Tax administrations in most countries have more corporate and personal information than any other government office. Data mining techniques can be used in many different problems due to the large amount of tax returns received every year. In the present work we show an essay of the Brazilian Tax Administration on using Bayesian networks to predict taxpayers behavior based on historical analysis of income tax compliance. More specifically, we tried to improve a previous risk based audit selection which detects a large amount of taxpayers as high risk. However, in its current form it identifies much more cases than the tax auditors can handle. Our first results are promising, considerably improving tax audit performance.
INTRODUCTION
Tax administrations have more information on people
and companies than any other government office. Tax
returns, bank transactions, and invoices arrive as hundreds
of millions of records every year. The Secretariat of
Federal Revenue of Brazil (RFB) is the Brazilian Tax
Administration and Brazilian Customs as well. This
combination is a major leverage and also a challenge.
Basically, there are two types of taxes: sales taxes and
income taxes. Sales taxes includes value-added taxes and
they are based on the value of the product being sold.
Income tax is based on how much a person or a company
earns. In most countries, sales taxes amount are
considerably larger than income taxes
In order to facilitate and prioritize tax audits on personal income tax, RFB created the concept of a “fiscal lattice”. One can understand the fiscal lattice as a first audit selection based on historical risk analysis of tax compliance by taxpayers. This lattice is a complex process in which many tax auditors specialized in personal income tax frauds create risk based rules for audit selection. The main difference between a regular audit and fiscal lattice audit is that the former has a much simpler process of analysis in order to determine whether to punish a taxpayer or not.
Since the number of taxpayers has increased, and the
ratio between tax auditors and citizens has been reducing
Data mining techniques can help better selecting taxpayers for audit and the present work offers one solution to improve the selection of this kind of audits. In Section 2.1 we discuss how Bayesian networks can be used as a classification algorithm in order to create predictive models.
The document is organized as follows: Section 2 describes some background information about Bayesian networks; Section 3 details the solution for the tax audit selection problem, from its methodology to our first results; Section 4 presents the conclusion and future work. 2
BACKGROUND In this section we bring some tax administration concepts, formulate the problem assessed by the present work, and discuss Bayesian networks for prediction. 2.1
As stated by
Bayesian networks are useful to learn from data and
discover causalities between variables and it can be used
as a classifier algorithm. It is being used for
prediction in many different problems, from genetics
Na¨ıve Bayes is the most simple version of Bayesian
network. It uses strong connections between the nodes and
it considers all explanatory variables (nodes) as
independent. Despite its simpleness it has many applications
with good results and great run performance as stated in
Tree-Augmented Na¨ıve Bayes (TAN), as explained in
A source of such information, case studies,
methodologies, and best practices are intergovernmental
organizations. For tax administrations and customs the World
Customs Organization (WCO) and the Organization for
Economic Cooperation and Development (OECD) are
important sources. In a recent survey that gathered
many countries, OECD presented a comparative chart
that shows the use of data mining to detect tax fraud
Tax Administrations internal publications also present many studies that can be applied by other countries and many of them have developed methodologies based on statistical analysis and data mining to create tax compliance risk systems. Most countries use data mining for taxpayers classification considering its risks of noncompliance.
Some studies, however, reveal different data analysis
approach being held in tax administration. The US
Internal Revenue Service (IRS) uses data mining for
different purposes, according to
Another related reference is Jani Martikainens master
thesis
More related to the present work Gupta et al. in
SOLUTION AND FIRST RESULTS In this section we describe the methodology used in the present work and detail each step of the data analysis from the information and data gathering to the construction of predictive models for improvement of tax audit selection. 3.1
METHODOLOGY
The methodology of the present work follows the
wellknown CRISP-DM (CRISP-DM). The Cross Industry
Standard Process for Data Mining is a
technologyindependent methodology and reference model to
implement data mining process in every business. It
describes each phase every data mining work should pass.
Each phase is equally relevant for the success of the data
analysis process and should not be underestimated. The
process has six phases and it is possible to perform the
same step more than once. The phases of CRISP-DM
are
Once the business questions are clear, it is time to understand the required information to perform the changes needed in the business process and achieve the goals identified in the previous phase. In data understanding, all sources of information needed to perform the analysis are determined. The first insights and main patterns are also identified in the first contact with the data available from the possible sources. Each business question needs to be mapped to every data source (systems, databases, webpages, etc.) in order to address every goal and identify possible gaps and lack of information.
In
The data preparation phase covers all activities to construct the final dataset (data that will be fed into the modeling tool(s)) from the initial raw data. Data preparation tasks are likely to be performed multiple times, and not in any prescribed order. Tasks include table, record, and attribute selection, data cleaning, construction of new attributes, and transformation of data for modeling tools.
In this phase, various modeling techniques are selected and applied, and their parameters are calibrated to optimal values. Typically, there are several techniques for the same data mining problem type. Some techniques require specific data formats. There is a close link between data preparation and modeling. Often, one realizes data problems while modeling or one gets ideas for constructing new data.
At this stage in the project you have built one or more models that appear to have high quality, from a data analysis perspective. Before proceeding to final deployment of the model, it is important to more thoroughly evaluate the model, and review the steps executed to construct the model, to be certain it properly achieves the business objectives. A key objective is to determine if there is some important business issue that has not been sufficiently considered. At the end of this phase, a decision on the use of the data mining results should be reached.
Creation of the model is generally not the end of the project. Usually, the knowledge gained will need to be organized and presented in a way that the customer can use it. Depending on the requirements, the deployment phase can be as simple as generating a report or as complex as implementing a repeatable data mining process. In many cases it will be the user, not the data analyst, who will carry out the deployment steps. In any case, it is important to understand up front what actions will need to be carried out in order to actually make use of the created models. 3.2
BUSINESS UNDERSTANDING Our main goal is to improve individuals tax audit selection. We try to achieve a better audit process performance by better using the tax auditors knowledge and time available to perform these audits. As in any tax administration, there are far more taxpayers returns and information to analyze than tax officers, and to achieve the revenue goals and tax fairness it is major that the selection of audit is as risk based as possible.
In Brazil, personal taxpayers pay their income taxes every month. Since the tax is calculated on a year based, by April of the next year, taxpayers are obliged to send their income tax return in order to adjust their debt (or credit). Every year, tens of million of returns are sent to RFB, much more than it could handle if there were no risk based selection.
RFB created the concept of “fiscal lattice” to select personal income tax returns based on tax compliance risk. In this technique personal income tax fraud experts analyze the historical of all taxpayers and their previous knowledge in order to come up with parameters to select the tax returns for audit. Once caught on “fiscal lattice”, only a tax officer could release the tax return, preventing fraudsters from receiving a possible credit. There are three main purposes in using this technique: • to better select taxpayers based on tax compliance risk; • to facilitate the verification of tax auditors, since each parameter has a well defined analysis and treatment activities; • to ease the auto-correction of tax returns by taxpayers, since many of them were caught due to filling
Besides all Brazilian tax administration efforts to select
the individuals tax audits, the number of audits selected
by fiscal lattice has increased from 569,000 in 2011 to
937,000 in 20141 in contrast with the number of tax
officers, that decreased from 12,273 in 2010 to 10,419 in
2015
More specifically, we intend to use data mining techniques to discharge as many taxpayers as possible of fiscal lattice, with the minimum compliance risk to tax administration. With thousands of audits already finalized by experienced tax auditors, it is possible to assess this problem with machine learning tools and achieve best results in letting go those taxpayers that offer less risk of tax compliance.
In our first approach on trying data mining techniques to address the problem, we selected a certain RFB’s unit that has been suffering from the large number of fiscal lattice audits. The “Delegacia Especial de Pessoa Fsica” (DERPF) or “Individual Taxpayers Special Office” is an individual taxpayer specialized unit located at Sao Paulo City, the Brazilian biggest city, in the most economically active federation unit (State of Sao Paulo). This unit has come to its limit of fiscal audits since its creation in 2014, and has the largest number of this kind of audits in the whole country. It was a natural choice for our first experiments. 3.3
To answer the business question on how to improve the selection of individual taxpayers caught in fiscal lattice, we evaluated the sources of the information needed to perform the data mining analysis. Our sample was taken from audits performed by DERPF from years 2014 to 2016.
Basically, all individuals taxpayer information was taken from internal systems, from online systems to datamarts and datawarehouses. Most of taxpayer information caught in fiscal lattice is available from tax returns, but some information is taken from invoices and financial operations. The exact properties retrieved by the data extraction as well as the fraud/non-compliance rate are classified information.
The final taxpayer table has 25,322 taxpayer’s returns analyzed by tax auditors and classified as compliant or non-compliant. Each line has, besides the dependent 1In 2015 this number decreased to 670,000 due to efforts in better selecting individuals tax returns for audits variable (compliant) other 20 characteristics of taxpayers and information retrieved from returns and other systems. From these, 13,547 are women and 10,730 are men. Other explanatory variables are information of tax return and unfortunately cannot be specified because it could present classified information, since the result of the analysis could lead taxpayers to learn fraud patterns and use that information to avoid being caught. For preparation, all independent variables were analyzed in order to remove the incomplete rows and to discretize continuous ones to comply with the Bayesian network algorithms constraints. The numeric variables where classified within bands in terms of average multipliers (one average, half average, three times average, etc.). After data preparation the final number of individual taxpayers returns was 24,277.
All data preparation took place using R language2 and its packages. 3.4
We used bnlearn R package3 in order to run the Bayesian network algorithms. Specifically the functions naive.bayes and tree.bayes where chosen to create the predictive models. The first is the well-known Na¨ıve Bayes algorithm, which does not take parameters for customizing the models and the former is an implementation of the Tree-Augmented Na¨ıve Bayes (TAN) algorithm. The TAN algorithm takes white list (force the inclusion of arcs in Bayesian network), black list (force the exclusion of arcs in Bayesian network), and mi4 parameters. To create the predictive models we took the compliant variable as dependent and the other 35 (thirty five) information as independent variables. The sample of 24,277 where divided into training (80%) and test (20%). No validation sample was needed since we used 10fold cross-validation technique with bnlearn’s function bn.cv().
As stated in bn.cv() documentation
2https://www.r-project.org/. 3http://www.bnlearn.com/.
4The estimator used for the mutual information coefficients for the Chow-Liu algorithm in TAN. Possible values are mi (discrete mutual information) and mi-g (Gaussian mutual information). We use discrete since all explanatory variables have been discretized Since the proportion of compliant/non-compliant taxpayers is classified information, we present the results of the predictive models in terms of improvements from the actual process of discharging taxpayers from fiscal lattice. Since our dependent variable is compliant/noncompliant, we are interested in evaluating the models by specificity more than sensitivity, since it is more dangerous to let a non-compliant taxpayer go away without being audited than to select one that is compliant to be audited.
Each Brazilian tax administration local unit is autonomous and may choose whatever criteria it finds best to dismiss taxpayers from fiscal lattice. So, to a matter of possible comparison with our proposal, we consider a linear cut (random selection) of taxpayers until it reaches a units capacity. If, for example, an office has the capacity to audit 2,000 taxpayers per month, and there are 3,000, we consider the actual process to randomly choose the 1,000 to be dismissed. The overall taxpayers wrongly dismissed, is the same as the proportion between noncompliant taxpayers from overall caught on fiscal lattice. Our goal is to better predict if a taxpayer caught on fiscal lattice is compliant or not. If we come to a specificity considerably better than random selection, we achieve our goal to let go as few non-compliant taxpayers as possible.
As we learn from Table 1, using Na¨ıve Bayes is already a good tool to select those taxpayers which can and cannot be dismissed from being audited. Tree-Augmented Na¨ıve Bayes had no major advantages, despite the customization of parameters (root chose automatically or user defined). Therefore, the predictive models in this first results showed optimistic results, resulting in a increase of more then 30% in tax audit selection in comparison to randomly discharging taxpayers. It is major to recollect that the taxpayers caught in fiscal lattice have already been through a risk based process of selection and any increase in this criteria is a leverage in using Bayesian networks to build models of tax compliance.
CONCLUSION AND FUTURE WORK Brazil has been through a major crisis and the responsibility of the RFB as a tax administration has also increased in order to guarantee the revenue for public policies. A better selection of tax audits save resources and increase the performance of the collecting tax process. Our approach on creating predictive models to improve the risk based selection of the so called “fiscal lattice” proved to be a promising one based on the first results. We intend to use different approaches and Bayesian networks algorithms in order to create compliance risk scores and leave the decision of taxpayers being compliant or not to the tax officers and possibly increase the specificity. The approach in the present work delegates this decision to the prediction algorithm.
Furthermore we will try and build Bayesian networks with larger samples and more tax units and include more information about the taxpayer, since in this work we basically used income tax returns and registry information. Financial transactions and invoice data could be interesting explanatory variables and will be used in future applications.
Acknowledgements The authors would like to thank RFB, specially DERPF, for providing the resources necessary to work in this research, as well as for allowing its publication.