## Annex I. Main Analytical Tools for Cost Accounting (wpiea2025159-source-pdf)

## Source details

**Canonical URL:** [Annex I. Main Analytical Tools for Cost Accounting (wpiea2025159-source-pdf)](https://www.imf.org/-/media/files/publications/wp/2025/english/wpiea2025159-source-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2025/english/wpiea2025159-source-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2025/english/wpiea2025159-source-pdf.pdf.json)

---

### Machine learning for anomaly detection (Box 8)
- Overview and rationale
  - ML models can detect anomalies, reveal cost patterns, and uncover signals difficult to identify manually in cost accounting datasets.
  - ML applications in Public Financial Management (PFM) are nascent but gaining traction, including in low-income countries as digital capacity grows.
  - ML can support detection and prevention of mismanagement and corruption, and make cost accounting data more actionable for managers and policymakers.
- Implementation considerations
  - ML is a complementary analytical layer built on structured cost accounting data; design data structures with future analytical use in mind.
  - Prerequisites: suitably structured, reconciled, and interoperable cost accounting data; increasing digital capacity; interoperable data schemas.
  - ML complements BI and other analytics by adapting and predicting rather than only visualizing.
- Limitations and contextual notes
  - Usefulness depends on data quality; nascent maturity across countries; poor data can lead to false alarms or missed anomalies.
- Illustrative elementary school infrastructure example
  - Dataset: synthetically augmented; "The original dataset was augmented using Generative AI."
  - Unsupervised algorithm chosen: Isolation Forest (preferred for efficiency and lower tuning requirements versus One-Class SVM).
  - Testing results: out of 10,000 records, 20 were identified as potential anomalies.
  - Anomalous vs normal characteristics:
    - Average student body: around 39 students (anomalous) versus average of 1,259 students (normal).
    - Total costs: averaging approximately 6.95 million (anomalous) compared to the normal average of 4.50 million.
  - Interpretation: much lower student numbers combined with disproportionately higher costs mark distinct outliers.
- Semi-supervised pipeline and supervised refinement
  - Steps: (1) unsupervised detection, (2) expert investigation/validation, (3) create labeled dataset, (4) train supervised models (Random Forest, XGBoost, K-Nearest Neighbors).
  - Model evaluation: confusion matrix, cross-validation, active learning, semi-supervised ensembles, XAI methods (SHAP, LIME).
- Unsupervised algorithm options
  - Isolation Forest; DBSCAN; One-Class SVM.
- Implementation checklist
  - Feature selection; hyperparameter tuning; cross-validation; rigorous data validation and preprocessing.

### Earned value management indicators (Box 1)
- Earned value formulas
  - Cost variance = budgeted cost for work performed (BCWP) - actual cost of work performed (ACWP)
  - Schedule variance = BCWP - budgeted cost for work schedule (BCWS)
  - Variance at completion = budget at completion (BAC) - estimated cost at completion (EAC)
  - Cost performance index (CPI) = BCWP / ACWP
  - To complete performance index (TCPI) = (BAC - BCWP) / (EAC - ACWP)
- Uses and diagnostic interpretation
  - Enables integration and monitoring of cost, schedule, and performance to ensure execution follows plan, mitigate risks, identify root causes, and support corrective measures.
  - Examples:
    - Low CPI may indicate cost inefficiency, unanticipated travel, expensive labor, or technical issues.
    - CPI lower than TCPI may indicate overly optimistic future productivity assumptions at budget approval.
- Additional indicators signaling credibility problems
  - Large and frequent cost or schedule variances.
  - Significant difference between estimated cost to complete and budget for remaining work.
  - Insufficient budgets, including management or contingency reserve, for remaining work.
  - Recurring data errors or anomalies (e.g., negative values for actual costs).
- Role of cost accounting and transparency in budget credibility
  - Requirements for credible estimates: continuous stakeholder participation; clear scope and lifecycle understanding; access to reliable historical data; robust risk and sensitivity assessment; sufficient validation; detailed documentation.
  - Evidence cited: 24 percent increase in budget transparency scores since 2008 (IBP, 2023); 15 OECD countries (48 percent) indicated improving allocation of resources was a main driver for performance budgeting; only six (21 percent) consider high effectiveness in informing allocation (OECD, 2018).
- Recommendations to improve PBB relevance
  - Drill down on strategic indicators; develop digital solutions for granular data while keeping strategic view; include cost data in performance indicators (e.g., outputs divided by cost of inputs).

### Integrating cost accounting into performance-based budgeting (Box 2)
- Uses of cost data
  - Inform allocations, pricing, program decisions, and user fees.
  - Preconditions: accounting systems must record expenditure by program as well as by economic and administrative classifications.
  - Cost analysis enables identification of cost drivers and cost-saving opportunities; can inform expenditure baselines.
- Technology and algorithmic support
  - Algorithms can analyze variations in cost allocations, identify sudden unexplained changes, detect insufficient documentation, flag disproportionate cost variances, and detect high overheads without justification.
  - Implementation note (footnote 17): illustrative algorithm executed in Python; connection options to FMIS include "pymysql", "PyMongo", APIs, or manual CSV updates.
- Performance measurement and spending reviews
  - Cost accounting allows comparison of actual costs with budgeted costs to identify variances and conduct spending reviews.
  - Example trigger: consistent underperformance despite high costs, suggesting further evaluation.
- Procurement and contract management support
  - Cost accounting helps estimate total cost of ownership, support performance-based contracts, monitor supplier performance, and assess procurement risks.
  - Illustrative examples:
    - Denmark: TCO calculation for 40 municipalities added three-year power consumption costs; total estimated savings for the three-year period were almost 11 million DKK (about 1.6 million USD), 7,250 MWh, and 3,625 tons of CO2.
- Detection of corruption and mismanagement
  - Cost accounting enables benchmarking provider costs and identifying unexplained outliers potentially linked to corruption.
  - Algorithms can highlight deviations in pricing, quantities, purchase orders, and cash-flow patterns.

### Red flags and procurement anomaly detection (Box 4)
- Scale and transparency context
  - Public contracts constitute one-third of government spending, totaling about 13 trillion USD per year globally.
  - Less than three percent of public contracts are published openly.
  - In OECD member countries: public procurement spending estimated at 6 trillion EUR (about 7 trillion USD) per year, representing 12 percent of GDP and 29 percent of government expenditure.
- How analytics and ML can help
  - ML and cost accounting data can enable real-time detection of irregularities, improve accountability, monitor inventory and assets, detect accrual manipulation and unusual activities, and analyze trends for fraud signals.
  - Data limitations affect implementation: four out of 43 identified flags were implemented in Colombia due to data limitations, while 28 were implemented in Paraguay.
- Illustrative red-flag indicators
  - Line-item bids are unreasonably high / low
  - Tender value is higher or lower than the category average
  - Agents charge excessive fees or overcharge for the work performed
  - Small initial purchase from supplier is followed by much larger purchases
  - Awarded price is higher than the price of the input under open contract
  - Margins on a particular item or category of items is consistently low or high
  - There is a large difference between contract award and final contract amount
  - Payments of unjustified high prices relative to historical average are made
  - Total payments to a contractor exceed total contract or purchase order amounts
- Interpretation caveats
  - Anomalies may have benign explanations (e.g., scope changes, limited competition); further investigation is required to detect corruption, defined as abuse of public office for private gain (IMF, 2024).
- Alternative tools
  - Corruption Cost Tracker calculates a Corruption Risk Index based on frequency of seven red flags (single bidder; non-open procedures; lack of publication of call for tenders; long/short submission periods; concentration in bidders; share of suppliers with limited transparency).

### Institutional integration, analytical tools, and implementation priorities (Section 1 & Annex I)
- Policy recommendations and institutional priorities
  - 1. Institutional Integration: Embed cost accounting into core PFM processes (budget formulation, procurement planning, performance reviews); requires clear mandates, cross-sector coordination, and sustained leadership.
  - 2. Capacity Building: Invest in training for technical staff and decision-makers to ensure cost data is understood and used.
  - 3. Iterative Implementation: Achieve early wins (e.g., detecting anomalies, improving procurement efficiency) to build support.
- Main analytical tools for cost accounting (Annex I)
  - Descriptive statistics: summarize distributions, detect outliers; computationally inexpensive but limited in complexity.
  - BI dashboards: interactive KPI monitoring; require setup and may not be real-time.
  - Process Mining: discover and optimize processes from event logs; requires specific event log formats.
  - Linear and non-linear models: for prediction and modeling relationships; non-linear models capture complex effects but risk overfitting.
  - ML: predictive analytics, anomaly detection, optimization; requires considerable data and preprocessing.
- Advantages and disadvantages (method-level highlights)
  - Descriptive statistics: easy initial use; limited to simple insights.
  - BI dashboards: decision-friendly visuals; may not prescribe actions.
  - Process Mining: identifies bottlenecks; interpretation requires expertise.
  - Linear models: simple interpretability; limited to linear relationships.
  - Non-linear models: capture complex dynamics; more complex to interpret.
  - ML: high accuracy and adaptability; interpretation and preprocessing can be intensive.
- Types of ML algorithms (Annex II)
  - Supervised: task-driven; requires labeled data; applications include classification and regression.
  - Unsupervised: data-driven; useful for exploration, clustering, anomaly detection.
  - Semi-supervised: hybrid; leverages labeled and unlabeled data.
  - Reinforcement: environment-driven; generally not preferable for straightforward cost accounting problems.

### Data structures, sample data dictionary, and interoperability (Annex III & Section 3)
- Establish common identifiers to link records
  - Organizational identifiers; Program/project codes; Service IDs; Time periods (standardized fiscal year or monthly timestamps).
- Integrate data using data warehouse or data lake: extraction, transformation, loading, and clear relationships (e.g., Program_ID, Region_ID, Time_Period).
- Sample data dictionary core variables (selected examples; exact types and values preserved)
  - Year / String / "2023"
  - ID_Ministry / String / "4"
  - Ministry / String / "Ministry of Education"
  - ID_School / Integer / 12345
  - School / String / "School_567"
  - ID_District / Integer / 78901
  - District / String / "District_78901"
  - School_size / Integer / 750
  - Number_of_students / Integer / 1500
  - Students_per_room / Float / 150.0
  - Type_of_cost / String / "Fence installation - service"
  - Total_costs / Integer / 4500000
  - Cost_per_student / Float / 3000.0
- Recommended columns for anomaly detection: ID_School; Type of cost; School characteristics; ID_District; Economic classification; Functional classification (COFOG).
- Interoperability considerations for FMIS and administrative records
  - Benefits: link expenditure to outcomes; usability concerns require clear naming conventions and documentation.
  - Performance and security: mitigate latency; implement robust authentication and authorization.
  - Data privacy: handle sensitive personal data carefully; ensure compliance with laws and regulations.

### Isolation Forest technical details and practical steps (Annex IV)
- Core concept: isolate anomalies via recursive partitioning; anomalies tend to have shorter path lengths in isolation trees.
- Algorithm steps (technical)
  1. Tree Construction: randomly select feature and split value between min and max; repeat until isolation or depth limit.
  2. Path Length Calculation: ℎ(푥) = number of edges from root to external node; anomalies have shorter ℎ(푥).
  3. Anomaly Score Calculation: s(x,n) = 2^( - E(h(x)) / c(n) ) using
     - E(h(x)) = average path length of x across forest
     - c(n) ≈ 2( ln( n−1 ) + γ ) − 2( n−1 ) / n, with γ ≈ 0.5772 (Euler–Mascheroni constant)
     - Normalization makes scores comparable across dataset sizes; anomalies near 1, normal points near 0.5 or below.
  4. Model Training and Prediction: unsupervised on entire dataset; assigns anomaly scores and classifies short-path points as anomalies.
  5. Parameter Tuning: n_estimators; contamination; max_samples; max_features.
  6. Labeling for supervised learning: use regulatory flags, expert labeling with consensus; train Random Forest, XGBoost, or K-Nearest Neighbors.
- Example implementation note
  - Illustrative Python example using pandas and sklearn.ensemble.IsolationForest with contamination=0.002 to predict anomalies and export results.

### Conclusions, applicability, and recommended next steps
- Key findings
  - Cost accounting can improve budget credibility, PBB, public procurement, and detection of corruption and mismanagement when properly integrated.
  - Barriers: fragmented data systems, unclear institutional mandates, perceptions of resource intensity.
  - Technical enablers: interoperability, automation, standardized data schema to enable real-time flows without major system overhauls.
  - Advanced analytics, including ML, expand the utility of cost accounting for anomaly detection, pattern discovery, and predictive analysis.
  - Semi-supervised learning is promising where labeled data are scarce or biased.
- Applicability and next steps for implementers
  - Align technical solutions with institutional realities.
  - Assess existing data infrastructure, identify interoperability opportunities, and map cost centers and service lines to administrative records.
  - Pilot in sectors with relatively structured data to demonstrate value and build momentum.
  - Leverage digital infrastructure and prioritize data quality, common identifiers, and documentation for scalable ML applications.

*Source: wpiea2025159-source-pdf — Annex I. Main Analytical Tools for Cost Accounting (selected boxes and annexes).*

### Annex I. Main Analytical Tools for Cost Accounting .....................................................................

### Annex I. Main Analytical Tools for Cost Accounting

### Contents overview
- Annex I. Main Analytical Tools for Cost Accounting ..................................................................................... 30
- Annex II. Types of Machine Learning Algorithms Considered ..................................................................... 32
- Annex III. Sample Data Dictionary of Costs for Anomaly Detection ............................................................ 33
- Annex IV. Technical Note on Isolation Forest for Anomaly Detection ......................................................... 35
- References ......................................................................................................................................................... 37

### Listed boxes and focal points
- Box 1. Indicators for variance analysis in earned value management ................................................................ 10
- Box 2. Advantages and challenges of integrating cost accounting into PBB ...................................................... 14
- Box 3. Use of cost accounting in public procurement ......................................................................................... 17
- Box 4. Example of possible red flags in public procurement ............................................................................... 18
- Box 5. Interoperability mechanisms to collect cost accounting data ................................................................... 21
- Box 6. Schema to standardize cost centers and services with different levels of data granularity ..................... 22
- Box 7. Linking administrative records to measure performance under PBB ....................................................... 23

*Source: wpiea2025159-source-pdf - Annex I. Main Analytical Tools for Cost Accounting*

### Box 8. Using unsupervised ML to detect anomalies in cost accounting data ...............................................

### Box 8. Using unsupervised ML to detect anomalies in cost accounting data

### Overview
- Machine learning (ML) models are explored as tools to help detect anomalies, reveal cost patterns, and uncover signals that are difficult to identify manually in cost accounting datasets.
- ML applications in Public Financial Management (PFM) remain nascent but are gaining traction, including in low-income countries as digital capacity grows.
- Figure referenced: "Data distribution (anomalies vs rest of data)" illustrating the distinction between anomalous observations and the remainder of the dataset.

### Rationale and intended benefits
- ML can bridge the gap between existing cost data and actionable insights when traditional methods fail to extract complex patterns.
- Potential uses highlighted:
  - Detection and prevention of mismanagement and corruption by revealing anomalies in cost data.
  - Support for data-driven decision-making by uncovering cost patterns not evident through manual review.
  - Complementing cost accounting efforts to make data more actionable for managers and policymakers.

### Implementation considerations (as discussed)
- ML is presented as a complementary analytical layer built on top of structured cost accounting data; designing data structures with future analytical use in mind is encouraged.
- Increasing digital capacity and interoperable data schemas are implied prerequisites to effectively deploy ML methods in PFM contexts.
- The text notes that ML applications are part of a broader push to make cost data more timely, accessible, and actionable.

### Limitations and contextual notes
- ML applications in PFM are described as nascent, suggesting limited maturity and experience across countries.
- The usefulness of ML depends on the availability of suitably structured, reconciled, and interoperable cost accounting data.

*Source: IMF Working Paper — Box 8. Using unsupervised ML to detect anomalies in cost accounting data (from wpiea2025159-source-pdf).*

### Box 1. Indicators for variance analysis in earned value management

### Box 1. Indicators for variance analysis in earned value management

### Earned value measures and formulas
- Cost variance = budgeted cost for work performed (BCWP) - actual cost of work performed (ACWP)
- Schedule variance = BCWP - budgeted cost for work schedule (BCWS)
- Variance at completion = budget at completion (BAC) - estimated cost at completion (EAC)
- Cost performance index (CPI) = BCWP / ACWP
- To complete performance index (TCPI) = (BAC - BCWP) / (EAC - ACWP)

### Uses and diagnostic interpretation
- Integrating and monitoring cost, schedule, and performance data through earned value management enables comprehensive variance analyses that:
  - Help ensure execution follows original plan and budget
  - Mitigate risks
  - Identify root causes and impact of deviations
  - Support application of corrective measures
- Measures may be disaggregated at the level of cost elements, plotted for trends, compared to each other or against a threshold, and investigated further to identify causes and test assumptions.
- Diagnostic example:
  - Low CPI may indicate cost inefficiency, unanticipated travel, expensive labor, or technical issues; managers should assess full meaning to apply corrective measures.
  - CPI that is lower than TCPI may indicate that higher productivity was assumed for the future at the time of budget approval and may need to be revisited.

### Additional indicators that signal data, execution, or budget credibility problems
- Large and frequent cost or schedule variances (either at program or cost element level)
- Significant difference between estimated cost to complete and budget for remaining work
- Insufficient budgets, including management or contingency reserve, for the remaining work
- Program status and schedule that are not comparable to historical data or original plan
- Logic sequence and durations for forecasted work that vary significantly from plan
- Recurring data errors or anomalies (e.g., negative values for actual costs)

### Role of cost accounting, data, and transparency in budget credibility
- Producing credible cost estimates requires:
  - continuous participation of diverse stakeholders
  - clear understanding of scope, project life cycle, and all relevant cost elements
  - access to reliable historical data that can serve as the basis
  - robust risk and sensitivity assessment
  - sufficient validation
  - detailed documentation (Government of Canada, 2023)
- Quantification of cost risks and uncertainties is crucial but often inadequate (U.S. GAO, 2020); risk and uncertainty analysis can:
  - predict probability of achieving the estimate and executing within budget
  - inform contingency levels to provide acceptable confidence
  - sensitivity analysis can show potential effects of changing assumptions
- Governments can adopt practices to improve credibility of cost estimates and budgets:
  - require updates of risk and uncertainty analysis prior to budget submissions and at other key decision points
  - use past outcomes, similar initiatives, and benchmarks to validate cost estimates
  - improve availability and quality of relevant data (historical and current costs, program schedule, data required for normalization such as inflation rates)
  - ensure budgetary financial data can be used for estimating costs through standardizing necessary accounting adjustments and defining appropriate level of detail
- Open data and documentation improve credibility and help minimize principal-agent problems by enabling verification and public scrutiny
- Documentation should be done throughout the cost accounting process, capturing ground rules and assumptions, methods, data sources, and other key details to ensure data can be understood, traced, audited, replicated, or updated

### Links to performance-based budgeting (PBB) and cost accounting
- PBB defined as the “systematic use of performance information to inform budget decisions” (OECD, 2023)
- Evidence cited:
  - 24 percent increase in budget transparency scores since 2008 (IBP, 2023)
  - 15 OECD countries (48 percent) indicated improving allocation of resources was a main driver for performance budgeting; only six (21 percent) consider a high effectiveness in informing allocation and prioritization of resources (OECD, 2018)
- Challenges in PBB:
  - notable lack of application of increased performance information in decision-making
  - scarcity of relevant performance indicators that enable comparison across programs for budget decisions
  - many systems compile indicators in aggregate forms, missing actionable individual-level data
- Recommendations for enhancing PBB relevance:
  - allow drilling down on strategic indicators to provide more detailed, attributable measures (example: Australia revised outcome language to add sub-levels for attribution)
  - develop digital solutions that allow drilling down to granular data while maintaining strategic view
  - include cost data in performance indicators (e.g., outputs divided by cost of inputs) to assess financial implications and trade-offs

### Advantages and challenges of integrating cost accounting into PBB
- Advantages:
  - Simple and practical information; helps simplify presentation of complex interactions (e.g., activity-based costing)
  - Allocation of indirect costs to programs and products with reasonably accurate mechanisms
  - Cross-sectional cost benchmarking across agencies, programs, or external entities
  - Linkage between funding and results; useful for medium-term expenditure estimates, informed performance targets, and formula funding
- Challenges:
  - Oversimplified unit costs: production level affects costs; solution is using complete cost function or average variable costs
  - Differentiated costs by geographical location (e.g., higher costs in rural schools and hospitals)
  - Difficulties measuring outcomes vs. outputs; cost-based formulas for budget may distort results
  - External factors and contingency actions introduce complexity (e.g., fire services cannot predict demand ex ante)

### Methodological notes and contributions
- Different methodologies to include costs in performance evaluation:
  - acknowledge tradeoffs between service and costs
  - emphasize reduction of costs potentially without focus on consequences
  - compare actual costs with standard costs to identify variations worth understanding
- Cost accounting can contribute primarily to three objectives through cost attribution:
  - enable identification and allocation of costs to specific activities, programs, or projects
  - assess actual expenses associated with different performance measures or outcomes
  - support functions of performance measurement including evaluate, control, budget, motivate, promote, celebrate, learn, and improve (Behn, 2003)

*Source: Authors, based mainly on U.S. GAO (2020).*

### Box 2. Advantages and challenges of integrating cost accounting into PBB

### Box 2. Advantages and challenges of integrating cost accounting into PBB

### Cost data for allocations, pricing, and program decisions
- Cost accounting provides insights into the cost structure and resource requirements for different programs, allowing governments to allocate resources based on the expected costs of achieving desired outcomes.
- Costs are important for deciding on user prices and setting fees for the provision of goods and services (Rogošić, 2021).
- Preconditions: the accounting system needs to “record expenditure on a continuing basis by program (as well as by the established economic and administrative classifications)” (Robinson & Last, 2009, p. 6).
- A purchaser-provider system approach—where line ministries are paid prices for the services they deliver—can determine costs but has been successful only in small, controlled sectors such as hospitals.
- Cost-based analysis and budgeting may be challenging for complex activities with mixed inputs; allocations based on pre-defined formulas are normally useful to compare programs that produce similar inputs in controlled environments.
- Cost analysis enables detailed examination of cost drivers, identifying areas of inefficiency and cost-saving opportunities, and by linking costs to performance outcomes governments can assess program cost-effectiveness and define expenditure baselines, noting that cost may change even if policies remain constant (Rahim et al., 2022).

### Technology and algorithms to improve cost allocation
- Introduction of technology can enhance the accuracy and transparency of cost allocation processes.
- An algorithm could analyze the variations in cost allocations for different cost categories over time using raw data, preferably integrated into an Integrated Financial Management Information System (FMIS).17
- Algorithms can identify sudden or unexplained changes that exceed a predetermined threshold, which may indicate errors, manipulation, or misreporting in cost attribution.
- Algorithms can detect lack of documentation or inadequate evidence for cost allocations by assessing the availability and quality of supporting documentation and issuing warnings when evidence is insufficient.
- By comparing costs for similar programs, algorithms can identify disproportionate cost variances and issue warnings when costs are significantly higher or lower compared to similar programs, suggesting potential inefficiencies or anomalies.
- Algorithms can detect high levels of overhead costs without clear justifications or impact on performance and help implement cost-effectiveness evaluations.

- Footnote 17 (from source): As explained further in Chapter IV, the algorithm that we propose in this paper is executed in Python. There can be different ways in which Phyton can connect to the FMIS, including the installation of specific packages to connect with SQL and not SQL databases (including ”pymysql” or "PyMongo", among others), connect to an API, publish the data in a URL and connect to that data periodically or even manually update a local csv file and use it to run the algorithm in Python.

### Performance measurement, variance analysis, and spending reviews
- Cost accounting provides a foundation for measuring and monitoring performance against budgeted targets by enabling comparison of actual costs with budgeted costs and identifying variances for corrective action.
- Integrating cost accounting data with performance indicators allows governments to evaluate efficiency and effectiveness of programs and conduct spending reviews—especially program reviews—focused on delivering efficiency savings (Doherty & Sayegh, 2022).
- Example indicator: consistent underperformance or failure to achieve targeted performance outcomes despite high costs; lack of correlation between cost increases and improvements in performance outcomes can trigger more detailed impact evaluation.18

- Footnote 18 (from source): A challenge that could arise for performance measurement is that costs might be calculated in different units than the level of granularity for budgets. While costs are normally calculated at the level of the unit that provides the services (e.g., schools, hospitals), budgets are normally prepared at the services level. A consolidation would be required, which can be addressed through the digital tools presented in the following section of this Working Paper.

### Support to public procurement
- Cost accounting can enhance procurement decisions, cost control, and contract management across the procurement cycle—from planning and tender to award, contract, implementation, and evaluation.
- Cost estimates can serve as the basis for contract budgets, evaluations, and negotiations and help develop more realistic and reliable budgets for goods, services, or infrastructure projects.
- Cost accounting helps evaluate the total cost of ownership (including maintenance, operation, and disposal) to compare procurement options considering life-cycle costs, quality, and sustainability.
- Cost accounting facilitates performance-based contracts by integrating performance metrics and costs, enabling payment linkages to achievement of agreed targets and supporting development of measurable, realistic, and data-driven rates and deliverables.19
- Cost accounting data can aid in evaluating and monitoring supplier performance by tracking actual costs against contracted costs to identify cost overruns, variances, or discrepancies.
- Cost accounting contributes to procurement risk assessment and mitigation by identifying cost-related risks such as cost escalation, budget overruns, or inadequate cost controls.
- Cost accounting helps detect anomalies in procurement by comparing actual costs to market benchmarks, historical costs, or budgeted costs and disaggregating data to indicate origins of deviations (e.g., specific cost elements).

- Footnote 19 (from source): Governments should carefully assess whether it would be appropriate and take into consideration any risks involved. For example, new programs may not have the historical data to inform the negotiation of objective and measurable performance targets. In such cases, governments could consider a hybrid model, to provide time to collect the necessary data. As suppliers also risk incurring non-reimbursable indirect costs (e.g., costs associated with tracking deliverable), governments may need to provide additional assistance to ensure compliance. For more information, see Nonprofit Resiliency Committee (2019).

- Illustrative Box 3 examples (from source):
  - Canada: cost accounting guides price negotiation and helps determine contract costs when market-based pricing is not possible.
  - Denmark: total cost of ownership (TCO) calculation for 40 municipalities added three-year power consumption costs; total estimated savings for the three-year period were almost 11 million DKK (about 1.6 million USD), 7,250 MWh, and 3,625 tons of CO2.
  - United States: U.S. Government Accountability Office (GAO) reviews of the Department of Defense (DOD)’s F-35 program have supported actions to reduce cost overruns and improve credibility of cost estimates, including independent cost estimates and automated tools providing real-time access to software development metrics.

### Detection of corruption and mismanagement
- Cost accounting offers relevant insights for detecting and deterring various forms of corruption and mismanagement that affect PFM.21
- Corruption can take multiple forms (bureaucratic or political, cost-reducing or benefit-enhancing, coercive or collusive) and methods (bribery, extortion, cronyism, embezzlement, fraud22), and it can influence budget choices and execution (e.g., diversion to activities offering kickbacks, exploitation of procurement weaknesses, off-budget spending).
- Cost accounting data can help define indicators to detect possible cases of corruption by enabling calculation of the costs of producing a given good or service to serve as benchmarks for comparing providers or activities and identifying unexplained cost outliers that may be linked to corruption (Box 4).
- Cost accounting can support anomaly detection by highlighting unexplained cost savings or outliers, deviations in pricing, quantities, and purchase-order information, and unusual cash-flow patterns; disaggregated analysis can indicate origins of deviations and inform targeted investigations or interventions.

- Footnotes from source:
  - 20 (referenced in procurement section) indicates that variance trends and patterns may require further investigation to inform decisions and interventions.
  - 21 and 22 (referenced in corruption section) annotate the relevance and varied manifestations of corruption in PFM.

*Source: IMF Working Papers — Making Cost Data Work for Public Financial Management, Box 2.*

### Box 4. Example of possible red flags in public procurement

### Box 4. Example of possible red flags in public procurement

### Scale and transparency challenges in public procurement
- Public contracts constitute one-third of government spending, totaling about 13 trillion USD per year on a global scale.
- Less than three percent of public contracts are published openly—and key data needed for useful analyses (e.g., contract amount, start and end dates) are made available for an even smaller share.
- In OECD member countries, public procurement spending was estimated to be 6 trillion EUR (about 7 trillion USD) per year, representing 12 percent of GDP and 29 percent of government expenditure.
- Lack of transparency undermines competition and effective oversight, with potential dire consequences for the quality of public goods, works, and services acquired.

### How cost accounting and analytics can strengthen procurement integrity
- Advanced uses of cost accounting data can support real-time detection of irregularities and improve accountability and outcomes in public procurement.
- Several countries are experimenting or implementing this approach (examples cited: Ukraine, Paraguay, Dominican Republic, Guatemala).
- An alert system requires quality, real-time, and standardized data for calculations; improving data availability and quality through cost accounting enables implementation of red flags.
- ML algorithms can:
  - Identify disorganization through analysis of large datasets that detect inconsistencies or deviation patterns (e.g., significant variations in prices of similar goods or services).
  - Ensure transactions are properly recorded and accounted for, enhancing monitoring of inventory and assets.
  - Discover accrual manipulation and unusual activities such as unexpected growth in employee productivity or discretionary accruals (Aryal, 2022).
  - Analyze trends and deviations in expected financial outcomes to detect fraud (examples: analysis of peaks in wages to create red flags).
- Examples where data limitations affected implementation of flags: four out of 43 identified flags were implemented in Colombia due to data limitations, while 28 were implemented in Paraguay.

### Illustrative triggers and red-flag indicators for an alert system
- Line-item bids are unreasonably high / low
- Tender value is higher or lower than the category average
- Agents charge excessive fees or overcharge for the work performed
- Small initial purchase from supplier is followed by much larger purchases
- Awarded price is higher than the price of the input under open contract
- Margins on a particular item or category of items is consistently low or high
- There is a large difference between contract award and final contract amount
- Payments of unjustified high prices relative to historical average are made
- Total payments to a contractor exceed total contract or purchase order amounts

### Interpretation caveats and investigative context
- Further investigation may be necessary to detect corruption. Examples of possible alternative explanations:
  - An abnormal spike in labor costs could reflect an overly optimistic assumption about required labor, a change in project scope, or fraudulent activities (e.g., over-reporting or inflating of labor charges).
  - Selection of suppliers with poor past outcomes could indicate conflicts of interest or improper practices, but could also result from limited competition and lack of performance-based contracting.
- Corruption is defined as the abuse of public office for private gain (IMF, 2024) and has large economic consequences (IMF, 2016), reducing resources available for expenditures and undermining impact.
- Fraud is defined as any act, including omission or misrepresentation of facts, that aims to deceive and induce a party to act to their detriment (World Bank, 2010).
- Given the linkage between corruption/mismanagement and PFM, cost accounting can help detect and deter such issues though detection may require deeper investigation.

### Alternative and complementary tools
- In the absence of an alert system, countries could consider using the Corruption Cost Tracker (Basdevant & Fazekas, 2023), which calculates the Corruption Risk Index based on the frequency of seven red flags across all available contracts:
  - (i) single bidder
  - (ii) non-open procedures
  - (iii) lack of publication of call for tenders
  - (iv) and (v) long or short period for submission and decision
  - (vi) concentration in bidders
  - (vii) share of suppliers with limited transparency

### Data requirements and links to digital enablers
- Effective alert systems and red-flag calculations depend on quality, real-time, standardized data; cost accounting improves availability and quality of such data.
- Interoperability with administrative records and standardized protocols are key to generating the data needed for procurement analytics (see subsequent discussion on APIs, schemas, and administrative system linkage).
- Improvements in data standardization and linkage also enable automated calculation of performance indicators for PBB and broader PFM uses.

*Source: Authors based on OCP (2016 & 2020), OCP blogs, RICG (2022), and Basdevant & Fazekas (2023).*

### 3. Establish common identifiers

### 3. Establish common identifiers

### Standard identifiers to link records
- Create a set of common identifiers across datasets to link records.
- Examples:
  - Organizational identifiers: Unique codes for ministries, departments, or agencies.
  - Program/project codes: Unique identifiers for programs or initiatives.
  - Service IDs: Unique codes for specific services delivered.
  - Time periods: Standardized fiscal year or monthly timestamps.

### Interoperability considerations for FMIS and administrative records
- Benefits: Interoperability of FMIS with administrative records of line ministries offers significant benefits.
- Usability and complexity:
  - Integration of various systems can create usability concerns, emphasizing the need for clear naming conventions and comprehensive documentation.
  - Developers should manage learning curves through proper training and resources.
- Performance and security:
  - Potential performance issues, such as latency, can be mitigated through optimized system design.
  - Robust authentication and authorization mechanisms should be put in place to ensure security and effective data protection during data exchanges.
- Data privacy:
  - Attention should be paid to data privacy, particularly in handling sensitive information such as personal data, across systems.
- Compliance:
  - Ensure effective and secure interoperability, as well as compliance with relevant laws and regulations.

### Integrate data (using data warehouse or data lake)
- Develop an integration platform to consolidate and link data:
  - Data extraction: Pull relevant data from administrative records.
  - Transformation: Standardize formats, clean data, and resolve inconsistencies.
  - Load: Store linked datasets in a central repository with clearly defined relationships.
- Example linkages:
  - Program_ID: To associate costs with specific programs.
  - Region_ID: To link expenditure and outcomes in geographic areas.
  - Time_Period: To align costs and performance metrics over time.

### Calculate performance metrics
- Create workflows or queries to compute metrics.
- Example formula as presented:
  - Cost efficiency = 
    Total service units delivered
    Total cost

### Machine learning to enhance uses of cost accounting data
- ML strengths:
  - ML can process large, complex datasets and uncover subtle patterns that traditional models may overlook.
  - ML adapts to evolving datasets, continuously improving its ability to identify irregularities, inconsistencies, or fraud.
- Business Intelligence vs ML:
  - Business Intelligence tools visualize and summarize cost data (dashboards and reports) but do not inherently detect anomalies or learn from data.
  - ML goes beyond visualization by adapting and predicting.
- Data quality precondition:
  - Ensure data quality through rigorous validation and quality assurance (cross-checking cost estimates against historical benchmarks, verifying data completeness, removing duplicates or misreported entries).
  - Poor data can lead ML models to generate false alarms or overlook genuine irregularities.

### ML model types and selection for cost accounting
- Four basic ML model types: supervised, unsupervised, semi-supervised, and reinforcement learning.
- Considerations in model selection:
  - Match model to data structure and objectives.
  - Ensemble methods (combine multiple models) identified as promising for prediction accuracy but can be computationally intensive and primarily geared toward prediction.
  - Neural networks are widely used but better suited for large, unstructured datasets; many management accounting datasets are structured and modest in size.
  - Supervised models require labeled data that define mismanagement or corruption risk in advance; labeling can introduce human bias.
  - Unsupervised models (clustering, classification) can detect anomalies without prior labeling and can be refined using supervised learning.
  - Semi-supervised learning combines unsupervised pseudo-labeling with supervised refinement; useful when labeled data is scarce or biased.
  - Reinforcement learning is less suitable for cost accounting, given its focus on sequences of actions.

### Illustrative example: elementary school infrastructure transfers program
- Dataset and augmentation:
  - Used a synthetically augmented dataset; "The original dataset was augmented using Generative AI."
  - Worked initially with an unlabeled dataset reflecting main variables related to identifying cost anomalies.
  - The creation of an ideal dataset avoided noise in training data; practical implementations will require careful preprocessing, feature selection, and hyperparameter tuning.
- Unsupervised algorithm choice:
  - Isolation Forest chosen for unsupervised anomaly detection due to advantages in handling large datasets efficiently and being less computationally intensive and requiring less parameter tuning than One-Class SVM.
  - Other unsupervised options mentioned: DBSCAN, One-Class SVM.
- Testing results:
  - The Isolation Forest algorithm was tested with the sample dataset.
  - Out of 10,000 records, 20 were identified as potential anomalies.
  - Characteristics of anomalous schools:
    - Average student body: around 39 students (anomalous) versus average of 1,259 students (normal).
    - Total costs: averaging approximately 6.95 million (anomalous) compared to the normal average of 4.50 million.
  - The combination of much lower student numbers and disproportionately higher costs marks them as distinct outliers.

### Semi-supervised pipeline and supervised refinement
- Semi-supervised pipeline steps:
  1. Unsupervised anomaly detection identifies suspicious observations.
  2. Domain experts investigate and validate these cases (e.g., audit teams).
  3. The results of investigations feed into a labeled dataset.
  4. Supervised learning algorithms are trained on this labeled data to generalize detection of future anomalies.
- Supervised models for classification after labeling: Random Forest, XGBoost, K-Nearest Neighbors.
- Model evaluation:
  - Use a confusion matrix to assess performance (true positives, false positives, etc.) against validated outcomes.
  - Employ cross-validation to gauge model performance on unseen data and prevent overfitting.
- Pipeline extensions and tools:
  - Active Learning: model queries for the most informative labels from auditors.
  - Semi-supervised ensemble methods: combine weakly-labeled and high-confidence samples for model training.
  - Explainable AI (XAI) methods such as SHAP or LIME to interpret model outputs and enhance institutional adoption.

### Box 8 — Unsupervised ML algorithm options and implementation considerations
- Algorithm options:
  - Isolation Forest: builds multiple decision trees or isolation trees with random splits to isolate anomalies.
  - DBSCAN (Density-Based Spatial Clustering of Applications with Noise): clusters densely packed data points; points outside dense regions are noise.
  - One-Class SVM: constructs a boundary around normal data points; deviations are anomalies.
- Implementation considerations and processes:
  - Feature selection: identify the most relevant variables for predictive power.
  - Hyperparameter tuning: adjust model parameters to enhance accuracy and efficiency.
  - Cross-validation: assess how well the model generalizes to independent datasets, reducing overfitting risk.

### Conclusion, applicability, and next steps
- Summary of paper’s findings:
  - Cost accounting can improve budget credibility, PBB, public procurement, and detection of corruption and mismanagement when properly integrated.
  - Cost accounting remains underutilized due to fragmented data systems, unclear institutional mandates, and perceptions of being resource-intensive.
  - Leveraging existing digital infrastructure—especially interoperability and automation—can overcome barriers.
  - A standardized data schema offers a practical solution for integrating cost accounting with administrative records and enabling real-time data flows without major system overhauls.
  - Advanced analytical tools, including ML, expand utility of cost accounting data for anomaly detection, hidden pattern discovery, and predictive analysis.
  - Semi-supervised learning is promising where labeled data is scarce or biased.
- Applicability and next steps:
  - Align technical solutions with institutional realities.
  - Begin by assessing existing data infrastructure, identifying opportunities for interoperability, and mapping cost centers and service lines to administrative records.
  - Pilot in sectors with relatively structured data to demonstrate value and build momentum.
  - Implementers should consider three key areas (not enumerated in the supplied excerpt but noted as a focus for implementers).

*Source: Authors.*

### 1. Institutional Integration: Embed cost accounting into core PFM processes, including budget

### 1. Institutional Integration: Embed cost accounting into core PFM processes, including budget

### Policy recommendations and institutional priorities
- 1. Institutional Integration: Embed cost accounting into core PFM processes, including budget formulation, procurement planning, and performance reviews. This requires clear mandates, cross-sectoral coordination, and sustained leadership.
- 2. Capacity Building: Invest in training for both technical staff and decision-makers to ensure that cost data is not only collected but also understood and used.
- 3. Iterative Implementation: Early wins—such as identifying cost anomalies or improving procurement efficiency—can help build support and justify further investment.
- Implementation emphasis:
  - Focus on practical integration, institutional ownership, and user-centered design.
  - Success depends on consistent use of cost data to inform decisions rather than on the sophistication of tools.
  - Objectives: improve public sector efficiency, accountability, and service delivery.

### Main analytical tools for cost accounting (Annex I)
- Descriptive statistics
  - Role: organize, summarize, and present data visually (measures of central tendency, variability, histograms, box plots, scatter plots).
  - Use in cost accounting: understand distribution of costs and identify unusual data points or trends.
- BI dashboards
  - Role: compile and display data interactively; enable real-time monitoring of KPIs.
  - Use in cost accounting: track costs, performance, revenues, and financial metrics for decision-makers.
- Process Mining
  - Role: use event log data to understand how existing processes operate.
  - Use in cost accounting: optimize processes for efficiency and compliance; identify bottlenecks and deviations.
- Linear models
  - Role: assume linear relationships for prediction and understanding (e.g., how a change in production volume could affect costs).
- Non-linear models
  - Role: model complex relationships such as economies of scale or diminishing returns.
- ML (Machine Learning)
  - Role: predictive analytics, anomaly detection, and optimization to enhance budget forecasting, identify potential fraud or errors, and optimize resource allocation.

### Advantages and disadvantages (Annex Table 1 – method-level highlights)
- Descriptive statistics
  - Advantages: helpful for initial approach; detect outliers; simple to implement periodically; computationally inexpensive.
  - Disadvantages: limited scope; cannot analyze complex relationships; insights may not be directly actionable; cannot be used straightforwardly for PFM decision-making.
- BI dashboards
  - Advantages: easy to read by decision-makers; allow comparison of raw data and complex indicators; easy to share and justify decisions with visuals.
  - Disadvantages: limited to presenting a clear course of action; require initial setup/expert knowledge; may not reflect real-time state depending on refresh cycles.
- Process Mining
  - Advantages: discovers, monitors, and improves actual processes from event logs; identifies bottlenecks and inefficiencies; useful for conformance checking and optimization.
  - Disadvantages: requires event log data in a specific format; interpretation can be difficult for non-experts; may raise privacy and security concerns.
- Linear models
  - Advantages: detect patterns via linear equations; helpful for prediction and comparison to observed results.
  - Disadvantages: limited to linear relationships; cannot detect unforeseen patterns; rely on assumptions (homoscedasticity, normality).
- Non-linear models
  - Advantages: capture more complex relationships; useful for prediction and comparison.
  - Disadvantages: based on theoretical approaches (cannot detect unforeseen patterns); more complex to model and interpret; risk of overfitting if not regularized.
- ML
  - Advantages: high accuracy for detecting relationships and patterns; self-learning; adaptable to different data and prediction needs.
  - Disadvantages: requires considerable data; interpretation is complex; preprocessing and feature engineering can be labor-intensive.

### Types of Machine Learning algorithms (Annex II)
- Supervised (Task-driven)
  - Features: learn mapping from input to output using labeled training data; useful for generating predictions; gets human feedback.
  - Applications: text classification; classification (discrete target); regression (continuous target).
- Unsupervised (Data-driven)
  - Features: analyzes unlabeled datasets without human labels; useful for exploration when goals are unclear.
  - Applications: extracting features; identifying trends and structures; clustering; density estimation; feature learning; dimensionality reduction; association rules; anomaly detection.
- Semi-supervised (Hybrid)
  - Features: combination of supervised and unsupervised; uses labeled and unlabeled data; can improve prediction using extensive unlabeled data.
  - Applications: machine translation; fraud detection; labeling data; text classification; learns from combined data.
- Reinforcement (Environment-driven)
  - Features: agents learn optimal behavior via rewards/penalties through trial and error; not given examples of optimal outputs.
  - Applications: robotics; autonomous driving; manufacturing and supply chain logistics.
  - Note: "Not preferable to use it for solving the basic or straightforward problems."

### Sample data dictionary for anomaly detection (Annex III)
- Dataset composition: variables on administrative, economic, functional, programmatic, geographical classifications; features of school facilities; cost data.
- Core variables (Variable / Type / Example Value):
  - Year / String / "2023"
  - ID_Ministry / String / "4"
  - Ministry / String / "Ministry of Education"
  - ID_School / Integer / 12345
  - School / String / "School_567"
  - ID_District / Integer / 78901
  - District / String / "District_78901"
  - ID_COFOG_Division / String / "9"
  - COFOG_Division / String / "Education"
  - ID_COFOG_Group / String / "9.3"
  - COFOG_Group / String / "Post-secondary non-tertiary education"
  - ID_Class / String / "09.3.0"
  - Class / String / "Post-secondary non-tertiary education"
  - ID_Budgetary_Program / String / "3"
  - Budgetary_Program / String / "Staff Training"
  - ID_Main_economic / String / "27"
  - Main_economic_classification / String / "Social benefits"
  - ID_Economic_sub / String / "272"
  - Economic_subclassification / String / "Social assistance benefits"
  - ID_revenue / String / "13"
  - Source_of_revenue / String / "Grants"
  - School_size / Integer / 750
  - Number_of_students / Integer / 1500
  - Number_of_rooms / Integer / 10
  - Students_per_room / Float / 150.0
  - Type_of_cost / String / "Fence installation - service"
  - ID_Agreement_contact / Integer / 98765
  - Total_costs / Integer / 4500000
  - Cost_per_student / Float / 3000.0
- Recommended columns for anomaly detection:
  - ID_School; Type of cost; School characteristics (size, number of students); ID_District; Economic classification; Functional classification (COFOG features).
- Rationale: school ID and costs are the most relevant categories; model applicability across contexts and maturity levels of alignment towards the Government Finance Statistics Manual 2014.

### Isolation Forest technical note (Annex IV)
- Core concept: isolate anomalies rather than profile normal points using recursive partitioning in tree-based models.
- Algorithm steps:
  1. Tree Construction: randomly select a feature and split value between feature minimum and maximum; repeat recursively until instance isolated or depth limit reached.
  2. Path Length Calculation: path length (ℎ(푥)) is the number of edges traversed from root to external node; anomalies generally have shorter path lengths.
  3. Anomaly Score Calculation: the path length is used to calculate anomaly score. The anomaly score, 푠(푥,푛), for a data point 푥 in a forest of 푛 trees is defined as:
     - 푠(푥,푛)=2
       퐸(ℎ푥)
       푐(푛)
     - Where:
       - 퐸(ℎ(푥)) is the average path length of 푥 across the forest.
       - 푐(푛) is the average path length of an unsuccessful searches in a Binary Search Tree (BST) of size 푛, and is approximated by:
         - 푐(푛)≈2( ln( n−1 ) +γ ) − 2(푛−1) 푛
       - γ≈0.5772 is the Euler–Mascheroni constant.
     - Note: normalization ensures anomaly scores are comparable across dataset sizes; anomalies tend to have scores closer to 1, normal points closer to 0.5 or below.
  4. Model Training and Prediction: does not require labeled data; trained on entire dataset; assigns anomaly scores and classifies short-path points as anomalies.
  5. Parameter Tuning: primary parameters:
     - n_estimators: Number of trees in the forest,
     - contamination: Expected proportion of anomalies in the dataset (used to set the threshold),
     - max_samples: Number of samples used to build each tree,
     - max_features: Number of features to draw from when looking for the best split.
  6. Labeling Data for Supervised Learning: after anomalies detected, label subset using:
     - Regulatory flags (known cases from audit institutions or anticorruption agencies).
     - Expert labeling with consensus-based decisions; exclude ambiguous cases.
     - Labeled examples can train supervised algorithms like Random Forest, XGBoost, or K-Nearest Neighbors in a semi-supervised pipeline.

### Example code snippet (Annex Box 1)
- The document provides an illustrative python example using pandas, sklearn.ensemble.IsolationForest, and simple preprocessing to fit a model with contamination=0.002, predict anomalies, and export results. (Code appears as an illustrative example in the source.)

*Source: IMF Working Papers — Making Cost Data Work for Public Financial Management (content from the supplied PDF chapter/section).*

### References

### wpiea2025159-source-pdf - References

### Major thematic clusters in the references
- Machine learning, data science, and fraud/anomaly detection
  - Barboza, F., Kimura, H., & Altman, E. (2017). Machine learning models and bankruptcy prediction. Expert Systems With Applications, 83, pp. 405-417. https://doi.org/10.1016/j.eswa.2017.04.006.
  - Bhattacharya, I. & Lindgreen, E. R. (2020). A Semi-Supervised ML Approach to Detect Anomalies in Big Accounting Data. European Conference on Information Systems.
  - Brady, E. S., Leider, J. P., Resnick, B. A., Alfonso, Y. N. & Bishai, D. (2017). Machine-Learning Algorithms to Code Public Health Spending Accounts. Public Health Reports, 132(3), pp. 350-356. doi.org/10.1177/0033354917700356
  - Green, B. P. & Choi, J. H. (1997). Assessing the Risk of Management Fraud through Neural Network Technology. Auditing: A Journal of Practice and Theory, 16(1), pp. 14-28.
  - Kondo, S., Miyakawa, D., Shiraki, K., Suga, M. & Usuki, T. (2019). Using Machine Learning to Detect and Forecast Accounting Fraud. Discussion papers 19-E-103, RIETI. https://www.rieti.go.jp/jp/publications/dp/19e103.pdf
  - Nielsen, S. (2020). Management accounting and the idea of machine learning. Economics Working Papers No. 2020-09.
  - Ranta, M., Ylinen, M. & Järvenpää, M. (2022). ML in Management Accounting Research: Literature Review and Pathways for the Future. European Accounting Review, 32(2), pp. 1-30.
  - Sarker, I. H. (2021). Data Science and Analytics: An Overview from Data-Driven Smart Computing, Decision-Making and Applications Perspective. SN Computer Science, 2(5).

- Public sector cost accounting, costing methods, and implementation cases
  - Alqudah, O. M. & Mohd Salleh, S. I. (2023). The Realities and Expectations of Calculating Government Service Costs: An Analysis of Jordanian Hospitals. International Journal of Economics and Financial Issues 13 (1): 65–72.
  - Brazil Secretaria do Tesouro Nacional (2025). Contabilidade de Custos no Setor Público: Experiências e Diretrizes. Brasília: Ministério da Fazenda. https://sisweb.tesouro.gov.br/apex/f?p=2501:9::::9:P9_ID_PUBLICACAO:51563.
  - Chan, J. A. & Pessoa, M. (2016). Finding the Costs of Public Services: The Experience of the State of São Paulo in Implementing a Cost System. In Pessoa, M. & Pimenta, C. (Eds.), Public Financial Management in Latin America: The Key to Efficiency and Transparency. Inter-American Development Bank.
  - Jovanović, T., Dražić Lutilsky, I. & Vašiček, D. (2019). Implementation of cost accounting as the economic pillar of management accounting systems in healthcare: a comparison of Slovenia and Croatia. Economic Research-Ekonomska Istraživanja, vol. 32(1), pages 3754-3772, January.
  - Malmmose, M., & Lydersen, J. P. (2021). From centralized DRG costing to decentralized TDABC-assessing the feasibility of hospital cost accounting for decision-making in Denmark. BMC health services research, 21(1), 835. https://doi.org/10.1186/s12913-021-06807-4
  - Miller, A. D. (2009). Managerial Cost Accounting in the Federal Government: Providing Useful Information for Decision Making. AGA CPAG Research Series: Report No. 22.
  - New Zealand Treasury. (2001). A Guide to Output Costing. Wellington: New Zealand. https://www.treasury.govt.nz/publications/guide/output-costing.
  - Pessoa, M. & Uña, G. (2017). Implementing a Public Services Cost System in São Paulo. IMF PFM Blog.
  - Ribeiro, R. et al. (2024). Management Public Accounting: An Analysis of Cost Information in Light of Public Managers’ Perception. Brazilian Business Review, 21(1).
  - Rivenbark, W. C. (2005). A Historical Overview of Cost Accounting in Local Government. State and Local Government Review, 37(3), pp. 217-227.
  - Rivenbark, W. C. (2017). The Intrinsic Value of Cost Accounting for Benchmarking Service Efficiency. In Mohr, Z. T. (Ed.), Cost Accounting in Government: Theory and Applications. Routledge.
  - Rivero del Paso, L., Bonari, D. & Iles, C. (2022). Consolidation of the Cost Accounting System – CUSTOS. Technical Report. IMF.
  - Rivero del Paso, L., Pattanayak, S., Uña, G. & Tourpe, H. (2023). Digital Solutions Guidelines for Public Financial Management. Technical Notes and Manuals 2023/007. IMF.
  - Secretaria do Tesouro Nacional. (2024). Relatório Foco em Custos – RFC 2023. Ministério da Fazenda, Brasil. https://www.tesourotransparente.gov.br/

- Performance budgeting, performance measurement, and linking costs to results
  - Behn, R. (2003). Why Measure Performance? Different Purposes Require Different Measures. Public Administration Review, 63(5), pp. 586-606. https://doi.org/10.1111/1540-6210.00322.
  - Breul, J. (2007). GPRA—a Foundation for Performance Budgeting. Public performance & Management Review, 30(3), pp. 312-331. https://doi.org/10.2753/pmr1530-9576300301.
  - Joyce, P. (2011). The Obama administration and PBB: building on the legacy of federal performance‐informed budgeting? Public Administration Review, 71(3), pp. 356-367. https://doi.org/10.1111/j.1540-6210.2011.02355.x
  - Mohr, Z. T. (2016). Performance Measurement and Cost Accounting: Are They Complementary or Competing Systems of Control? Public Administration Review, 76(4), pp. 616-625.
  - Mohr, Z. T., Raudla, R., & Douglas, J. W. (2018). Is Cost Accounting Used with Other NPM Practices? Evidence from European Countries. Public Performance and Management Review, 41(4), pp. 696-722.
  - OECD. (2018). OECD Best Practices for Performance Budgeting. GOV/PGC/SBO(2018)7.
  - OECD. (2023). OECD Performance Budgeting Framework. GOV/SBO(2023)1.
  - Robinson, M. (2007). Performance Budgeting Models and Mechanisms. Performance Budgeting: Linking Funding and Results, pp. 1-18. https://doi.org/10.1057/9781137001528_1.
  - Robinson, M. & Last, D. (2009). A Basic Model of Performance-Based Budgeting. Technical Notes and Manuals 09/01. IMF.
  - Trenovski, B. & Nikolov, M. (2015). Cost-benefit analysis of Performance Based Budgeting Implementation. CEA Journal of Economics, 10(2).

- Budget credibility, fiscal outcomes, and public financial management (PFM) systems
  - De Renzio, P. & Cho, C. (2020). Exploring the Determinants of Budget Credibility. International Budget Partnership (IBP).
  - De Renzio, P., Lakin, J. & Cho, C. (2019). How Deviations are Affecting Spending on Social Priorities. IBP.
  - Gurazada, S., Kirova, S., Alazmi, N., Rahman, A., & Pirlea, A. F. (2022). Government budget credibility and the impact of COVID-19. https://datatopics.worldbank.org/world-development-indicators/stories/government-budget-credibility-and-the-impact-of-covid-19.html.
  - IBP. (2023). Open Budget Survey 2023. https://internationalbudget.org/open-budget-survey-2023/.
  - IBP. (2023). Initiatives – Budget Credibility. https://internationalbudget.org/initiative/budget-credibility/.
  - Mustapha, S. (2019). Budget Credibility, Fiscal Outcomes, and PFM Performance in Fragile and Nonfragile Countries. In PEFA, Public Financial Management, and Good Governance. World Bank.
  - Rahim, F., Wendling, C. & Pedastsaar, E. (2022). How to Prepare Expenditure Baselines. How to Notes 2022/002. IMF.
  - Simson, R. & Welham, B. (2014). Incredible Budgets – Budget credibility in theory and practice. ODI.
  - Public Expenditure and Financial Accountability (PEFA) Secretariat. (2018). PEFA Handbook.

- Corruption, public procurement integrity, and open contracting tools
  - Abdou, A., Basdevant, O., David-Barrett, E., & Fazekas, M. (2022). Assessing Vulnerabilities to Corruption in Public Procurement and Their Price Impact. IMF Working Papers, 2022(094).
  - Basdevant, O. & Fazekas, M. (2023). The Corruption Cost Tracker: An Online Tool to Assess Corruption Risks in Public Procurement. Technical Notes and Manuals 2023/001. IMF.
  - Open Contracting Partnership. (2016). Red flags for integrity: Giving the green light to open data solutions.
  - Open Contracting Partnership. (2020). How governments spend: Opening up the value of global public procurement. Open Contracting Partnership and Spend Network.
  - Nestulia, V. (2020). How do you build a network of citizen corruption fighters? Ask Ukraine’s revolutionaries. OCP.
  - Salazar, C. & McKinney, J. (2024). Cardinal, an open-source library to calculate public procurement red flags. https://www.open-contracting.org/2024/06/12/cardinal-an-open-source-library-to-calculate-public-procurement-red-flags/
  - Red Interamericana de Compras Gubernamentales. (2022). Informe de implementación de Algoritmos de sistemas de alertas tempranas de irregularidades y banderas rojas en la compra pública.
  - World Bank. (2010). Fraud and Corruption Awareness Handbook.
  - IMF. (2019). Curbing Corruption. Fiscal Monitor (April 2019).
  - IMF. (2024). Governand and Anti-Corruption. https://www.imf.org/en/Topics/governance-and-anti-corruption.

- International organizations, technical guidance, and country practices
  - Federal Accounting Standards Advisory Board (FASAB). (2024). FASAB Handbook of Federal Accounting Standards and Other Pronouncements, as Amended.
  - Government of Canada. (2023). Guide to Costing. https://www.tbs-sct.canada.ca/pol/doc-eng.aspx?id=32600.
  - Public Services and Procurement Canada. (2022). Practitioner’s Guide for Procurement Pricing.
  - United States Government Accountability Office (U.S. GAO). (2020). Cost Estimating and Assessment Guide.
  - U.S. GAO. (2023). F-35 Joint Strike Fighter: More Actions Needed to Explain Cost Growth and Support Engine Modernization Decision.
  - UNICEF. (2021). AdaMM: Administrative Data Maturity Model Ver 1.0.
  - World Bank (2017). Improving Public Sector Performance through Innovation and Inter-Agency Coordination: Malaysia’s Experience with PEMANDU. Washington, DC: World Bank. https://documents1.worldbank.org/curated/en/318041492513503891/pdf/114270-WP-PUBLIC-13-4-2017-13-0-58-WorldBankReportPemanduFAFULLWeb.pdf.
  - OECD. (2024). Fiscal Sustainability of Health Systems: How to Finance More Resilient Health Systems When Money Is Tight? OECD Publishing. Paris, https://doi.org/10.1787/880f3195-en.
  - OECD (2021). Resourcing Higher Education in Denmark. Paris: OECD Publishing. https://www.oecd.org/content/dam/oecd/en/publications/reports/2021/12/resourcing-higher-education-in-denmark_6dab4b07/c8217325-en.pdf

### Cross-cutting practical guides, toolkits, and IMF products cited
- IMF How to Notes and Technical Notes and Manuals
  - Doherty, L. & Sayegh, A. (2022). How to Design and Institutionalize Spending Reviews. How to Notes 2022/004. IMF.
  - Rahim, F., Wendling, C. & Pedastsaar, E. (2022). How to Prepare Expenditure Baselines. How to Notes 2022/002. IMF.
  - Basdevant, O. & Fazekas, M. (2023). The Corruption Cost Tracker. Technical Notes and Manuals 2023/001. IMF.
  - Rivero del Paso, L., Pattanayak, S., Uña, G. & Tourpe, H. (2023). Digital Solutions Guidelines for Public Financial Management. Technical Notes and Manuals 2023/007. IMF.
  - Rivero del Paso, L., Bonari, D. & Iles, C. (2022). Consolidation of the Cost Accounting System – CUSTOS. Technical Report. IMF.

### Observable emphases for the Working Paper’s literature foundation
- Adoption of machine learning and data-science approaches for accounting, fraud detection, and anomaly detection.
- Extensive country-level and sectoral case studies on implementing cost accounting systems (healthcare, defense, higher education, subnational government).
- Integration of cost accounting with performance budgeting, budget credibility, and PFM reforms.
- Tools and open-source initiatives for procurement red-flagging, corruption risk assessment, and digital PFM solutions.
- Guidance from international standard setters and multilateral institutions on costing, costing guidance, and fiscal/administrative data maturity.

*References list as provided in the source PDF.*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2025/english/wpiea2025159-source-pdf.pdf_