IEEE Computational Intelligence Magazine - May 2021 - 65

Torgler [33] found strong empirical evidence supporting mutual influence
between taxpayers. The investigations by
Blaufus et al. [34] and Calvet and Alm
[35] show other perspectives of horizontal factors, such as tax privacy, empathy, and sympathy.
Fedeli and Forte [36] studied the
bargaining problem between agents to
decide the level of (under) reporting of
sales and purchases along the different
steps of the supply chain, combining
sales and income tax. Each pair of agents
has to decide the amount of trade
between themselves and the optimal
level of (under) reporting. They concluded that the best policy to control
fraud is to increase the inspection
-probability rather than to increase penalties against fraudulent activities. On the
other hand, by acting against one agent,
the tax agency may eliminate the incentives for misreporting along the entire
chain, thus reducing the enforcement
costs for increasing the inspection probability for all agents. Abraham et al. [37]
offered another example of modeling
based on a non-cooperative game-theoretic model and demonstrated that social
norms are more important for explaining behavior in tax fraud than in situations where agents act independently.
Evolutionary game theory is commonly supported by ABM, a CI simulation paradigm. These CI models start
with the conviction that, especially in
the field of sales taxes, fraud always
involves more than one agent. Boadway
et al. [38] presented the payoffs of a
group of potential evaders in the form of
a prisoner's dilemma game. With collaboration, they could increase their revenues; but if one of the agents decides not
to collaborate, the rest can be negatively
affected. Following the results of their
model, we can say that a fundamental
determinant of tax fraud that requires
collaboration between two agents is the
capacity of each agent to collaborate in
the evasion activity. In the above model,
the collaboration capacity depends on
the agent's tolerance of dishonesty.
There are several advantages of using
ABM in the analysis of tax fraud. The
ABM approach enables the inclusion of

There are several advantages of using ABM in the
analysis of tax fraud. The ABM approach enables the
inclusion of agents with very different response patterns,
which allows for a dynamic bottom-up structure ... On
the other hand, the agents included in an agent-based
model do not maintain static positions and reactions.
agents with very different response patterns, which allows for a dynamic bottomup structure that may converge in a global
equilibrium that houses an intricate set of
individual realities. On the other hand, the
agents included in an agent-based model
do not maintain static positions and reactions, but can evolve throughout the simulation, learning from their own situation
and those of the other agents.This interaction between agents (i.e., social network)
is, therefore, one of the fundamental components of this type of models.
ABM examples in the study of tax
fraud can be found in the work of
Bloomquist [39], which described this
methodology as relatively recent and highlighted the apparent lack of -acceptance by
the mainstream social scientists. Hokamp
et al. [40] also discussed recent advances
in the modeling of tax fraud analysis.
The different attitudes of agents may
manifest themselves in individual perceptions of both the risk of apprehension and the audit probability [41], as in
the model presented by Korobow and
Axtell [42]. In relation to tax compliance, their model suggested that the
outcome depends on how strongly
agents are influenced by their neighbor's
results. If agents do not significantly
value the potential gains of evaders, it is
possible to achieve general compliance
even if the model realistically incorporates low enforcement measures. Another example with a similar pattern was
provided by Hashimzade et al. [43].
Combining behavioral economics and
social networks, their model reflects
how the connection between agents can
form a subjective audit probability,
which clearly exceeds the objective levels that stem from real levels of fiscal
auditing activities. The compliance decisions are modeled to include benefits for

the agents for following the social
norms of honesty. In a similar fashion,
Hashimzade et al. [44] described their
starting point in terms of attitudes,
beliefs and network effects based on two
features of empirical analysis on tax
fraud. First, there seems to be strong evidence to indicate that individual compliance attitude is clearly affected by
social norms. Second, agents do not
normally have access to the real audit
probability. Therefore, they end up generating their own subjective probability
of being subject to inspection and being
identified as an evader. A similar basis
was used by Hashimzade and Myles [45]
in their effort to incorporate predictive
analysis and behavioral economics in
modeling the tax compliance decision.
Previous efforts, such as those in the
field of experimental analysis, are very relevant and should continue to be so.
Bloomquist [46] presented a good example of how the results of experimental
economics can be exploited in the development of ABM. New efforts such as
those coming from the field of econophysics should be closely observed in order to
identify possible c- omplementarities [47]. It
has already been mentioned that network
tools are an essential part of ABM; therefore, delving into how the network structure influences the results of these models
is an important avenue of study. The first
step in this direction came from Andrei
et al. [48], who explored the impact of
different network structures on the levels
of tax evasion in an agent-based model.
Reality is complex and models seeking to describe it must be able to reflect
this complexity. Calibrating the models
to real data is not an easy task when
dealing with ABM and tax fraud. However, linking the main elements of an
agent-based model to real data or

MAY 2021 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE

IEEE Computational Intelligence Magazine - May 2021

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - May 2021