## tarea2024011

## Source details

**Canonical URL:** [tarea2024011](https://www.imf.org/-/media/files/publications/tar/2024/english/tarea2024011.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/tar/2024/english/tarea2024011.pdf.md)
- [Structured JSON version](/-/media/files/publications/tar/2024/english/tarea2024011.pdf.json)

---

### I. Introduction and mission purpose
- Follow-up to the November 2022 field-based assignment to help improve risk assessment capabilities for the Republic of Slovenia Revenue Administration (SFA).
- Short assignment (virtual) focused on resolving implementation questions as SFA develops modernized approaches to risk assessment for corporate income tax (CIT).
- Key implementation topics reiterated:
  - Ideal organizational arrangements to support data analytics
  - How to access and utilize better and more data, including third party data
  - Data workflows and processes with a view to streamlining and improving them
  - How to improve data integrity
  - Using data science and big data analytics to strengthen risk assessment
  - Leveraging the value of country-by-country reports
- Mission observations:
  - STX impressed by rapid adoption of tools and methods from November mission and high engagement of dedicated staff.
  - SFA progressing on: a) emphasizing taxpayer size when determining risk; b) introducing empirically founded business risk rules when determining audits; c) building advanced predictive models.
  - Recommendation to introduce empirical tools gradually, starting with lowest-hanging fruits (simple analytics and weighting by firm size) before full deployment of advanced machine-learning models.

### II. Summary of advice given on November 2022 mission
- Foundational strengths identified:
  - Highly competent “Analysis unit” and timely investments in data cleaning, manipulation and storage.
  - High-quality data on audits, tax returns, Country-by-Country-Reports, business risk rules and balance sheets, all mergeable using unique taxpayer identifiers.
- Priorities and practical steps:
  - Utilize existing data sources more systematically to automatically identify risks via data-driven models.
  - Business risk indicators set up for transfer pricing were accurate in predicting yield of CIT comprehensive audits—opportunity to reuse existing work for CIT audit selection and for risk profiles of largest taxpayers.
  - Previous audit strategies/business risk rules were not being systematically evaluated; recommend simple historical evaluations (e.g., scatterplots).
  - Use simple benchmarks and standard deviations (industry profitability, standard deviations) to root business risk in systematic analysis rather than intuition.
  - Introduce thresholds in measure of strike rates so marginal corrections (e.g., a correction of €20) do not contaminate success measures; strike rates should reflect meaningful adjustments.
  - The 30 largest taxpayers are always a priority and should be continuously observed; create automated risk profiles for the 30 largest taxpayers and establish one-to-one client account managers.
- Pilot models constructed (data/code/workflows handed over to SFA):
  - An advanced (random forest) CIT audit selection model (incl. deployment workflow)
  - An advanced (random forest) VAT audit selection model
  - Workflow to evaluate existing business risk rules against historical data (VAT/CIT)
  - Country-by-Country-Reporting anomaly detection tool
- Model objectives and results:
  - All models targeted real-time risk assessment using all available data.
  - Demonstration: the CIT audit selection model could increase strike rates by 20-40 pct.-pts.
- Recommended priorities for the coming year (in order):
  - Create risk profiles for the 30 largest taxpayers.
  - Create business risk indicators for CIT.
  - Put larger weight on firm size and redefine strike rates to reflect firm size.
  - Evaluate existing business risk rules using historical data.
  - After these adjustments, proceed with more advanced modelling.

### III. Summary of technical assistance advice (workshop findings and SFA plans)
- Background and SFA implementation plan:
  - SFA requested follow-up work one month after initial assignment, indicating commitment.
  - SFA short-term plan for 2023:
    - Emphasize large taxpayers in CIT; implement existing and new strategies for selection of taxpayers for CIT supervision using internal and external data.
    - Use known tools (e.g., SAP Fraud Management) and include CIT risk indicators recently identified; test predictive model in practice on a smaller scale.
    - Adapt annual supervision plan of the Special Financial Office (SFO); comprehensive audits of all taxes planned by teams including transfer pricing specialists.
    - Use automatic risk analysis in transfer pricing combined with predictive analytics and current case selection processes.
    - Use questionnaires, preventive visits, and cooperative compliance approaches to promote voluntary compliance for large taxpayers.
  - STX endorsed incremental approach from simple to advanced empirical analysis.
- Workshop outcomes and model performance:
  - SFA built upon the Random Forest model from November 2022 using KNIME®.
  - Advanced model substantially outperformed current audit selection models for CIT:
    - Increased predicted strike rate (measured as corrections larger than $5 thousand) from 30 to 70 percent.
- Areas for further model improvement identified by mission team:
  - Separate models should be developed for comprehensive audits and partial audits; also separate models for each type of partial audit (e.g., interest deduction audits, desk audits). Models must use comparable outcome/intervention in training data.
  - Model training currently uses corrected tax returns (post-audit) rather than initial returns; audit selection in practice is based on initial returns, so feeding corrected returns hampers prediction of non-compliance.
- Additional recommendations:
  - Proceed with simpler risk analyses before deploying most advanced models; test simple CIT risk rules against past audit results using KNIME® workflows.
  - Use RCTs and “double-blinded horse races” for model evaluation during gradual implementation.
  - Build advanced risk assessment models for Customs and VAT first (abundance of cases makes them easier to predict).
  - Technical correspondence on AI/Random Forest and KNIME® summarized in Appendix I.

### IV. Proposed next steps (same as in November 2022 mission)
- Prioritization / senior management endorsement (Short term: within next 12 months)
  - Prioritize/endorse high-value data analytics work; create protected time for development and deprioritize low-value tasks.
  - Move towards “top-down” risk assessments in CIT by defining and using existing business risk indicators.
  - Increase emphasis on taxpayer size in risk assessments: risk profile the 30 largest taxpayers; create client owners; increase threshold for “strikes”; weigh risk rules according to consequence.
  - Endorse multifaceted compliance risk management (audits plus service/education).
  - Move towards automatic exchange of third-party information from banks, regulators and other stakeholders (Medium term: within two to three years) — may require legislative action.
- Improvements in data analytics and way-of-work (Short term: within next 12 months)
  - Readjust risk rules to reflect expected adjustment/risk by multiplying current risk rules with measures of size (turnover, tax paid, employees).
  - Build automated reports for the largest taxpayers (Analysis unit and large taxpayer unit); reports to indicate suggestive actions such as “contact taxpayer” or “audit”.
  - Build CIT business risk indicators (begin with current TP business risk rules).
  - Evaluate existing risk rules based on historical data and advanced models; update accordingly (see KNIME workflow).
  - Expand use of anomaly detection in transfer pricing using CbCR-data + other data sources; inform which of the >2,300 cases of firms outside safe-harbor thin capitalization rules to pursue; prioritize by taxpayer size.
  - Advanced data analytics projects after implementing low-cost investments (1-5), e.g., random forest models; use pilot models from mission as inspiration.
  - Evaluate implementation of data analytics projects and devise remedying actions if necessary; anticipate and address resistance to change.
- Organizational structure (Short term: within next 12 months)
  - Ensure the Analysis unit supports the entire SFA (potentially by referring directly to the DG or as part of a CRMU).
  - Reestablish the Compliance Risk Management Unit.
- Staffing and training (Short term: within next 12 months)
  - Prioritize end-user education in all existing applications; no application should be built without prioritizing end-user education and feedback.
  - Establish learning programs where data scientists teach non-experts in data analytics to alleviate Analysis unit workload and spread capabilities; KNIME® suggested as a low-barrier training tool.
  - Consider improving terms for data analysts to ensure recruitment and retention (longer contracts, time-conditioned compensation “Golden handcuffs”).
- Implementation timelines and responsibilities
  - Multiple actions assigned deadlines of “Short term (within next 12 months)” and one action assigned “Medium term (within two to three years)”.
  - Responsibilities allocated to Director General (DG)/Heads of department, Analysis unit, relevant business owners, IT, HR, Working Groups (WG), large taxpayer unit, and relevant stakeholders as specified in Table 1.

### Retrain and embedding programs (Section 2)
- Retrain 1-3 non-data scientists from SFA to join the Analysis unit
  - Time horizon: Short term (within next 12 months)
  - Responsible parties: Data Science Team and HR
  - Purpose:
    - The intention is two-fold: 1) bring business knowledge into the Data Science Team 2) expand the capacity of the Data Science Team.
  - Implementation note:
    - One possibility is to exploit the KNIME® software and available courses to ease this process.
- Start “tours” program where non-data analysts from SFA are embedded in the Analysis unit and vice versa
  - Time horizon: Medium term (within two to three years)
  - Responsible parties: DG and relevant heads of division in consultation with Analysis units
  - Implementation guidance:
    - Make sure this aligns with the projects in the pipeline – e.g. if a customs selection project is taking off, then it make sense to collaborate with customs.

*Source: tarea2024011*

### Section 1

### tarea2024011 - Section 1

### I. Introduction and mission purpose
- Follow-up to the November 2022 field-based assignment to help improve risk assessment capabilities for the Republic of Slovenia Revenue Administration (SFA).
- Short assignment (virtual) focused on resolving implementation questions as SFA develops modernized approaches to risk assessment for corporate income tax (CIT).
- Key implementation topics from prior mission reiterated:
  - Ideal organizational arrangements to support data analytics
  - How to access and utilize better and more data, including third party data
  - Data workflows and processes with a view to streamlining and improving them
  - How to improve data integrity
  - Using data science and big data analytics to strengthen risk assessment
  - Leveraging the value of country-by-country reports
- Mission observations:
  - STX impressed by rapid adoption of tools and methods from November mission and high engagement of dedicated staff.
  - SFA progressing on: a) emphasizing taxpayer size when determining risk; b) introducing empirically founded business risk rules when determining audits; c) building advanced predictive models.
  - Recommendation to introduce empirical tools gradually, starting with lowest-hanging fruits (simple analytics and weighting by firm size) before full deployment of advanced machine-learning models.

### II. Summary of advice given on November 2022 mission
- Foundational strengths identified:
  - Highly competent “Analysis unit” and timely investments in data cleaning, manipulation and storage.
  - High-quality data on audits, tax returns, Country-by-Country-Reports, business risk rules and balance sheets, all mergeable using unique taxpayer identifiers.
- Priorities and practical steps:
  - Utilize existing data sources more systematically to automatically identify risks via data-driven models.
  - Business risk indicators set up for transfer pricing were accurate in predicting yield of CIT comprehensive audits—opportunity to reuse existing work for CIT audit selection and for risk profiles of largest taxpayers.
  - Previous audit strategies/business risk rules were not being systematically evaluated; recommend simple historical evaluations (e.g., scatterplots).
  - Use simple benchmarks and standard deviations (industry profitability, standard deviations) to root business risk in systematic analysis rather than intuition.
  - Introduce thresholds in measure of strike rates so marginal corrections (e.g., a correction of €20) do not contaminate success measures; strike rates should reflect meaningful adjustments.
  - The 30 largest taxpayers are always a priority and should be continuously observed; create automated risk profiles for the 30 largest taxpayers and establish one-to-one client account managers.
- Pilot models constructed (data/code/workflows handed over to SFA):
  - An advanced (random forest) CIT audit selection model (incl. deployment workflow)
  - An advanced (random forest) VAT audit selection model
  - Workflow to evaluate existing business risk rules against historical data (VAT/CIT)
  - Country-by-Country-Reporting anomaly detection tool
- Model objectives and results:
  - All models targeted real-time risk assessment using all available data.
  - Demonstration: the CIT audit selection model could increase strike rates by 20-40 pct.-pts.
- Recommended priorities for the coming year (in order):
  - Create risk profiles for the 30 largest taxpayers.
  - Create business risk indicators for CIT.
  - Put larger weight on firm size and redefine strike rates to reflect firm size.
  - Evaluate existing business risk rules using historical data.
  - After these adjustments, proceed with more advanced modelling.

### III. Summary of technical assistance advice (workshop findings and SFA plans)
- Background and SFA implementation plan:
  - SFA requested follow-up work one month after initial assignment, indicating commitment.
  - SFA short-term plan for 2023:
    - Emphasize large taxpayers in CIT; implement existing and new strategies for selection of taxpayers for CIT supervision using internal and external data.
    - Use known tools (e.g., SAP Fraud Management) and include CIT risk indicators recently identified; test predictive model in practice on a smaller scale.
    - Adapt annual supervision plan of the Special Financial Office (SFO); comprehensive audits of all taxes planned by teams including transfer pricing specialists.
    - Use automatic risk analysis in transfer pricing combined with predictive analytics and current case selection processes.
    - Use questionnaires, preventive visits, and cooperative compliance approaches to promote voluntary compliance for large taxpayers.
  - STX endorsed incremental approach from simple to advanced empirical analysis.
- Workshop outcomes and model performance:
  - SFA built upon the Random Forest model from November 2022 using KNIME®.
  - Advanced model substantially outperformed current audit selection models for CIT:
    - Increased predicted strike rate (measured as corrections larger than $5 thousand) from 30 to 70 percent.
- Areas for further model improvement identified by mission team:
  - Separate models should be developed for comprehensive audits and partial audits; also separate models for each type of partial audit (e.g., interest deduction audits, desk audits). Models must use comparable outcome/intervention in training data.
  - Model training currently uses corrected tax returns (post-audit) rather than initial returns; audit selection in practice is based on initial returns, so feeding corrected returns hampers prediction of non-compliance.
- Additional recommendations:
  - Proceed with simpler risk analyses before deploying most advanced models; test simple CIT risk rules against past audit results using KNIME® workflows.
  - Use RCTs and “double-blinded horse races” for model evaluation during gradual implementation.
  - Build advanced risk assessment models for Customs and VAT first (abundance of cases makes them easier to predict).
  - Technical correspondence on AI/Random Forest and KNIME® summarized in Appendix I.

### IV. Proposed next steps (same as in November 2022 mission)
- Prioritization / senior management endorsement (Short term: within next 12 months)
  - Prioritize/endorse high-value data analytics work; create protected time for development and deprioritize low-value tasks.
  - Move towards “top-down” risk assessments in CIT by defining and using existing business risk indicators.
  - Increase emphasis on taxpayer size in risk assessments: risk profile the 30 largest taxpayers; create client owners; increase threshold for “strikes”; weigh risk rules according to consequence.
  - Endorse multifaceted compliance risk management (audits plus service/education).
  - Move towards automatic exchange of third-party information from banks, regulators and other stakeholders (Medium term: within two to three years) — may require legislative action.
- Improvements in data analytics and way-of-work (Short term: within next 12 months)
  - Readjust risk rules to reflect expected adjustment/risk by multiplying current risk rules with measures of size (turnover, tax paid, employees).
  - Build automated reports for the largest taxpayers (Analysis unit and large taxpayer unit); reports to indicate suggestive actions such as “contact taxpayer” or “audit”.
  - Build CIT business risk indicators (begin with current TP business risk rules).
  - Evaluate existing risk rules based on historical data and advanced models; update accordingly (see KNIME workflow).
  - Expand use of anomaly detection in transfer pricing using CbCR-data + other data sources; inform which of the >2,300 cases of firms outside safe-harbor thin capitalization rules to pursue; prioritize by taxpayer size.
  - Advanced data analytics projects after implementing low-cost investments (1-5), e.g., random forest models; use pilot models from mission as inspiration.
  - Evaluate implementation of data analytics projects and devise remedying actions if necessary; anticipate and address resistance to change.
- Organizational structure (Short term: within next 12 months)
  - Ensure the Analysis unit supports the entire SFA (potentially by referring directly to the DG or as part of a CRMU).
  - Reestablish the Compliance Risk Management Unit.
- Staffing and training (Short term: within next 12 months)
  - Prioritize end-user education in all existing applications; no application should be built without prioritizing end-user education and feedback.
  - Establish learning programs where data scientists teach non-experts in data analytics to alleviate Analysis unit workload and spread capabilities; KNIME® suggested as a low-barrier training tool.
  - Consider improving terms for data analysts to ensure recruitment and retention (longer contracts, time-conditioned compensation “Golden handcuffs”).
- Implementation timelines and responsibilities
  - Multiple actions assigned deadlines of “Short term (within next 12 months)” and one action assigned “Medium term (within two to three years)”.
  - Responsibilities allocated to Director General (DG)/Heads of department, Analysis unit, relevant business owners, IT, HR, Working Groups (WG), large taxpayer unit, and relevant stakeholders as specified in Table 1.

*TECHNICAL ASSISTANCE REPORT — Republic of Slovenia — Data-Driven Risk Assessment (Follow-Up) — February 2022 — Prepared By Ludvig Wier*

### Section 2

### tarea2024011 - Section 2

### Retrain 1-3 non-data scientists from SFA to join the Analysis unit
- Time horizon: Short term (within next 12 months)
- Responsible parties: Data Science Team and HR
- Purpose:
  - The intention is two-fold: 1) bring business knowledge into the Data Science Team 2) expand the capacity of the Data Science Team.
- Implementation note:
  - One possibility is to exploit the KNIME® software and available courses to ease this process.

### Start “tours” program where non-data analysts from SFA are embedded in the Analysis unit and vice versa
- Time horizon: Medium term (within two to three years)
- Responsible parties: DG and relevant heads of division in consultation with Analysis units
- Implementation guidance:
  - Make sure this aligns with the projects in the pipeline – e.g. if a customs selection project is taking off, then it make sense to collaborate with customs.

*Source: tarea2024011 - Section 2*

---


_Source: https://www.imf.org/-/media/files/publications/tar/2024/english/tarea2024011.pdf_
