## 1armea2020001

## Source details

**Canonical URL:** [1armea2020001](https://www.imf.org/-/media/files/publications/cr/2020/english/1armea2020001.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/cr/2020/english/1armea2020001.pdf.md)
- [Structured JSON version](/-/media/files/publications/cr/2020/english/1armea2020001.pdf.json)

---

### Executive summary — key findings and strategic direction
- Mission: Fiscal Affairs Department (FAD) mission visited Yerevan during February 6–19, 2019 to advise on strategic choices for tax administration and compliance risk management (CRM).
- Leadership and participants: led by Mr. Vincent de Paul Koukpaizan (FAD); included Messrs. Enriko Aav (FAD), John Crotty (FAD external expert), and Stuart Hamilton (FAD external expert).
- Main observations and strategic direction:
  - Armenia’s tax policy setting is likely to create new noncompliance opportunities and result in revenue leakages.
  - SRC should link business strategies to major functional operations and ensure basic functions of tax administration operate effectively.
  - On average, 35 percent of large taxpayers submitted their value added tax (VAT) returns after the due dates during 2018; good practice aims for all large taxpayers to file returns on time.
  - Accuracy of the taxpayer register and audit coverage (through diversification of interventions, including issue-based audits) require priority attention.
  - Two practices should be discontinued:
    - Concessional treatment of voluntary disclosures made up to and including the conduct of an audit.
    - Advance publication of the names of taxpayers included in the annual audit program.
  - SRC draft strategic plan for 2019-2022 relies excessively on extensive and fully automated data matching and envisages a “closed system” capturing all business transactions regardless of size and risk; this approach is costly and risks driving transactions into the informal economy.
  - SRC needs better segmentation of compliance management by risk and taxpayer size; additive risk rule scoring should be supplemented by predictive modeling.
  - Basic SRC operational data (e.g. filing) were not available until late in the mission, indicating insufficient use of operational information for decision-making.
  - Integrity and good governance in SRC processes are essential to reduce vulnerability to corruption and build legitimacy in revenue collection.
- Capacity development (CD):
  - A new 24-month EU-funded project launched in January 2019 will help design and implement analytical tools to enhance SRC compliance and risk management and improve tax auditing.
  - CD delivery: headquarters missions, staff and short-term expert (STX) visits; first STX visit planned for June 2019.
  - Success of CD requires timely provision of requested data.

### Summary of recommendations (condensed)
- Short-Term Measures:
  - Set up formal CRM processes for planning, executing, and monitoring SRC activities: develop a compliance strategy defining main compliance risks and SRC responses; establish governance arrangements; measure impact of SRC activities on tax compliance.
  - Initiate targeted actions, including strong sanctions and other disincentives to deter late filing.
  - Develop performance management to ensure SRC units achieve on-time reporting and full payment.
- Medium-Term Measures — Tax system vulnerabilities:
  - Reduce complexities of Armenia’s tax policy setting.
  - Withdraw the annual option for taxpayers to switch between presumptive and general tax regimes.
  - Consider legislative changes to clarify tax liabilities of “self-employed” persons to prevent circumvention of PAYE withholding.
  - Consider legislative changes to address company loans and similar arrangements used to arbitrage CIT and PIT liabilities.
- Medium-Term Measures — Strategic and operational choices:
  - Use the report’s framework to reduce vulnerabilities to corruption and increase taxpayer confidence.
  - Establish procedures for ongoing regular cleansing and updating of the taxpayer register.
  - Discontinue annual publication of the list of taxpayers selected for audit.
  - Eliminate preferential treatment for voluntary disclosures after initiation of an audit.
  - Engage taxpayers through targeted segment-oriented approaches and compliance campaigns.
  - Develop estimates of relative tax gaps by industry to prioritize compliance efforts.

### I. Vulnerabilities of the tax system — critical challenges
- Policy and regime complexity:
  - Corporate income tax (CIT) preferences, special tax allowances, and tax rate differentials present major compliance management challenges.
  - Special or presumptive regimes (patent fee, turnover tax, self-employed, family company regimes) add complexity, inequity, and undermine fairness.
  - Planned rationalization commencing in 2020: establishment of a new regime for micro-businesses covering family company regime and self-employed persons; under proposed micro-business regime, earnings below 24 million Armenian Drams (AMD) would be exempt from income and turnover taxes.
- VAT thresholds and changes:
  - VAT threshold for 2018 was AMD 115 million but reduced to 58.3 million in 2019.
  - Government plans to reverse the reduction in forthcoming tax reform package.
- Turnover tax and IFRS effects:
  - Turnover tax may encourage understatement of income and income splitting; proposed exemption under 24 million AMD likely exacerbates this.
  - IFRS accounting requirements can contribute to income under-reporting by small-medium businesses seeking to avoid IFRS.
- Specific legislative vulnerabilities:
  - Tax regime switching: qualifying taxpayers can opt, by February 20 each year, for special/presumptive or general tax regime — creates tax planning opportunities and complicates SRC benchmarking and CRM.
  - Self-employment reclassification: shifting from employment to self-employment undermines PAYE and social security contributions; self-employed persons are not required to make social security contributions and are not being audited at this time.
  - CIT–PIT arbitrage: proposed reduction of the CIT rate to 18 percent while PIT remains at 23 percent may open arbitrage (e.g., substituting loans for dividends); unclear treatment of interest-free or low-interest shareholder loans.

### Fundamentals of tax administration strategies and CRM
- Draft SRC strategic plan 2019–22 aims to reduce informality and achieve sustained revenue growth via: (1) improved taxpayer services; (2) effective communication; (3) enhanced IT systems; and (4) enhanced human resources — but does not sufficiently inform compliance management direction.
- Compliance approach continuum:
  - Voluntary compliance characteristics: self-assessment, risk-oriented non-intrusive administration, limited information collection, cooperative relationships, lower costs.
  - Coerced compliance characteristics: no self-assessment, capture of all transactions, full monitoring and punishment, no trust, high administrative and taxpayer costs.
  - SRC’s current practice aiming at capturing information on all transactions leans toward a coerced compliance approach.
- CRM processes:
  - Good practice: establish formal CRM processes for detecting risks, planning, executing and monitoring interventions, with formal governance.
  - Mission did not identify existence of formal CRM processes in SRC.

### Ensuring integrity and good governance in revenue administration
- Corruption risks:
  - Revenue administration vulnerable to corruption due to broad powers, interaction with most economic agents, and operation in remote border areas.
  - Corruption reduces compliance, increases evasion, and fosters informality, placing government revenues at risk.
- Good governance framework features (selected highlights, as presented):
  - Sound Policy and Legislation: equity, efficiency/neutrality, simplicity, transparency; clear administrative/procedural laws; effective dispute settlement; limited discretionary power; legal framework supporting modern systems and reporting.
  - Modern Systems and Processes: public reporting of work plans and outcomes; streamlined collection systems; service-oriented approach; risk-based administration; special programs for largest contributors; automated risk assessment and monitoring.
  - Streamlined Organization and Management: independence of revenue administration; function-based design; strong headquarters oversight; internal audit and anti-corruption units; external oversight focused on performance.
  - Extensive Use of Technology: digitalized processes; automated internal controls and audit logs; technology to notify citizens and collect feedback, including integrity hotlines.
  - Leading People Management & Institutionalized Integrity: merit-based HR policies; fixed tenure for senior management; rotation policy; ongoing training; enforced Code of Conduct; technology to detect unethical behavior; legal sanctions publicly applied.
- Recommendation: use the framework provided as a reference to reduce vulnerabilities to corruption and increase taxpayer confidence and trust.

### Core tax administration processes, filing, register, and operational performance
- Core functions must operate effectively: taxpayer registration, filing, and payment.
- Taxpayer register:
  - A cleanup was undertaken in 2018; as law does not allow deregistration, SRC moved inactive taxpayers out of active register with restoration possible if activity resumes.
  - No procedure established for regular updating; need ongoing regular cleansing/updating procedures.
- Large business VAT filing performance (2018) — Table 2 (VAT Returns Submitted by Large Taxpayers during 2018, SRC):
  - January: On time 850, Late 581, Total 1431, % of late filings 40.60
  - February: On time 911, Late 523, Total 1434, % of late filings 36.47
  - March: On time 883, Late 553, Total 1436, % of late filings 38.51
  - April: On time 592, Late 847, Total 1439, % of late filings 58.86
  - May: On time 984, Late 468, Total 1452, % of late filings 32.23
  - June: On time 1007, Late 449, Total 1456, % of late filings 30.84
  - July: On time 1001, Late 428, Total 1429, % of late filings 29.95
  - August: On time 1053, Late 376, Total 1429, % of late filings 26.31
  - September: On time 753, Late 675, Total 1428, % of late filings 47.27
  - October: On time 1162, Late 271, Total 1433, % of late filings 18.91
  - November: On time 1262, Late 169, Total 1431, % of late filings 11.81
  - December: On time 745, Late 687, Total 1432, % of late filings 47.97
- Provisions undermining audit effectiveness:
  - Voluntary disclosure provisions that eliminate penalties when disclosure is made after audit initiation undermine audit effectiveness and act as ongoing compliance amnesty.
  - Annual publication of taxpayers selected for audit provides up to a year’s notice enabling hiding/altering records, may incentivize underreporting and cause reputational damage without redress, and reduces SRC’s ability to address emerging risks.
- Recommendations (short term):
  - Initiate targeted sanctions and disincentives for late filing.
  - Develop performance management to ensure on-time reporting and full payment.
- Recommendations (medium term):
  - Establish ongoing taxpayer register cleansing procedures.
  - Discontinue annual publication of audit selection lists.
  - Eliminate preferential treatment for voluntary disclosure after audit initiation.

### Analytic capabilities, transaction monitoring, and costs
- SRC analytic status:
  - Upgraded additive risk scoring but lags in predictive and risk-differentiated approaches (ML).
  - Uses census approach to transaction-level monitoring: matched some 14.5 billion transactions in an economy of roughly 3 million people and GDP of about 6,083 billion AMD (12.4 billion USD).
  - Comparator: Australian Tax Office matched 650 million transactions in population 25 million and GDP 1.5 trillion USD.
- Transaction-level monitoring trade-offs:
  - Reduces transactional noncompliance opportunities but imposes costs on compliant taxpayers and administrative burden on SRC.
  - SRC has good pre-filling and OLAP capabilities but needs development and deployment of ML approaches.
- Recommendations:
  - Develop and deploy ML approaches; supplement additive risk rules with predictive modeling.
  - Promote analytics unit as high-prestige talent program.
  - Use open-source analytical tools to contain software costs (examples cited: KNIME®, WEKA, ANACONDA, RATTLE).

### Risk-differentiated audit selection, empirical findings, and limitations of additive scoring
- Audit results 2015–18:
  - Mean audit adjustment increased from 9.8 million AMD to 16 million AMD between 2015-17 and 2017-18.
  - Median case outcome declined slightly from 2.25 million AMD to 1.9 million AMD — indicating a few extremely large cases drove mean increase.
- Additive risk scoring limitations:
  - Aggregate additive scores do not reliably approximate risk-adjusted value; distribution of audit adjustments is highly skewed (not Gaussian).
  - Regression of aggregate risk scores against outcome thresholds showed essentially no predictive capability for larger size adjustments for 2015–17; aggregate scores performed no better than random for targets of 1.5m AMD and 2.5m AMD.
  - 10% of cases > mean contain 82% of the value of adjustments.
- Recommendation:
  - Develop estimates of relative tax gaps by industry and use risk-differentiated approaches and compliance campaigns to prioritize interventions.

### Predictive modeling applications, methodology, and practical guidance (Appendix III–V)
- Potential applications of advanced analytics (pattern recognition, outlier detection, cluster analysis, experimental design, network analysis, text mining) across tax lifecycle:
  - Registration: predict who should be registered; predict revenue from non-registered persons.
  - Filing: predict late filers before event; predict revenue associated with late/non-filing.
  - Reporting: predict likelihood of non-compliance and size of potential adjustment; predict high-risk refunds; predict objections to amended assessments.
  - Payment: predict late payers before event; predict payment response to interventions; predict capacity to pay and business viability.
  - Service: inform channel design and self-service opportunities; improve service delivery via proactive messaging.
  - Border protection and policy: predict smuggling likelihood; tax gap measurement; forecast impact of tax policy changes.
- Practical model development tips (CRISP_DM and data guidance):
  - Prefer categorical to free text; ordinal over categorical; interval over ordinal where possible.
  - Manipulate mirror data sets, save SQL scripts, use representative sampling for very large datasets.
  - For rare targets (<10%), consider oversampling, under sampling, or SMOTE.
  - Use Random Forest initially; use ensemble methods; eliminate non-predictive variables; normalize where appropriate.
- Deployment and maintenance:
  - Once robust, rebuild the model using full training partition; use CART for explanatory output; revisit models every six months; include stratified random component to capture new risks and estimate tax gaps.
- Empirical experiment (Armenian data):
  - Created binary target ‘Strike Y/N’ with threshold $Adjustment$ >= 1500000.
  - Data partitioned 50/50 training/verification; Random Forest trained and tested.
  - ROC analysis showed additive aggregate scores were no better than random for Armenian targets of 1.5m AMD and 2.5m AMD.
- Comparator country example:
  - With richer data and Random Forest, ROC curves indicated significantly better case selection performance.
  - Typical predictive model dataset fields: Identifiers; Demographic Data; Interaction Data; Financial Data; Risk Data; Adjustment, Penalty Amount.
  - KNIME workflow example: Excel input → partition → Random Forest learner/predictor → ROC / confusion matrix → RF variable importance → CART for explanation → deploy.
- Single-tree decision-rule example (extracted rule):
  - Example decision rule: If 2746069.5 AND $Expenses VAT$ <= 5583015.5 AND $Sales Exempt$ <= 4.6373646E7 AND $Sector$ IN ('G', 'I') AND $Industry$ IN ('0099', '0018','0020', '2231','0226', '0011','0073', '0045' etc.) => “Y” probability 0.722.
  - Random Forest description: several hundred decision trees vote on classification; more accurate but less explainable; risk-adjusted value calculation R = L x C where L = likelihood (from model), C = consequence.
- Consequence estimation approaches:
  - Data matching with third-party information; industry benchmarking; cumulative curve analysis by turnover size to derive median expected adjustments as proxy for consequence.
  - RAV (Risk Adjusted Value) = L x C; prioritize by RAV to maximize long-term return in skewed populations.
- Model evaluation metrics:
  - Strike Rate (precision) = TP/(TP+FP); Miss Rate = FN/(TP+FN); Accuracy = (TP+TN)/(TP+TN+FP+FN); ROC area indicates model quality.
- Empirical model performance and revenue improvement potential:
  - Reported potential improvement in revenue from model-based case selection: between a 9% and 68% percentage point improvement in revenue if audit caseloads maintained (final outcome may be toward lower end depending on audit capability).
  - Example tabulated comparisons indicate substantial increases in Strike Rate (precision) and Revenue per Case when moving from original risk rules to data mining/model extrapolation.

### Technical assistance (IMF FAD / IMF-EU CD program) and planned activities
- FAD will help enhance SRC’s CRM through an IMF-EU CD program running January 1, 2019 to December 31, 2020 focused on strengthening Public Finance Management and fiscal governance in Eastern Partnership countries.
- For Armenia, CD aims to improve revenue mobilization through strengthened tax CRM and development of a modern compliance planning model and tools.
- Tentative FAD CD activity schedule (selected entries):
  - June 2019: CRM Expert visit — corporate priorities better managed through effective risk management.
  - September 2019: Follow up HQ staff visit — support SRC’s CRM program and STX’s work.
  - September 2019: CRM Expert visit — CRM continued.
  - November 2019: CRM Expert visit — CRM continued.
  - January 2020: Tax administration core processes HQ staff visit — audit and verification programs more effective.
  - March 2020: Follow up HQ mission — review progress and plan further FAD CD.
  - May 2020: Tax audits Expert visit — audit and verification programs more effective.
  - September 2020: CRM Expert visit — CRM continued.
- FAD agreed to help assess the VAT gap under IMF’s RA-GAP in response to MOF request. Further assistance subject to successful implementation and request.

### Key prioritized actions (operational checklist)
- Implement formal CRM processes: compliance strategy, governance, impact measurement.
- Address filing compliance in large business segment: target reduction of average 35 percent late VAT filing in 2018.
- Establish ongoing taxpayer register cleansing and updating procedures.
- Remove provisions undermining audits: eliminate post-audit voluntary disclosure concessions; stop publication of audit selection lists.
- Move from additive risk scoring toward predictive modeling (Random Forest, ensembles) and RAV-based prioritization (R = L x C).
- Build analytic capability using open-source tools (examples: KNIME®, WEKA, ANACONDA, RATTLE); treat data as an asset and institutionalize analytics unit.
- Use formal modeling processes (e.g., CRISP_DM), save SQL scripts, address class imbalance (SMOTE) for rare targets, and retrain models at least every six months with a stratified random component.

*Source: IMF FAD mission aide-mémoire (content unit 1armea2020001).*

### PREFACE _________________________________________________________________________________________ 5

### 1armea2020001 - PREFACE

### Preface: mission overview
- A Fiscal Affairs Department (FAD) mission visited Yerevan during the period February 6–19, 2019.
- Main purpose: advise on strategic choices for tax administration and compliance risk management (CRM).
- Mission leadership and participants:
  - Led by Mr. Vincent de Paul Koukpaizan (FAD).
  - Included Messrs. Enriko Aav (FAD), John Crotty (FAD external expert), and Stuart Hamilton (FAD external expert).
- Key interlocutors: Mr. Arman Poghosyan, Deputy Minister of Finance; Mr. Davit Ananyan, Chairman of the State Revenue Committee (SRC); Mr. Rafik Mashadyan, First Deputy Chairman of the SRC; Mr. Mikayel Pashayan, Deputy Chairman of the SRC; and several SRC officials.
- Acknowledgements: special thanks to Ms. Narine Fahradyan for facilitating meetings; gratitude to IMF Resident Representative Ms. Yulia Ustyugova and her staff.
- Structure of the aide-mémoire: Executive Summary and three sections — (I) Vulnerabilities of the Tax System; (II) Strategic Choices for Tax Administration; and (III) Technical Assistance Available.

### Executive summary: key findings and strategic direction
- Purpose: complement the March 2018 tax administration mission and focus on persistent issues hampering effective compliance management, including tax policy framework and SRC’s business processes.
- Primary observations:
  - Armenia’s tax policy setting is likely to create new noncompliance opportunities and result in revenue leakages.
  - SRC should link business strategies to major functional operations and ensure basic functions of tax administration operate effectively.
  - On average, 35 percent of large taxpayers submitted their value added tax (VAT) returns after the due dates during 2018; good practice aims for all large taxpayers to file returns on time.
  - Accuracy of the taxpayer register and audit coverage (through diversification of interventions, including issue-based audits) require priority attention.
  - Two practices identified that should be discontinued:
    - Concessional treatment of voluntary disclosures made up to and including the conduct of an audit.
    - Advance publication of the names of taxpayers included in the annual audit program.
  - SRC draft strategic plan for 2019-2022 relies excessively on extensive and fully automated data matching and envisages a “closed system” capturing all business transactions regardless of size and risk; this approach is costly and risks driving transactions into the informal economy.
  - SRC needs better segmentation of compliance management by risk and taxpayer size, consistent with May 2018 recommendations.
  - SRC’s additive risk rule scoring approaches need to be supplemented by predictive modeling; predictive modeling is envisaged in SRC’s draft strategic plan.
  - Basic SRC operational data (e.g. filing) were not available until late in the mission, which may indicate insufficient use of operational information for decision-making.
  - Integrity and good governance in SRC processes are essential to reduce vulnerability to corruption and build legitimacy in revenue collection.
- Capacity development (CD) support:
  - A new 24-month European Union (EU)-funded project launched in January 2019 will help design and implement analytical tools to enhance SRC compliance and risk management and improve tax auditing.
  - CD delivery: headquarters (HQ) missions, staff and short-term expert (STX) visits.
  - First STX visit planned for June 2019.
  - Success of CD will require timely provision of requested data.

### Summary of recommendations (Table 1 — condensed)
Short-Term Measures
- Set up formal CRM processes for planning, executing, and monitoring SRC activities to improve tax compliance, including:
  - developing a compliance strategy that defines the main compliance risks and SRC responses;
  - establishing governance arrangements for implementation of the compliance strategy;
  - measuring the impact of SRC activities on tax compliance.
- Initiate targeted actions, including strong sanctions and other disincentives to deter late filing.
- Develop a performance management process that monitors and ensures SRC units effectively achieve on-time reporting and full payment of tax obligations by registered taxpayers.

Medium-Term Measures — Vulnerabilities of the tax system
- Reduce complexities of Armenia’s tax policy setting, given revenue leakages, the impact on taxpayer compliance costs, and challenges for SRC in managing compliance.
- Withdraw the annual option available to certain taxpayers to switch from the presumptive to the general tax regime, or vice-versa, to take advantage of tax minimization opportunities.
- Consider legislative changes to clarify the tax liabilities of “self-employed” persons who seek to circumvent PAYE withholding obligations.
- Consider legislative changes to address the use of company loans and similar financial arrangements to arbitrage CIT and PIT liabilities.

Medium-Term Measures — Strategic and operational choices for tax administration
- Use the framework provided in the report to reduce vulnerabilities to corruption and increase taxpayer confidence and trust.
- Establish procedures for ongoing regular cleansing and updating of the taxpayer register.
- Discontinue annual publication of the list of taxpayers selected for audit.
- Eliminate preferential treatment for voluntary disclosure of underreporting after initiation of an audit.
- Engage with taxpayers through a more targeted segment-oriented approach and use compliance campaign approaches where appropriate to ‘break through’ entrenched non-compliance.
- Develop an estimate of the relative tax gaps associated with strategic compliance risks, using a consistent measure by industry to prioritize compliance efforts.

### I. Vulnerabilities of the tax system — critical challenges for tax administration
Main issues summarized
- Complex, inequitable, and distortionary aspects of Armenia’s tax policy framework arise from corporate income tax (CIT) preferences, special tax allowances, and tax rate differentials that present major compliance management challenges for SRC.
- Special or presumptive tax regimes (patent fee, turnover tax, self-employed individuals, special family company regimes) add complexity and inequity and undermine simplicity and fairness.
  - Planned rationalization commencing in 2020 includes establishment of a new regime for micro-businesses covering the current family company regime and self-employed persons.
  - Under the proposed micro-business regime, earnings below 24 million Armenian Drams (AMD) would be exempt from income and turnover taxes.
- VAT thresholds and changes:
  - The VAT threshold for 2018 was AMD 115 million but has been reduced to 58.3 million in 2019.
  - The government plans to reverse the reduction of the VAT threshold in the forthcoming tax reform package.
- Turnover tax may encourage understatement of income and/or income splitting to remain outside the general tax system; the proposed exemption under 24 million AMD is likely to exacerbate this.
- IFRS accounting requirements can contribute to income under-reporting by small-medium businesses seeking to avoid being required to use IFRS.

Specific vulnerabilities and legislative considerations
- Tax regime switching:
  - Qualifying taxpayers, by February 20 of each year, can opt for taxation under a special/presumptive or under the general tax regime, creating tax planning opportunities and arbitrage.
  - Switching complicates SRC’s industry benchmarking and case selection, making CRM more difficult.
- Self-employment reclassification:
  - Shifting from employment to self-employment undermines PAYE withholding collections and social security contributions.
  - The proposed micro-business exemption for earnings below 24 million AMD increases the incentive to assume self-employed status.
  - Self-employed persons are not required to make social security contributions and are not being audited at this time.
  - Legislative remedies in other countries include deeming rules based on deriving a predominant amount of income (for example, 90 percent) from a single economic group to require PAYE and social security obligations.
- CIT–PIT arbitrage:
  - Proposed reduction of the CIT rate to 18 percent while PIT remains at 23 percent may open further arbitrage opportunities (e.g., substituting loans for dividends).
  - It is unclear how Armenian tax law treats interest-free or low-interest loans to shareholders and executives made out of undistributed profits — whether such loans would be treated as “dividends” or “income.”
  - Some countries address this with deemed dividend and presumptive interest rate provisions.

Recommendations (medium term) — reiterated
- Reduce complexities of Armenia’s tax policy.
- Withdraw the annual option to switch between presumptive and general tax regimes.
- Consider legislative changes to clarify tax liabilities of “self-employed” persons who seek to circumvent PAYE withholding obligations.
- Consider legislative changes to address use of company loans and similar arrangements to arbitrage CIT and PIT liabilities.

_This summary is based on the PREFACE and Executive Summary sections of the IMF FAD mission aide-mémoire._

### 9.      This section advises on key areas of focus when developing tax administration strategies.

### 9. This section advises on key areas of focus when developing tax administration strategies.

### A. Fundamentals of Tax Administration Strategies

- The SRC has prepared a draft strategic plan to guide tax administration reforms in 2019–22.  
  - The plan aims to reduce informality and achieve sustained tax revenue growth through: (1) improved taxpayer services; (2) effective communication with the taxpayers; (3) enhanced IT systems; and (4) enhanced human resources.
  - Although it covers some key strategic areas, the plan does not inform on the SRC’s strategic direction to compliance management.
  - Footnote: The mission was provided the draft strategy plan but was not able to discuss it with the SRC management during its stay.

- Strategic choice for SRC:
  - SRC faces a strategic choice in building tax administration operations and compliance management; approaches vary from knowledge-based, research-driven strategies to less formal operational practice development.
  - Choice depends on environmental factors including: expectations of the government, culture of the public service and of the country, availability of resources, human capacity, and access to information on good practices.

- Compliance approach continuum:
  - Tax administration approaches to compliance management lie on a continuum between voluntary (unenforced) compliance and coerced (fully enforced) compliance.
  - Voluntary compliance features (as described):
    - Self-assessment of taxpayer obligations;
    - Risk-oriented, non-intrusive tax administration;
    - Emphasis on informing and supporting taxpayers to prevent non-compliance;
    - Volumes of information on taxpayer limited to strict needs of tax administration;
    - Ensuring a low compliance burden is one of the key performance objectives of tax administration;
    - Cooperative relationship between representatives of industries/taxpayers and tax administration;
    - Lower costs.
  - Coerced compliance features (as described):
    - No self-assessment of tax obligations;
    - Tax administration seeks to capture data on all transactions;
    - Emphasize on full monitoring and punishment;
    - No trust between taxpayers and tax administration;
    - Taxpayers burdened with requirements that distort their business operations;
    - Sizeable cost of investments in tax administration technology and operations.
  - Current SRC practice aiming at capturing information on all transactions of all taxpayers leans further in the direction of a coerced compliance approach.

- CRM (Compliance Risk Management) processes:
  - Good practice for developed tax administrations is to establish formal CRM processes: detecting compliance risks, planning, executing and monitoring activities to address those risks, with formal governance arrangements.
  - The mission did not identify the existence of formal CRM processes during its discussions with the SRC.
  - Appendix 1 (of source) provides further details on strategic planning cycle for compliance risks.

- Recommendation (short term):
  - Set up formal CRM processes for planning, executing, and monitoring SRC activities to improve tax compliance. These include developing a compliance strategy that defines the main compliance risks and SRC responses to these risks, governance arrangements for the implementation of the compliance strategy, and measurement of the impact of SRC activities on tax compliance.

### B. Ensuring the Integrity of Revenue Administration

- Context and priority:
  - Rooting out corrupt practices from public administration, including revenue administration, is one of the top declared priorities of the new government. This objective is highlighted in the draft government program presented recently to the Parliament.

- Risks and impacts of corruption in revenue administration:
  - Revenue administration is frequently vulnerable to corruption because it:
    - Affects economic interests of businesses and individuals through interactions with most members of society;
    - Safeguards significant government revenues and its officials are vested with broad powers;
    - Often operates in remote and isolated border areas.
  - Corruption in revenue administration risks being imputed to the rest of the public service and political leadership.
  - Corruption is influenced by broader environmental factors and ethical norms of a country; it is difficult to isolate an “island of integrity in a sea of corruption.”
  - Lack of integrity puts significant government revenues at risk: corruption is accompanied by evasion and low levels of taxpayers’ compliance resulting in significant revenue losses and constrained financing for priority government policies.
  - Low levels of compliance enabled by corrupt officials facilitate the development of the informal economy.

- Strategy to reduce vulnerabilities and build legitimacy:
  - Governments need to invest in modernizing tax administration and building greater legitimacy in the collection of revenues; a whole-of-government approach is crucial.

- Key features of a good governance framework in revenue administration (as presented):
  - Sound Policy and Legislation
    1. Revenue policy designed based on principles of equity, efficiency/neutrality, simplicity, and transparency.
    2. A common set of administrative and procedural laws that are simple and reliable for different tax types.
    3. Legal framework provides appropriate balance between rights of taxpayers and powers of revenue administration, supported by effective dispute settlement procedures (e.g. independent tribunal/court or tax ombudsman) and legal safeguards against the improper exercise of powers by revenue administration (e.g. opportunity for taxpayers to pay overdue taxes before forced sale of property seized through distraint).
    4. A system of tax self-assessment is in place promoting voluntary compliance by taxpayers.
    5. Clarity and stability of law, rules, and processes, including minimal discretionary power vested in the revenue administration, and where discretion is unavoidable, clear conditions on how discretion will be exercised.
    6. Legal and human resource frameworks allow for firing of officers behaving unethically and provide a suite of appropriate sanctions for cases of lower culpability, with prosecution for criminal activities.
    7. Legislation allows for adoption of modern systems, processes and technology in revenue administration and sets out key aspects of organization and management (including relationship between Ministry and the revenue administration), including express legislative requirements for revenue administration to provide and publish reports on its operations and financials on a regular basis.
  - Modern Systems and Processes
    8. Revenue administration work plans, budget, performance objectives, and outcomes are regularly publicly reported.
    9. Collection systems and procedures are streamlined to secure timely revenues without imposing undue compliance cost and inconvenience to the business.
    10. Service-oriented approach ensuring taxpayers have the information (quantity, quality, comprehensiveness) and support they need to meet their obligations voluntarily.
    11. Availability of a tax rulings function with clear and straightforward rules to avoid distinct tax treatments that deviate from the general rules and pose transparency concerns.
    12. A general risk-based approach is adopted in the administration aimed at detecting and acting on taxpayers who present the greatest risk to the revenue system.
    13. Special programs using modern and transparent approaches to manage the compliance of the largest contributors, including large businesses, high-wealth individuals, and high-income earners.
    14. Effective and impartial dispute resolution process is available and publicized.
  - Streamlined Organization and Management
    15. Revenue administration is established with independence from political direction, e.g., reports to Minister of Finance who has overall fiscal responsibility, rather than to the Prime Minister or President.
    16. A function-based organization design with separation of duties and appropriate numbers of staff assigned to each function based on workload.
    17. Strong headquarters function providing oversight and uniform operations across the field network.
    18. Streamlined field operations and organizational alignment to key taxpayer segments.
    19. Effective internal audit and investigation/anti-corruption units established, with relationships and cooperation with public service wide anti-corruption activities and bodies.
    20. Strong oversight of revenue administration by external bodies (General Audit Office, Ministry of Finance) focused on monitoring performance but not allowed to interfere in specific taxpayers’ affairs.
  - Extensive Use of Technology
    21. Revenue administration processes are digitalized and automated to the extent possible.
    22. Robust automated system of internal control checks and monitoring of processes, with access controls/audit logs.
    23. Automated risk assessment and case selection is in place.
    24. Technology supports notification of citizens about their obligations and correct procedures for revenue administration.
    25. Technology supports collection of feedback from the public on interactions with revenue administration staff, including reporting unethical behavior, e.g., through a dedicated integrity hotline.
  - Leading People Management
    26. Human resource policies and processes assure merit-based selection, appointment, appraisal, and promotion of revenue officials.
    27. Senior management of revenue administration is appointed for a fixed period (tenure).
    28. Management process built on minimal management layers with appropriate spans of control, and internal control is one of the core management functions.
    29. Salaries set at a sufficient and competitive level.
    30. A formal rotation policy supports staff development, with a cycle to allow staff to build expertise and contribute to the respective function’s performance.
    31. Ongoing staff training programs delivered so officials know their duties, conditions of service, and sanctions for wrongdoings.
  - Institutionalized Promotion of Integrity
    32. Staff is regularly informed about and supported in adopting positive behavior; corporate practice, including through an enforced Code of Conduct, strongly signals zero tolerance towards low staff integrity.
    33. Technology solutions to detect unethical behavior are routinely used.
    34. Legal sanctions are effectively applied on each detected corrupt behavior and publicly announced.

- Recommendation (medium term):
  - Use the framework provided in this report as a reference to reduce vulnerabilities to corruption and increase taxpayer confidence and trust in the tax system.

### C. Tax Administration Processes and Compliance Management

- Core functions:
  - Critical to taxpayer compliance management is development of core tax administration processes.
  - Basic functions—taxpayer registration, filing, and payment of taxes—must operate effectively; failure in these core functions will prevent securing due revenues.
  - Example: If taxpayers are not registered, they are likely to operate outside the tax net as part of the informal economy.

*Source: 1armea2020001 - section 9.*

### 22.      While SRC is eager to implement modern approaches in tax administration

### 1armea2020001 - 22.      While SRC is eager to implement modern approaches in tax administration

### Core tax management and compliance focus
- Modern analytic approaches (including artificial intelligence and other sophisticated technologies) are only effective if SRC has robust processes to ensure taxpayers comply with basic registration, filing, and payment obligations.
- The SRC must maintain a focus on core tax management processes before or alongside deploying advanced analytics.

### Taxpayer register
- A taxpayer register cleanup was undertaken in 2018.
- As the tax law does not allow taxpayer deregistration, the SRC moved inactive taxpayers out of the active taxpayers’ register; if SRC receives information that these taxpayers are not inactive they can be restored as active.
- It appears that no procedure has been established to ensure regular updating of the taxpayer register; the one-off cleanup was worthwhile but ongoing regular cleansing and updating procedures are needed.

### Filing and payment compliance — Large business VAT filing performance (2018)
- Filing compliance of the large business segment is low; on average, 35 percent of large taxpayers were regularly late with submission of VAT returns during 2018.
- Table 2. VAT Returns Submitted by Large Taxpayers during 2018 (SRC)
  - January: On time 850, Late 581, Total 1431, % of late filings 40.60
  - February: On time 911, Late 523, Total 1434, % of late filings 36.47
  - March: On time 883, Late 553, Total 1436, % of late filings 38.51
  - April: On time 592, Late 847, Total 1439, % of late filings 58.86
  - May: On time 984, Late 468, Total 1452, % of late filings 32.23
  - June: On time 1007, Late 449, Total 1456, % of late filings 30.84
  - July: On time 1001, Late 428, Total 1429, % of late filings 29.95
  - August: On time 1053, Late 376, Total 1429, % of late filings 26.31
  - September: On time 753, Late 675, Total 1428, % of late filings 47.27
  - October: On time 1162, Late 271, Total 1433, % of late filings 18.91
  - November: On time 1262, Late 169, Total 1431, % of late filings 11.81
  - December: On time 745, Late 687, Total 1432, % of late filings 47.97

### Provisions undermining audit effectiveness
- Current provisions permitting voluntary disclosure of underreported taxes undermine the effectiveness of SRC’s audit program because no penalties apply when disclosure is made after an audit has been initiated; this creates a systemic incentive for noncompliance and functions as an ongoing compliance amnesty, especially problematic given limited audit coverage.
- Publication of the list of taxpayers selected for inclusion in SRC’s annual audit plan on its website:
  - Promotes transparency and integrity but provides up to a year’s advanced notice to taxpayers, enabling hiding or altering of business records.
  - May create incentives for non-targeted taxpayers to underreport.
  - May cause reputational damage to listed taxpayers without redress even if SRC later finds their tax affairs in order.
  - Reduces SRC’s ability to dynamically address emerging compliance risks.

### Recommendations — Short term
- Initiate targeted actions, including strong sanctions and other disincentives to deter late filing.
- Develop a performance management process that monitors and ensures that SRC units effectively achieve on-time reporting and full payment of tax obligations by the registered taxpayers.

### Recommendations — Medium term
- Establish procedures for the ongoing regular cleansing and updating of the taxpayer register.
- Discontinue annual publication of the list of taxpayers selected for audit.
- Eliminate preferential treatment for voluntary disclosure of underreporting after initiation of an audit.

### Analytic capabilities for effective compliance risk management
- The SRC has upgraded its additive risk scoring system but could achieve higher compliance with predictive and risk-differentiated approaches.
- Armenia uses a census approach to transaction-level monitoring (all products and cash register transactions monitored). Examples and comparative statistics:
  - SRC matched some 14.5 billion transactions in an economy of roughly 3 million people and GDP of about 6,083 billion AMD (12.4 billion USD).
  - By contrast, the Australian Tax Office (ATO) matched 650 million transactions in a population of 25 million and a GDP of 1.5 trillion USD.
- Detailed transaction-level monitoring:
  - Reduces opportunity for transactional noncompliance but imposes costs on compliant taxpayers and administrative burden on SRC.
  - SRC appears to have very good pre-filling and On Line Analytic Processing (OLAP) capabilities but lags in development and use of analytical modeling (ML) approaches.
- SRC needs a determined commitment to enhance analytic capabilities:
  - ML approaches need to be developed and deployed to enhance SRC’s CRM.
  - Existing additive risk rule scoring approaches should be supplemented by predictive modeling to better predict and prioritize likelihood and potential consequences of non-compliance.
  - Sufficient audit data exists now to build much improved audit selection systems.
- Technical considerations:
  - Sophisticated software often expensive; suggestion to use open-source software that is free.
  - Administrations must actively manage data — collect, evaluate, manage, disseminate — and treat data as an asset; IT departments have a key role.
  - Example low-entry analytical package: KNIME®; other open-source examples: WEKA, ANACONDA, RATTLE.
  - SRC should promote the analytics unit as a high prestige talent program.

### Risk-differentiated approaches and audit case selection
- Risk-differentiated approaches should be pursued to lower compliance burden on compliant taxpayers while focusing on potential non-compliance; the SRC’s structure provides a basis but vigilance is needed to avoid one-size-fits-all responses.
- Audit results 2015-18:
  - Average (mean) result increased from 9.8 million AMD to 16 million AMD between 2015-17 and 2017-18.
  - Median case outcome declined slightly from 2.25 million AMD to 1.9 million AMD, indicating a few extremely large cases drove the increase in the average.
- Recommendations (medium term):
  - Develop an estimate of the relative tax gaps associated with strategic compliance risks, using a consistent measure by industry to prioritize compliance efforts.
  - Engage with taxpayers through a more targeted segment-oriented approach and use compliance campaign approaches where appropriate.

### Technical assistance available (IMF FAD / IMF-EU CD program)
- FAD will help enhance SRC’s CRM through an IMF-EU capacity development (CD) program running from January 1, 2019 to December 31, 2020; main focus is to strengthen Public Finance Management systems and fiscal governance in the Eastern Partnership countries.
- For Armenia, CD aims to improve revenue mobilization through strengthened tax compliance risk management and to develop a modern, sophisticated compliance planning model and related tools.
- Tentative plan for FAD CD activities in SRC for 2019–20 (selected entries):
  - June 2019: CRM Expert visit — Corporate priorities are better managed through effective risk management.
  - September 2019: Follow up HQ staff visit — Support the SRC’s CRM program and support of STX’s work.
  - September 2019: CRM Expert visit — Corporate priorities are better managed through effective risk management.
  - November 2019: CRM Expert visit — CRM continued.
  - January 2020: Tax administration core processes HQ staff visit — Audit and other verification programs more effectively ensure accuracy of reporting.
  - March 2020: Follow up HQ mission — Follow up and review of progress in tax administration CRM. Plans for further FAD CD program.
  - May 2020: Tax audits Expert visit — Audit and other verification programs more effectively ensure accuracy of reporting.
  - September 2020: CRM Expert visit — Corporate priorities are better managed through effective risk management.
- FAD agreed to help assess the VAT gap under the IMF’s Revenue Administration Gap Analysis Program (RA-GAP) in response to the MOF’s request.
- Further FAD technical assistance could be considered subject to successful implementation and a request from the authorities.

*Source: SRC. Excerpt from IMF/FAD mission report and IMF-EU CD program materials.*

### Appendix III. Potential Uses of Predictive Modeling in Tax

### Appendix III. Potential Uses of Predictive Modeling in Tax Administration

### Applications of advanced analytics
- Advanced analytics techniques cited: pattern recognition, outlier detection, cluster analysis, experimental design, network analysis, text mining.
- Other applications (apart from case selection) include:
  - Registration
    - Predict who should be registered.
    - Predict revenue associated with a non-registered person.
  - Filing
    - Predict who will file late before the event.
    - Predict who will file once they are late (self-finalize).
    - Predict revenue associated with late or non-filing.
  - Reporting
    - Predict who is non-compliant (likelihood) for each tax type.
    - Predict size of potential adjustment (consequence).
    - Predict high risk refunds.
    - Predict who will object to an amended assessment.
    - Support text and social network mining in audit cases.
  - Payment
    - Predict who will pay late before the event.
    - Predict who will pay late but before intervention (self-finalize).
    - Predict who will pay given alternative interventions (phone, mail, visit, court action etc.).
    - Predict capacity to pay and propensity to pay.
    - Predict business viability (see BVAT model on the ATO website).
  - Service
    - Taxpayer channel use to inform design decisions and identify self-service opportunities.
    - Improve service delivery using proactive messaging, calling, and other interventions.
  - Border protection
    - Predict likelihood of smuggling contraband, drugs etc.
    - Predict behavior of traders and passengers.
  - Policy
    - Tax gap measurement.
    - Assessing or forecasting the impact of changes in tax policy.

### Practical tips regarding predictive model development and use (Appendix IV)
- Use a formal process, such as CRISP_DM, to document data mining efforts for better analysis, consistent reuse and learning.

- Data guidance:
  - Where possible have/use categorical data rather than free text fields.
  - Where possible use ordinal data (categories that have an order) rather than categorical data.
  - Where possible use interval data (numeric with meaningful spacing) rather than ordinal data.
  - Manipulate and transform mirror data sets - not the original data.
  - Save SQL scripts for consistent data retrieval, reuse and process documentation.
  - If faced with extremely large data sets, use representative sampling to reduce the size analyzed.
  - If the target variable is rare (e.g. <10%) consider oversampling the target, under sampling the negative class or generate additional ‘synthetic’ examples using the SMOTE node in KNIME.
  - Always review the data and understand the distributions involved, particularly of target variables.

- Modeling / Mining guidance:
  - Use a good ‘out of the box’ algorithm, such as Random Forest’ initially. As expertise develops explore the use of other modeling approaches to see if they can improve predictions over part of the data. (It is usually hard to beat Random Forest in practice.)
  - Use ensemble approaches (taking the best predictions from multiple models) when appropriate.
  - To reduce ‘noise’ evaluate and then eliminate variables that don’t provide predictive ability. (The Meta Node provided for Random Forest ‘variable importance’ indicates the relative use of variables in the random forest.)
  - Some modeling approaches work best with normalized interval data. Explore whether such transformations improve the predictions. (KNIME Normalizer node.)
  - To reduce the number of low value cases selected iteratively raise the threshold for a ‘strike’ and evaluate the results until a suitable balance between strike rate and caseload is obtained. (The median strike value is often a good starting point.)
  - If the threshold reduces the target percentage below ~10% consider over / under sampling or SMOTE (synthetically generating targets using a SOM clustering algorithm) to rebalance the data set.

- Deployment guidance:
  - Once a robust predictive model has been built and evaluated, for the deployment build push the ‘training partition’ to 100% and re-execute the learner node to maximize the deployed models’ predictive ability.
  - Use a single regression tree (CART) to provide a broad explanation of what the more accurate ‘black box’ model is doing.
  - Revisit models every six months with additional data to see if they need to be rebuilt or enhanced.
  - Use a small random component (using stratified random sampling) to maintain intelligence on new risks and monitor model performance over time. A random component also assists in estimating tax gaps and prioritizing compliance campaigns.

### Analysis of SRC comprehensive audit case selection (Appendix V)
- Context and current SRC approach:
  - SRC uses both comprehensive audits and thematic reviews.
  - SRC uses a weighted additive risk scoring approach: a series of ‘risk rules’ are defined, taxpayer data is parsed against each rule, a score is given, and aggregate scores are used to prioritize cases.

- Empirical audit adjustment distribution (Armenian comprehensive audit results, 2015–2017):
  - Audit adjustments over three years 2015-2017 are highly skewed.
  - Descriptive statistics reported:
    - Mean adjustment: 9.8m AMD.
    - Modal adjustment threshold used in analysis: 1.5m AMD.
    - Median adjustment threshold used in analysis: 2.5m AMD.
  - Strike rates for the 2015–17 period using different thresholds:
    - Using raw risk rules: 93% adjustment rate with mean adjustment of 9.8m AMD.
    - With modal threshold of 1.5m AMD: strike rate 68 percent.
    - With median threshold of 2.5m AMD: strike rate 47 percent.
    - With mean threshold of 9.8m AMD: strike rate 13 percent.
  - Additional note: 10% of cases > mean contain 82% of the value of adjustments.

- Limitations of additive risk scoring:
  - Aggregate (additive) risk scores do not reliably approximate risk-adjusted value; Risk Score A + Risk Score B + Risk Score C does not equal the combined risk-adjusted value.
  - Additive scoring produces an approximately Gaussian distribution of risk scores, while underlying tax compliance risk distributions are almost always highly skewed and better approximated by a log-normal distribution.
  - Regression of aggregate risk scores against thresholds (mode, median, mean) showed essentially no predictive capability for indicating larger size adjustments for 2015–17; aggregate scores were effectively random within the audited case results.
  - Conclusion: the aggregate risk score more accurately functions as an “indicator of concern” rather than a reliable prioritization of the riskiest taxpayers.

- Predictive modeling alternative and experiment:
  - Predictive data mining approaches (e.g., Random Forest) can regress taxpayer features against a target variable such as whether there was an adjustment >= X (e.g., 1.5 million AMD).
  - Example experiment using available audited data:
    - A new column ‘Strike Y/N’ was appended using KNIME rules engine with code: $Adjustment$ >= 1500000 => “Y”.
    - Data partitioned 50/50 into training and verification sets.
    - Training set parsed through a Random Forest learner; model used to predict verification set results.
    - Efficacy of 2015–17 aggregate risk score examined via a Receiver Operating Characteristic (ROC) Curve approach.
  - Findings: essentially no relationship between aggregate risk score and likelihood of selecting a case that would produce median or modal audit results over 2015–18; ROC analysis showed aggregate scores were no better than random for targets of 1.5m AMD and 2.5m AMD in the Armenian data.

- Comparator country example and results:
  - In a comparator country using a richer dataset (demographic, interaction, financial and risk score data) and Random Forest, ROC curves indicated significantly better case selection performance than the Armenian additive approach.
  - Typical predictive model dataset fields listed (each type may have multiple columns):
    - Identifiers (Not used)
    - Demographic Data (Region & Industry)
    - Interaction Data (Files O/S, Debt late, Last audit)
    - Financial Data (VAT & CIT Amount, P&L & Balance Sheet Amount)
    - Risk Data (Score for each risk rule & overall)
    - Adjustment, Penalty Amount
  - KNIME workflow example:
    - Excel input → partition into training/verification → Random Forest Learner/Predictor → evaluate via ROC Curves and Confusion Matrices.
    - Produce RF Variable Importance, generate simple CART for explanatory purposes, deploy model to mine big cases.
    - Workflow components include: Excel Reader (XLS), Partitioning, Random Forest Learner, Random Forest Predictor, Decision Tree Learner (CART), ROC Curve, Confusion Matrix, Rule Engine, Excel Writer (XL), and other KNIME nodes.
  - Explainability: a simple decision tree (CART) used for explanatory purposes, e.g. an example rule fragment: “$Sales Exempt$ >766473.0 AND $Expenses Exempt$ <= ...” (workflow demonstrates generating human-interpretable rules alongside higher-performing “black box” models).

### Key recommendations and operational notes
- Move beyond additive risk scoring to predictive modeling (e.g., Random Forest, ensemble methods) for improved prioritization of audit resources.
- Use formal processes (CRISP_DM) and reproducible workflows (KNIME) including saved SQL scripts and documented partitions.
- Address class imbalance when target variable is rare (<10%) via oversampling, under sampling, or SMOTE.
- Regularly re-evaluate models (every six months) and include a stratified random component to capture new risks and estimate tax gaps.
- Use CART or single regression trees for explanatory output to complement higher-accuracy models.
- Consider thresholds (mode, median, mean) and iterate thresholds to balance strike rate and caseload; avoid relying solely on aggregate additive scores for prioritization.

*Source: Appendix III–V, "Potential Uses of Predictive Modeling in Tax Administration" (IMF content unit 1armea2020001).*

### 2746069.5 AND $Expenses VAT$ <= 5583015.5 AND $Sales Exempt$ <= 4.6373646E7 AND

### 1armea2020001 - 2746069.5 AND $Expenses VAT$ <= 5583015.5 AND $Sales Exempt$ <= 4.6373646E7 AND

### Decision-rule example and machine learning comparison
- A single decision-tree risk rule derived from the data:
  - If 2746069.5 AND $Expenses VAT$ <= 5583015.5 AND $Sales Exempt$ <= 4.6373646E7 AND $Sector$ IN ('G', 'I') AND $Industry$ IN ('0099', '0018','0020', '2231','0226', '0011','0073', '0045' etc.) => “Y” probability 0.722.
  - The text notes this single-tree rule is more complex and more accurate than simple additive risk scoring rules manually developed by the tax administration.
- Random Forest description:
  - Several hundred decision trees (the forest), produced by repeatedly random sampling the data set with a sample size of roughly the square root of the data set, vote on classification.
  - Random Forests are harder to explain (a ‘black box’) but often more accurate and produce a more robust likelihood of adjustment (L) for use in risk adjusted value calculation (R = L x C).

### Calculating potential audit Consequence (C) and Risk Adjusted Value (RAV)
- Approaches to estimate consequence (C):
  - Data matching with third-party information can define potential adjustment.
  - Industry sector benchmarking can provide estimates with uncertainty.
  - Cumulative curve analysis by turnover size can provide median expected adjustments as a proxy for consequence when direct indicators are unavailable.
- Risk Adjusted Value (RAV):
  - RAV = L x C.
  - Audit ranking (case prioritisation) should be by RAV to maximise long-term return on investment.
  - In skewed tax populations, prioritisation using RAV usually produces significantly superior revenue results compared with ranking by likelihood (strike rate) or cyclical auditing (e.g., once every three years).

### Empirical model results and case selection comparisons (figures from comparator country)
- Reported potential improvement in revenue from case selection models: somewhere between a 9% and 68% percentage point improvement in revenue if audit caseloads were maintained. Audit capability may push final outcome toward the lower end of that range.
- Example aggregate comparisons (multiple tabulations presented):
  - Original Risk Rules / Data Mining / Model Extrapolated (selected figures):
    - Case Load: 10,217  —  439  — 10,217 (various columns)
    - Adjustments > 100,000: 1,294  —  2305,353  — (varied)
    - Candidate Cases: 1,294  —  439  — 35,603
    - Total Revenue Raised: $3,501,448,807  $678,178,097  $15,783,475,199
    - Strike Rate (Precision): 12.7% 52.4% 52.4%
    - Revenue per Case: $342,708 $1,544,825 $1,544,825
    - Revenue per Strike: $2,705,911 $2,948,600 $2,948,600
  - Alternate tabulation (Original / Risk / Rules Data Mining / Model Extrapolated):
    - Case Load: 10,217  —  380  — 10,217
    - Adjustments >100,000: 1,294  —  2225,969
    - Candidate Cases: 1,294  —  380  — 32,877
    - Total Revenue: $3,501,448,807   $679,395,470 $18,266,798,728
    - Strike Rate ('Precision'): 12.70% 58.4% 58.4%
    - Revenue per Case: $342,708 $1,787,883 $1,787,883
    - Revenue per Strike: $2,705,911 $3,060,340 $3,060,340
- Reported R² for fitted relation: R² = 0.94589 with equation y = 20973x -1.004 (presented in text).
- Cumulative-case examples and median adjustments by turnover segment (selected values):
  - Segment thresholds and median adjustments:
    - Extra Large: >$1,400,000,000 => $2,041,964
    - Large: >$86,401,885 => $956,642
    - Medium: >$19,110,326 => $349,951
    - Small: >$28,744 => $138,144
  - Alternate numeric thresholds and medians shown: 1,400,000,000 86,400,000 19,100,000 2,042,000 350,000 957,000 138,000 30,000 (presented in table/chart format in source).

### Model evaluation metrics and confusion matrix example
- Definitions provided:
  - Strike Rate (precision) = TP/(TP+FP)
  - Miss Rate = FN/(TP+FN)
  - Accuracy = (TP+TN)/(TP+TN+FP+FN)
  - ROC Curve: Greater area under curve implies better selection model.
- Confusion Matrix example counts (as presented):
  - Selected / Not Selected breakdown:
    - TP 1538, FN 5423 (Non Compliant)
    - FP 1938, TN 11543 (Compliant)
    - Totals shown: 3476 and 15428 (model/prior change table layout in source)
  - Reported metrics in that example block:
    - Strike Rate 44%    29%   54%
    - Miss Rate 72%0%-72%
    - Accuracy 69%    29%  141%
    - Revenue 95%   100%-5%
    - Revenue case 977,991  189,602  416%
    - Caseload 18%   100%-82%

### Cumulative curve and inflection-point approach
- Cumulative Curve Inflection Point Approach:
  - Identify turning points on matched cumulative curves of turnover and adjusted tax.
  - Use median adjustment between turning points as a proxy for consequence in R = L x C.
  - This reduces overestimation of consequence in highly skewed populations and produces stepwise consequence proxies by company size to improve prioritization reliability.

### Glossary of technical terms and tools referenced
- CART: Classification and Regression Tree — decision tree algorithm; explainable but tends to over-fit.
- Random Forest: ensemble of decision trees (e.g., created by repeated sampling, sometimes 500 times); relatively fast, copes with missing values and non-numeric data, resistant to over-fitting, but less explainable.
- Self-Organizing Map: descriptive clustering algorithm to identify taxpayers deviating from peer behavior.
- KNIME: KoNstanz Information MinEr — open source data analytics, reporting and integration platform; node-based workflows and meta nodes used to simplify complex workflows.
- Cumulative curve, ROC curve, confusion matrix described with usage for model evaluation.
- Median adjustment by turnover example nodes:
  - Median Adj values: 72.944, 55.071, 28.651 mapped to turnover buckets (>$340000, <$340000 & >100000, <=100000 respectively) and corresponding $Turnover$ rules shown.

*Italic: Source: 1armea2020001 - 2746069.5 AND $Expenses VAT$ <= 5583015.5 AND $Sales Exempt$ <= 4.6373646E7 AND (PDF chapter content).*

---


_Source: https://www.imf.org/-/media/files/publications/cr/2020/english/1armea2020001.pdf_