’s managers and the companies Green Apple System Limited/Ltd and GWT Systems Limited by relying on nothing more than an address (Suite 102, Aarti Chambers, Mont Fleuri, Victoria, Mahé, Seychelles).
As Table 2 shows, the details of this address are subject to considerable variation in the database, with changes falling under three broad categories: punctuation, capitalization, and (deliberate) spelling errors. Without a semantic approximation model, one would have to multiply the queries to search for other possible occurrences of the same suite under diefrent names. However, thanks to the modeling we developed in cooperation with DERIVO, we were able to identify not one but twelve occurrences of the same suite (Table 3). Moreover, we found other suites with multiple occurrences at the same residence, as well as in other locations throughout the Indian Ocean (Seychelles, Mauritius, etc.).
Since the data is not time-stamped, it is dificult to infer whether or not
the relationships between all of these companies were simultaneous – it
appears possible, and indeed likely, that we are dealing with a series of successive
domiciliations. But in any case, what becomes evident here is an important
mechanism of tax evasion, namely the use of concealment procedures within
protected databases in accordance with Hervé Falciani’s
Were we to use ICIJ’s approach to analyze the same data, we would find
that a total of eighty-one addresses remain after the subtraction of four
duplicates, all of which appear as separate entities. What becomes obvious here
is that while ICIJ managed to extract a large amount of information
concerning the mechanisms of tax evasion (and, as we have seen, generated
impressively large numbers in the process), the knowledge it produced is actually
much more superficial than it appears. In other words, it is quite clear that
the cognitive strategy employed to convert information extracted from the
data into knowledge
As the previous example has shown, the uncovering of tax avoidance structures entails identifying the location nodes of FVCs and using these nodes to reconstruct corporate networks based on criteria such as company owners, intermediaries, and assembly operators. Instead of relying on the soft
No. of No. of compan- doublets ies 67 3 1 Lettering of the address ware made available online by ICIJ with all of its limitations, I would argue that it is much more expedient to investigate evasion circuits on the basis of relational chains.
The analysis of networks in terms of relational chains, first proposed by
Stanley
Let us consider an example of such a chain: the employer knows A, who
knows B, who knows C, who knows the potential recruit; B passes the latter’s
name along to A together with his assessment of C’s qualification to give a
recommendation, and the likelihood of his giving a frank opinion; A in turn
passes this information on to the employer, along with his own assessment
of B. Of course, chains of this kind become more and more impractical the
longer they are – even the three-step recommendation chain from our
example has a slightly implausible ring to it. But Milgram’s work holds out
hope that short chains could in fact be the norm: his investigation showed
that randomly chosen pairs of Americans (one from Massachusetts and the
other from Nebraska) need a mere six to eight links on average in order to
connect them with one another
In the case of the Panama Papers, the relational chains are not always chains of operators in the strictest sense of the word, since it is not uncommon for the final recipient to be the initial customer at the end of a perfectly circular array of intermediaries. This poses several problems if we are to conduct a true network analysis.
The first obstacle is gaining access to the governance structures of companies residing in tax havens, as managers often enter account lines in shell companies that are nothing more than mailboxes hosted by intermediaries. This raises the question of which sources can be used to trace these financial lfows – after all, the concerned parties have no incentive whatsoever to cooperate. And what of the various other challenges associated with modeling these complex and sometimes flat-out paradoxical relationships? Can one be connected to a fictitious double of oneself via an intermediary one has never met? And if so, to what extent is that double actually fictitious? Are there no formal obligations to be fulfilled (signature, face-to-face procedures, etc.), no practical precautions to be taken to facilitate eventual liquidation?
For the purposes of our investigation, the lack of complete accounting data means that carefully planned qualitative explorations are necessary to understand the structure of the relational chains being examined. One particularly instructive case involves a number of companies linked to Portcullis, an assembly operator that is registered at 113 diefrent addresses in 9 countries (Table 4).
These addresses are in turn linked to 1,735 oficers, i.e. professional assembly operators (Table 6). The majority of these oficers are located in tax havens, but some are based in European countries that have not traditionally fulfilled such a role, including Italy. The two Italian oficers are registered at 18 diefrent addresses in their home country and are linked to an additional 22 oficers, one of whom is a legal entity called Sharecorp. When we proceed to examine the number of beneficiaries of this company, we discover no less than 1,610 further entities, distributed geographically over several tax havens (Table 5).
We can thus regress to infinity. What we have uncovered here, then, is an intricate partitioning system that organizes transnational address sharing in the form of concentric circles; an arrangement that is dificult if not impossible to penetrate for all those who do not have access to the database of participating law firms, most of which are based in the British Virgin Islands. Graph-based visualization, however, can help to facilitate the search for an original pattern. Beginning with what appears to be a kind of statistical anomaly, it becomes possible to identify forms of organization that would go unnoticed by classical detection algorithms – spotting their relevance is not so much a matter of statistical knowledge of graph structures, but a question of familiarity with the object of inquiry. 4.3
Our analysis revealed two characteristic traits of the assemblies in question: high-intensity backrest turnover and pronounced diversity of the carrying vehicles. The use of shipping companies and shipping addresses seems to be especially favored (Table 6).
The available data is testament to the massive scale on which asset management companies have been incorporated (72,720 in the Panama Papers and a further 4,260 in the Ofshore Leaks database). Figure 2 shows the distribution of these companies by country.
Of course, a chart based on the organizing principle of companies per country has significant limitations, as it does not reflect the volume of assets carried. Nevertheless, it does show that the extensive use of shipping companies is favored by Western companies and their traditional tax havens, whereas it is much less prevalent in Asia (with the notable exception of Hong Kong, which has traditionally served as the financial interface between major Western banks and Asian companies).
The 76,980 companies in question own a total of 35,345 entities, of which Country British Virgin Islands China Cook Islands Hong Kong Malaysia Samoa Seychelles Singapore Taiwan 31,549 were set up by Mossack Fonseca, and the rest by Portcullis. The status profile of these entities shown in Table 7 illustrates the intensity of turnover.
With companies registered as an international business company (IBC) in
a tax haven, the situation is straightforward: they are considered to be classic
instruments of tax evasion. Other vehicles, such as the porting of addresses
for consulting firms, blur the line between tax evasion and money laundering.
Indeed, the scheme of using a multitude of ‘shipping addresses’ is very much
in line with the findings of Fabian
To cite just one example, a relational chain based on the addresses connected to a specific porting address alerted us to complex financial dealings between Russia, Dubai, and the United Kingdom. After researching identities and companies, we found that one of the individuals (subsequently referred to as operator X) involved in the scheme was already being prosecuted in the United Kingdom and Moldova for money laundering on a scale of several billion euros. Country Australia Bahrain British Virgin Islands Canada China Cook Islands Djibouti France Germany Greece Guernsey Hong Kong India Indonesia Italy Japan Kazakhstan Kenya Malaysia Mauritius Myanmar Philippines Russia Samoa Seychelles Singapore South Africa South Korea Spain Sri Lanka Switzerland Taiwan Thailand Turkey United Arab Emirates United Kingdom United States 1 Status Defaulted Active Entity Dissolved Changed agent Struck / Defunct / Deregistered Inactivated Resignated as agent Dead Relocated in new jurisdiction In transition Discontinued Transferred Out Shelf company Bad debt account Liquidated Not to be renewed / In deregistration Shelf company not possible to sell In liquidation Unregistered Change in administration pending Redmoicited Trash company Profile Standard International Company Standard Company under IBC* Act Business Company Limited by Shares Nominee Only Entity Bahamas IBC* Turks Geographical location of the entities carried Entity in Switzerland Entity in Luxembourg Entity in British Virgin Islands Entity in Hong Kong Entity in USA Entity in Germany Entity in France o U i t e R d : n
a 3 , e i r a u b ig u F D
Yet the connected addresses revealed arrangements that went well beyond the indictments brought in this case: in the Mossack database, operator X appears as the manager of 103 companies located in ten diefrent countries: 82 in the British Virgin Islands, 70 in Russia, 9 in Cyprus, 7 in Samoa, 6 in Ukraine, 4 in the United Kingdom, 3 in the US, 2 in the Seychelles, 1 in Hong Kong, and 1 in the United Arab Emirates.
Operator X as an individual is domiciled in a series of addresses in Dubai that are artificially distinguished according to the method described above. Some of these addresses are shared by the anonymous directors of Cyprusdomiciled carrier companies. After a query of the operators sharing the address of the apartment in Dubai, it became apparent that one of them, a British national (Y), is one of several British shareholders in six companies domiciled in Russia and Cyprus, neither of which is owned or managed by X. However, when we explored X’s own activities in Cyprus, we found that X manages other companies in the country. In turn, these companies are all mediated by a Russian consultancy firm registered at an address in Cyprus – and one of the directors of this firm is Y, the person who shares X’s address in Dubai (Figure 3).
The chain thus operates as follows: the British partners in Russia, linked to the Dubai address via their business partner Y, manage the business relations of Russian and foreign companies established in Russia with other companies in which Y is involved via a consultancy company registered in Cyprus. The management of Cypriot companies from Dubai in turn enables the establishment of holding companies that carry out transactions via Mossack Fonseca in the British Virgin Islands.
What makes this chain so interesting is that it is connected, via its main operator, to a large-scale Russian money laundering network. In this case, 19 British companies have already been prosecuted for money laundering to the tune of 20 billion pounds. The schemes described above may be legal, and the presumption of innocence must be respected, but their discovery illustrates the fragility of the boundaries between legally permissible financial optimization and money laundering – some consulting firms such as Mossack Fonseca have certainly been complicit in large-scale money laundering operations, as the recently leaked FinCEN files show. 5
The questions and findings that have arisen over the course of our project are indicative of a much broader need for tools that enable the investigative exploration of open data in the humanities and social sciences (HSS) – a technological infrastructure we refer to as an Investigation Support System (ISS). We believe that two key building blocks are needed to establish such an ISS: visual analytics and knowledge graph data mining.
Visual analytics integrates new theoretical approaches and computational tools with innovative interactive techniques and visual representations to enable human‐information discourse. In our case, where visual analytics is applied to knowledge graphs or semantic graphs constituted from open data, DERIVO’s GraphScale and SemSpect tools have proven to be ideally suited to the task at hand.
When it comes to knowledge graph mining, relational data mining
(RDM) is an especially promising approach. Unlike traditional data
mining algorithms, which look for patterns in a single table (propositional
patterns), RDM algorithms look for patterns among multiple tables (relational
patterns). Davide Mottin’s groundbreaking work in the field allows
researchers to apply the sociological method of analytic induction to substantial
bodies of open data. By constructing exemplar patterns,
And yet, one significant methodological challenge remains. Graph pattern mining aims at identifying structures that appear frequently in large graphs, under the assumption that frequency signifies importance. Several measures of frequency have been proposed that respect the a priori property, which is essential for an eficient search of the patterns. This property states that the number of appearances of a pattern in a graph cannot be larger than the frequency of any of its sub-patterns. In real life, however, there are many graphs with weighted nodes and/or edges, in which case it would be clearly sensible for the importance (score) of a pattern to be determined not only by the number of its appearances, but also by the weights on the nodes/edges of those appearances.
Removing this obstacle will require both a rigorous methodology of frequency scoring, which Mottin et al. are in the process of developing, and suitable methods for case-building, which GraphScale and SemSpect already permit. In light of this, SemSpect’s plugging with Neo4j may well prove to have been the decisive step towards the large-scale application of analytic induction to open data.
The author would like to thank Thorsten Liebig, co-founder and CEO of DERIVO, for modeling the initial semantic graph of the ICIJ’s online database, and Vincent Vialard, senior engineer at DERIVO and co-presenter of this paper at the 2019 Graph Technologies conference, for his collaboration and astute assessment of the research process.