Risk Based Tax Audit - PDF

Document Sample
Risk Based Tax Audit - PDF Powered By Docstoc
					Charles Vellutini

Risk Based: Tax MINING Selection Methods

Washington, June 2, 2010                             1

1. Why Risk Based Audit Selection?
2. How to do it? Statistical Methods for RBA
3. A case study: regression-based predictive

1. Why Risk Based Audit

The role of audits

 Detect and redress individual cases of non-
 Promote voluntary compliance by increasing the
  probability of detection and penalties for non-compliant
      This impact critically depends on a properly designed audit selection
       strategy focusing on high-risk taxpayers  threaten to risky, bad
       taxpayers, not others

      Audits provide a good opportunity for the tax administration to
       educate taxpayers on their legal obligations or book-keeping
       requirements, thereby improving future compliance.

 Gather information on both the health of the tax
  system (by measuring the share of non-compliant
  taxpayers and the amount of unpaid taxes, for example)
  and the evasion techniques used by taxpayers.

Audit selection methods

 Manual selection of audit cases by auditors based on
  their own knowledge of the taxpayers‟ behavior and
      Uses auditor‟s knowledge of taxpayers and environment

      Cannot uncover patterns of non-compliance hidden in the history of
       non-compliance in the same area, sector, or as determined by other
       taxpayer attributes;

      Discretionary, subjective approach  favors rent-seeking and
       corruption in the tax administration

      Has fallen into disfavor in OECD countries

 Random selection, including stratified sampling
      No bias in audit selection  useful to fight corruption

      Perceived as fair by taxpayers

      But clearly not focused on highest risks  high opportunity cost if
       used as sole selection method

 Risk-based selection = identify those taxpayers that
  are the most likely to be non-compliant
Risk based selection

 Select the most „risky‟ cases for detailed audit:
      Better use of scarse resources of tax administrations

      Fewer cases audited = Lower cost of tax collection

      Most professionally competent officers can be deputed to tax audit

 Data-driven: case selection based on objective criteria,
  not left to the discretion of the tax official
 Reduce opportunities for rent seeking behavior :
      Reduce interface between tax inspectors and taxpayers

      Help fight corruption

 …But costs in terms of data, IT systems and training

Separating audit case selection from audit

         Traditional audits           Risk-based audits

                                           Central risk
                                           analysis &

                      Local                               Local Tax
   Taxpayer                     Taxpayer       Audit
                     Tax Unit                               Unit

          International experience in risk-based audits

Country          Manual                 Random                         Risk-based
                 selection              selection                      selection
USA                 No                    Stratified sampling by        Discriminate function
                                            taxpayer category, for         (DIF score), based on
                                            income tax;                    prior random samples.
                                           Around 2.000 random
                                            audits every year for
                                            learning purposes
                                            (unbiased data)
                                           Long experience of
                                            randomly-selected audits
                                            (first program in 1963).
Canada              No                    Stratified sampling by        Macro-level analysis
                                            industry and revenue           using time series
                                            range                          econometrics
                                           1000-2000 yearly              Audit selection based on
                                            random audits                  data mining techniques
UK                  Yes, for 55 % of      Simple random sampling        Extensive data matching
                     audits.                for self-assessment            and data mining
                                            taxpayers;                     techniques (incl. decision
                                           Yearly program: 6,800          trees and neural networks)
                                            random audits per year        Score-based risk
                                            (10 % of the audits).          assessment (estimation
                                                                           method unknown);
                                                                          35 % of the audits;
                                                                          Regression analysis for
                                                                           VAT returns.         8
           International experience in risk-based audits

Country           Manual          Random audits    Risk-based
                  selection                        audits
Kazakhstan        Abandoned in    No program          Centralized, based
                  2010                                 on manually
                                                       designed risk
Tanzania          Abandoned in    No program       Audit selection based
                  2007                             on data mining since

India             Only in cases   No program       Audit selection based
                  of external                      on risk profiling
                  information                      developed through
                                                   external data
                                                   sources, various
Turkey            No              Yes, to gather    Centralized at
                                  unbiased data       national level
                                                    Implemented by
                                                      tax instruments
                                                      (VAT, CIT, etc.)
2. How to do it: Statistical
    Methods for RBA

The problem: using data to assess
risk compliance

Data matching: tracking down observed

Checking consistency of tax returns
with other sources:
  Other taxpayers
Tax-specific: i.e. VAT – check returns
filed by both sides of a transaction
Very effective but IT intensive :
amount of data and processing power is a
major challenge
Modeling options for predicting risk(i):
Data mining techniques

 Decision tree        Neural network        Clustering

Automated segmentation techniques with no structural
underlying models: looking for correlations in the data
Works well with very large data sets
“Black box” effect: useless to suggest causal
relationships and hard to interpret
Difficult to correct selection biases
Modeling options for predicting risk(ii):

 Classical linear regression analysis is most
 commonly used when the dependant variable is
 continuous = relevant to audit case selection
 Long tradition in selection bias correction
 Robust to small datasets (from 40/50
 observations if randomly drawn)
 Proven techniques, likely to be well-known to
 local analysts and included in standard
 econometric software
 Can use data from data-m atching mechanisms
 as explanatory variables
 Other more recent econometric methods include
 matching techniques and propensity scores
         Predictive techniques used in India

Model                    Method                       Target Variable

 1      Classification tree                                -compliance

 2                                              Compliance/Non
        Same classification tree as Model 1(after            -compliance
        dropping cases with missing data)

 3      Classification tree                   Amount of underreported tax (3

 4      Logistic regression                                -compliance

 5      Discriminant function                              -compliance

 6      Discriminant function                 Amount of underreported tax (3

 7      Hybrid model based onmodels 6 and a Amount of underreported tax (3
        classification tree                 categories)

 8      Hybrid model based on models 6 & 2    Amount of underreported tax (3

Need for a country-specific strategy

 No single “right” approach : a wide
array of choices
Right strategy a function of
  IT resources
  Analytical resources
A dynamic learning process:
  several approaches can be developed
  in parallel
  but need for clear focus when
  analytical resources are a constraint

3. A case study: regression-
 based predictive modeling

        Data (1)

Data from tax returns forms filled in by taxpayers,
which are :

  VAT returns

  Profit tax returns, which include information on
  turnover, profit and expenses.

Data on tax payments, described by type of tax

Data on audit results, including detailed
information regarding which type of tax resulted in
tax adjustment, penalty or interest.

          Data (2)

Number of taxpayers and of audited taxpayers

                  tax_year        8     999

                      2005       42    201
                                 42    121

                      2006      295    208
                                225    119

                      2007      903    202
                                158     53

                      2008   1,100     190
                                44       6

1st row: number of taxpayers
2nd row: number of audited taxpayers
999: Large Taxpayers Unit
8: Region X                                    19
       Modeling approach : the Heckman
1. The audit selection equation is estimated (probit
   model); the dependent variable is a binary variable
   taking the value “1” if the taxpayer has been audited
   and “0” otherwise:
   •   Sample with audited and non-audited taxpayers.
   •   The predicted likelihood of being audited is
2. The compliance equation is estimated on the sample
   of audited taxpayers only.
   •   The predicted likelihood of being audited is included
       as a control variable.
   •    The selection bias is corrected.

           The Compliance equation


     is the monetary outcome of the audit of taxpayer

       are n attributes and behavioral variables of i
– including predicted selection likelihood,

       is a set of parameters to be estimated and
is an error term.

            Predicting compliance risk
                                            (1)          (2)
VARIABLES                                audit_tot      select
Annual turnover                         0.00100***
Growth rate of profits                     1828        -0.00171
                                          (0.168)      (0.152)
Deviation of profit rate from sector    -8.318e+06     1.322**
                                          (0.188)      (0.0157)
legal_form_code==A                     1.378e+06***
legal_form_code==JSC                   2.655e+06**
legal_form_code==LLC                   2.746e+06**
legal_form_code==SOE                   2.125e+06**
1 if audited in n-1                                   0.549***
tax_year==2006                                        -0.368***
tax_year==2007                                        -0.792***
tax_year==2008                                        -1.617***
Observations                               384           384       22
Rho                                       -0.936        -0.936
Findings from Compliance Equation (col 1
 Turnover = strong determinant of audit_tot =
simple size effect

Growth rate of profits influences audit_tot.

Deviation of the profit rate from sector mean =
typical explanatory variable used by many tax
administrations as an indicator of risk. Works well here

Legal status also influences compliance risk.

Findings from Selection Equation (col 2)
 The deviation of the profit rate with sector
averages explains how taxpayers were selected for
audit, which is intuitive enough.

The presence of an audit in year n-1 explains the
likelihood that it will audited again in year n.

The growth of profits also influences audit selection

Yearly dum mies are important, as audit practices
have changed in recent years

Parameter rho with a very significant negative value
of -0.936  audit selection is negatively correlated
with actual audit results

Computing risks scores by deciles
           Summary results by decile

      Decile Number Average
             of firms predicted
                      audit results
           1       11     -739 459
           2       12    1 448 202
           3        7    2 162 345
           4       11    2 477 869
           5       11    2 731 789
           6        8    3 014 082
           7       12    3 338 449
           8        9    3 827 023
           9       15    4 521 087
         10        15 22 690 300
      Total       111    5 301 577

 Scores to be used in audit planning,
  matching risk levels with adequate
  audit actions
 Random audits still needed for
  benchmarking and « learning » of the
 Model to be refined and tested against
  next wave of audits
 Data collection, model estimation and
  score computation to be automated        26
The risk assessment cycle


Shared By:
Description: Risk Based Tax Audit document sample