## ppea2021003

## Source details

**Canonical URL:** [ppea2021003](https://www.imf.org/-/media/files/publications/pp/2021/english/ppea2021003.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/pp/2021/english/ppea2021003.pdf.md)
- [Structured JSON version](/-/media/files/publications/pp/2021/english/ppea2021003.pdf.json)

---

### Executive Board Assessment — key decisions and views
- Framework renamed “Sovereign Risk and Debt Sustainability Framework for Market Access Countries” (MAC SRDSF).
- Directors broadly supported reforms to:
  - improve prediction of sovereign stress,
  - enhance transparency and communication of results,
  - align with the three-zone sustainability assessment required under the exceptional access framework.
- Default institutional coverage: General Government (GG) debt per GFSM 2014; some Directors suggested phased implementation given that two-fifths of EMs currently report central government only.
- Public sector liquid financial assets accepted as a mitigating factor.
- Risk-based inclusion of central bank liabilities and/or SOE contingent liabilities supported; some Directors urged broader assets and net public debt concepts.
- Capacity development needed to improve country data coverage.
- Timeline: Most Directors supported roll-out in Q4 2021 or Q1 2022; views varied on speed and ambition.
- Disclosure and publication:
  - In program context: staff reports should contain full range of risk-of-sovereign-stress outputs for medium and long term (but not near term), plus an overall risk assessment.
  - In surveillance/precautionary arrangement cases: most Directors endorsed full disclosure to the Board but limited public disclosure (omitting near-term risk signal and assessment) for a 12-month period, with reassessment thereafter.
  - Implementing limited disclosure options would require a targeted modification to the Transparency Policy on a lapse-of-time basis.
- Use of realism tools and data:
  - Support for expanded realism toolkit and three-horizon risk assessment (short, medium, long).
  - Support for probabilistic DSAs for Fund-supported programs and to evaluate restructuring targets.
  - Concerns about exchange rate analysis in pegged regimes and use of perceptions-based third-party indicators; Directors asked to leave room for judgment and cross-checks.
- Sustainability assessments:
  - Required for arrangements involving GRA resources (including precautionary arrangements) and for the PCI.
  - Optional in surveillance; some favored mandatory in high-risk surveillance (Board-only disclosure).
  - Compromise: Board disclosure of three-zone assessments in both normal and exceptional access cases; public disclosure only in exceptional access cases; reassess after 12 months.
- Precautionary arrangements:
  - Sovereign risk assessments informed by baseline; sustainability assessments informed by baseline and, when appropriate, by an adverse (full drawing) scenario.

### Executive Summary — proposed framework and rationale
- Replace current MAC DSA with methodology based on three-horizon risk assessments:
  - Near-term (1–2 years): multivariate (logit) model predicting sovereign stress.
  - Medium term (5 years): (i) debt fanchart for debt stabilization prospects, (ii) gross financing needs (GFN) module for rollover risk, (iii) triggered stress-tests (natural disasters, commodity price shocks, banking stress).
  - Optional long-term tools (beyond 5 years).
- New framework requires additional data and disclosure: debt coverage, liquid assets, holder and maturity profile, country-specific risks.
- Outputs communicated using mechanical risk signals (high/moderate/low) at near- and medium-term horizons, feeding into judgment-based final assessments at all three horizons.
- Applications:
  - Derive statements about debt stabilization under current policies (mandatory in surveillance cases).
  - Three-zone debt sustainability assessments mandatory in program cases and optional in surveillance cases.
- Extensive testing indicates substantially better predictive accuracy than current framework.
- Renaming to “Sovereign Risk and Debt Sustainability Framework”.

### Proposed reforms — overview and objectives
- Objectives:
  - (i) increase robustness of sovereign risk analysis via broader debt coverage, longer projection horizon, enhanced realism tools;
  - (ii) improve prediction of sovereign stress using three-horizon analytical tools that account for structural characteristics and use continuous metrics;
  - (iii) enhance transparency in judgment and horizon-based bottom-line assessments.
- Additional aim: support probabilistic debt sustainability assessments as required by Fund lending framework.
- History and rationale:
  - MAC DSA introduced in 2002; 2011–13 review added realism tools, heatmap thresholds, fancharts, but shortcomings persisted (predictive power, coverage, optimism, unclear bottom-line assessments).

### Structure and analytic innovations
- Architecture: risk tools at three horizons with supporting modules.
  - Near-term: multivariate logit model (1–2 years) using parsimonious set of regressors in five buckets.
  - Medium-term (up to 5 years): improved debt fanchart, GFN module, triggered stress-tests.
  - Long-term: 10-year projections and optional modules for aging, natural resources, amortizations, climate change.
- Reporting: mechanical signals to improve transparency, followed by staff judgment; require explanation where judgment deviates from mechanical signals.

### Near-term tool (Logit) — specification, diagnostics, and signals
- Model: multivariate logistic regression with 10 regressors organized in five buckets: institutional quality; stress history; cyclical; debt burden and buffers; global.
- Estimation diagnostics:
  - Number of Observations: 1,579
  - LR chi2: 246.70
  - Pseudo R2: 0.25
- Estimated coefficients (Coeff. and Std. Coeff.):
  - Institutional Quality: Coeff. -1.073 *** ; Std. Coeff. -0.377
  - Stress History: Coeff. 0.514 *** ; Std. Coeff. 0.1006
  - Current account balance/GDP: Coeff. -0.024 ** ; Std. Coeff. -0.095
  - REER (3-year change): Coeff. 0.013 ** ; Std. Coeff. 0.070
  - Credit/GDP gap (t-1) (if + ve): Coeff. 0.086 *** ; Std. Coeff. 0.258
  - Δ(Public debt/GDP): Coeff. 0.052 *** ; Std. Coeff. 0.1182
  - Public debt/revenue: Coeff. 0.002 *** ; Std. Coeff. 0.1213
  - FX public debt/GDP: Coeff. 0.024 *** ; Std. Coeff. 0.1601
  - International reserves/GDP: Coeff. -0.034 *** ; Std. Coeff. -0.2348
  - ΔVIX (global): Coeff. 0.015 *** ; Std. Coeff. 0.1373
- Near-term mechanical risk signal (Logit LSP):
  - Low/moderate/high cutoffs calibrated to 10 percent missed-crisis / false-alarm probabilities.
  - Stress probability cutoffs: 9 percent (low/moderate) and 20.5 percent (moderate/high).
  - Interpretation: Low risk = LSP < 9 percent; High risk = LSP > 20.5 percent.
  - Model performance metrics:
    - Near-term tool sensitivity = 0.55 (80/146), specificity = 0.73 (1034/1411).
    - Ex-post crisis probabilities: P(S | H) = 0.40; P(S | L) = 0.02.
  - In-sample AUC = 0.88; minimum total misspecification error (TME) = 37 percent (missed crisis rate = 10 percent; false alarm rate = 27 percent).

### Medium-term tools: Debt Fanchart, GFN Module, Triggered Stress-Tests, and MTI
- Debt fanchart:
  - Two-step procedure: Step 1 historical fanchart realism diagnosis (baseline below 20th percentile triggers scrutiny); Step 2 final fanchart either “standard” (symmetric) or “realism adjusted” (asymmetric).
  - Fanchart metrics used: probability debt does not stabilize in medium-term; fanchart width; debt level at t+5 controlling for debt-carrying capacity.
  - Debt Fanchart Index (DFI): AUC = 0.82; minimum TME = 43 percent.
  - Posterior average stress probability by DFI zone: high risk = 44 percent; moderate = 23 percent; low = 3 percent.
- GFN Module:
  - Three indicators aggregated into GFN Financeability Index (GFI): initial bank claims on government (% banking system assets); maximum cumulative change in bank claims under stress; average projected GFN/GDP.
  - GFI performance: AUC = 0.83; composite TME = 42 percent.
  - Posterior average stress probability by GFI zone: high risk = 42.1 percent; moderate = 4.1 percent; low = 3.5 percent.
- Triggered stress-tests: contingent liabilities, banking crisis, natural disasters, commodity price shock, REER shock; default calibrations with customization options.
- Medium-Term Index (MTI):
  - MTI = average of DFI and GFI, divided into low/moderate/high using 10 percent missed-crisis and false-alarm cutoffs.
  - MTI posterior average stress probability by zone: high risk = 43 percent; moderate = 9 percent; low = 4 percent.
  - Judgmental adjustment: triggered stress-test results can worsen MTI final assessment by one notch under specified conditions (scenario high probability in RAM; not fully captured by fanchart; generates debt above 75th percentile).

### Long-term tools and horizon extension
- Proposed baseline extension to 10-year debt and GFN projections for all cases; uncertainty modeling (fancharts, stress-tests) limited to first 5 years.
- Optional long-term modules (no mechanical signals) for:
  - Population aging (pensions, health), climate change (30-year implications), natural resource volume changes, large debt amortizations.
- Teams must report judgment as low/moderate/high for identified long-term risks.

### Aggregation, judgment, and reporting requirements
- Mechanical signals:
  - Near-term (Logit LSP): low/moderate/high.
  - Medium-term (MTI): low/moderate/high.
  - Long-term: judgment-based low/moderate/high (no mechanical signals).
- Staff judgment:
  - Teams may deviate from mechanical signals; deviations should be explained in SRDSA write-ups; general presumption that deviations will not exceed one notch in medium-term.
- Reporting:
  - Sovereign risk analysis required in both surveillance and program contexts.
  - In program staff reports: full range of medium- and long-term outputs plus overall risk assessment (not near-term outputs).
  - Surveillance and precautionary arrangements: full Board disclosure; options for limited public disclosure requiring Transparency Policy modification; reassessment after 12 months.
  - Sustainability assessments required for arrangements involving GRA resources (including precautionary) and for the PCI; optional in surveillance cases.

### Debt coverage, central bank treatment, contingent liabilities, and disclosures
- Gross debt remains core concept with enhanced role for liquid assets where data permit; otherwise enhanced guidance applies.
- General Government (GG) per GFSM 2014 set as default institutional coverage; rationale: GG encapsulates all non-market government-controlled entities and aligns with Fiscal Monitor.
- Countries reporting CG or incomplete GG must justify narrower coverage; mandatory contingent liability stress-test calibrated on historical nonfinancial contingent liabilities per FAD database when coverage narrower than GG.
- Central bank consolidation:
  - Consolidation only where central bank has large negative capital and/or significant direct monetary financing/quasi-fiscal activities.
  - Otherwise incorporate mitigating role of central bank holdings without consolidation; risk-based inclusion of central bank liabilities (liquidity papers, bilateral FX swap liabilities) when conditions indicate material fiscal risk.
- Contingent liabilities:
  - Include in projections when likelihood and materialization can be anticipated; government guarantees included fully as public debt if high likelihood of materialization (GFSM principles).
- Strengthened disclosures:
  - SRDSA to include metadata on institutional and instrument coverage; valuation method (nominal, face, or market value); consolidated tables by level of government and cross-holdings; breakdowns by domestic/foreign law, marketable/non-marketable, currency composition, and holder profile (foreign official, foreign private, domestic central bank, domestic commercial banks, domestic nonbank).
- Fund credit:
  - Fund credit disbursements included as public debt for SRDSA purposes.

### Predictive performance, calibration, and audit outcomes
- Overall predictive performance improves significantly relative to single-variable thresholds.
- Key performance excerpts (Table 3 style summary):
  - Logit model (1990-2015): AUC/metrics reported in source (in-sample AUC = 0.88; minimum TME = 37 percent).
  - Medium-Term Index (2014-15): AUC = 0.85 reported in table excerpts.
  - Debt fanchart (2010-15): AUC = 0.82.
  - GFN module (2014-15): AUC = 0.83.
  - Existing framework OR condition TME: 79 percent; AND condition produced NaN/100%/100%/0% table entries in source.
- Cutoff design:
  - High risk cutoff associated with false-alarm probability = 10 percent.
  - Low risk cutoff associated with missed-crisis probability = 10 percent.
- Audit and validation:
  - All analytical tools audited by independent RES/ICD team; Annexes IV and V respond to audit points and discuss modifications.

### Backtesting, datasets, and stress episode identification
- Stress identification:
  - 486 stress country-years corresponding to 139 distinct “stress episodes”.
  - Sample for near-term tool covers 1990-2017 and most MACs.
- Main triggers of stress episodes:
  - Defaults: 37 percent
  - Market stress: 32 percent
- Regressor selection and methodology:
  - Over 150 candidate variables considered; final logit chosen for interpretability and robust forecasting power.
  - Bayesian selection methods and robustness checks conducted; Random Forest (VEFM) provided complementary insights but logit selected as workhorse.
- Robustness checks:
  - Leave-one-country-out, rolling windows, fixed-effects considerations, heteroscedasticity corrections implemented and documented.

### GFN Module design, data needs, and simulated performance
- GFN module inputs: GFN-to-GDP projections, holder allocations (central bank, commercial banks, other domestic, foreign official, foreign private), default disaggregation where needed.
- Stress scenario design:
  - Generalized adverse shocks to macro-fiscal variables, maturities (shortening), and access to external debt markets (foreign private rollover to 67 percent and no new foreign private financing for two years).
  - Residual financing assumed absorbed by domestic banks unless customized.
- GFI performance and backtesting:
  - Composite index AUC = 0.83; TME = 42 percent (2014-15 backtest).
  - Posterior probabilities: GFI high risk posterior stress probability at least 24 percent; average high risk = 42.1 percent.
- Data requirements and practical pointers:
  - New data needs: amortization by debtholder, holder-profile of stock, 10-year projections, inputs for stress tests and long-term modules.
  - Use of Arslanalp-Tsuda methodology and centralized databases to approximate holder amortizations; teams encouraged to refine with authorities.

### Fanchart methodology and realism adjustment
- Fanchart two-step procedure:
  - Step 1: historical fanchart via block-bootstrap of debt driver joint empirical distribution; baseline below 20th percentile triggers realism scrutiny.
  - Step 2: generate standard symmetric fanchart or realism-adjusted asymmetric fanchart by adding skewed shocks until baseline percentile aligns with cross-country distribution; minimum adjustment percentile = 10th percentile.
- Covid-19 modification:
  - For 2021-22 recovery, modified historical fanchart partly dependent on team baseline for first two years to limit incorrect realism corrections.
- Limitations:
  - No feedback between debt drivers and debt level; foreign currency debt shares fixed at baseline; interest rate shocks calibrated on average effective rates—implies fanchart understates true uncertainty in some dimensions.
- Fanchart predictive metrics and DFI:
  - DFI AUC = 0.82; minimum TME = 43 percent; calibration to 10 percent missed/false rates.

### Implementation plan, training, and transition
- Materials to support implementation expected by second half of 2021: guidance note, software, Excel files.
- Early engagements: selected country teams to test tools in parallel with current framework; outreach to country authorities and debt management offices before rollout; external stakeholder engagement during 2021 Spring and Annual Meetings.
- Reporting format: published DSAs under SRDSF could be 7–8 pages with standardized tables/charts and staff commentaries.
- Transition timing: current framework used until rollout expected Q4 2021 or Q1 2022; SPR implementation team to support; guidance note to be adjusted based on transition experiences.
- Training and capacity building:
  - Comprehensive internal and external training planned; guidance note development expected [6-8] months post-approval per source.
  - Fund TA support for countries needing statistical capacity upgrades; transitional solutions for frontier LICs and recent PRGT graduates.

### Illustrative example — Ruritania (selected numeric points)
- Final assessment: Moderate.
- Near term:
  - Mechanical signal: High
  - Final assessment: Moderate — mitigating role of large liquid asset buffer.
- Medium term:
  - Mechanical signal: High
  - Final assessment: Moderate — liquid asset buffer cited; fanchart width major contributor.
  - Reported components: GFN: Moderate; Fanchart: High; Stress-test: Yes (commodity price shock)
- Long term:
  - Mechanical signal: Low
  - Final assessment: triggered module indicates scaling up of natural resources lowers debt long term.
- Selected projections and magnitudes (percent of GDP):
  - Nominal gross public debt: 53.5 58.7 61.1 62.2 65.1 69.3 73.1 70.0 62.4 54.8 47.2 39.6
  - Change in gross public sector debt: 6.7 5.1 2.5 1.1 3.0 4.1 3.9 -3.1 -7.6 -7.6 -7.6 -1.3
  - Identified debt-creating flows: 7.7 8.3 5.0 3.7 5.2 5.8 6.1 -0.9 -5.4 -5.4 -5.4 1.1
  - Primary deficit: 7.2 5.0 4.9 3.3 4.1 4.7 4.9 -2.1 -6.6 -6.6 -6.6 -0.1
  - Gross Financing Need: 14.8 15.2 13.7 12.6 15.5 13.0 13.0 13.0 13.0 13.0 13.0 13.0 13.5
  - Large realized/projected increase in public debt to finance oil/gas investments corresponding to US$20 billion or 25 percent of nominal GDP.
  - Liquid assets equal to 70 percent of total debt cited as mitigating factor.
  - Additional revenue expected after t+5 of about US$40 billion at current gas prices.
- Note: Sustainability assessment not required as country not in Fund-supported program; risk of sovereign stress reported only for surveillance cases.

### Issues for Board discussion (selected)
- Continue existing definition of debt sustainability?
- Endorse naming the proposed framework “Sovereign Risk and Debt Sustainability Framework”?
- Agreement with GG default coverage and proposed 10-year horizon?
- Agreement with expanded realism tools and realism adjustments?
- Support for horizon-based mechanical risk signals with required explanation for judgement deviations?
- Disclosure options:
  - For program cases: continue current practice (three-zone assessments only in exceptional access staff reports) OR disclose to Board in both normal/exceptional but public only in exceptional (Transparency Policy change)?
  - For surveillance reports: publish full outputs OR Board-only with limited public disclosure (Transparency Policy change)?
- Support use of new tools for debt restructurings and proposed DSA reporting format?
- Support proposed timeline for implementation?

*Source: MAC DSA REVIEW (selected excerpts from ppea2021003).*

### 13. The 2011–13 review introduced key features, including a risk-based approach through

### 13. The 2011–13 review introduced key features, including a risk-based approach through

### Executive Board Assessment — key decisions and views
- Framework to be renamed “Sovereign Risk and Debt Sustainability Framework for Market Access Countries” (MAC SRDSF).
- Directors broadly supported reforms to:
  - improve prediction of sovereign stress,
  - enhance transparency and communication of results,
  - align with the three-zone sustainability assessment required under the exceptional access framework.
- Default institutional coverage: General Government (GG) debt defined per GFSM 2014 classification, with some Directors suggesting phased implementation given that two-fifths of EMs currently report central government only.
- Inclusion and treatment of public sector liquid financial assets accepted as a mitigating factor.
- Risk-based approach supported for including central bank liabilities and/or SOE contingent liabilities in debt perimeter; some Directors urged broader public sector assets and wider adoption of net public debt concepts.
- Capacity development flagged as necessary to improve country data coverage.
- Timeline: Most Directors supported roll-out in Q4 2021 or Q1 2022; some preferred faster or judged it ambitious. Transition between old and new framework to be carefully managed.
- Disclosure and publication:
  - In program context: staff reports should contain full range of risk-of-sovereign-stress outputs for medium and long term (but not near term), plus an overall risk assessment.
  - In surveillance/precautionary arrangement cases: most Directors endorsed full disclosure to the Board but limited public disclosure (omitting near-term risk signal and assessment) for a 12-month period, with reassessment thereafter. Views varied from concern about market sensitivity to calls for immediate full public disclosure.
  - Implementing limited disclosure options would require a targeted modification to the Transparency Policy on a lapse-of-time basis.
- Use of realism tools and data:
  - Directors welcomed expanded realism toolkit and three-horizon risk assessment (short, medium, long).
  - Support for new tools to produce probabilistic DSAs for Fund-supported programs and to evaluate restructuring targets.
  - Concerns noted about: (i) expansion to exchange rate analysis in pegged regimes; (ii) use of perceptions-based third-party indicators for institutional quality — Directors asked to leave room for judgment and cross-check with alternative non-perception indicators.
- Sustainability assessments:
  - Agreed required for arrangements involving GRA resources (including precautionary arrangements) and for the PCI.
  - Most Directors: optional in surveillance; some favored mandatory in high-risk surveillance cases (disclosed to Board only).
  - For program cases: views varied; final compromise — disclosure to the Board of three-zone assessments in both normal and exceptional access cases, public disclosure only in exceptional access cases, with experience assessed after 12 months.
- Precautionary arrangements:
  - Sovereign risk assessments informed by baseline; sustainability assessments informed by baseline and, when appropriate, by an adverse (full drawing) scenario.
  - Adverse scenario appropriate in exceptional access cases (excluding FCL cases) if shocks triggering drawing are not captured by medium-term tools or doubts about realism remain.

### Executive Summary — proposed framework and rationale
- A careful review found significant scope to modernize the MAC DSA to better predict sovereign stress and to support three-zone sustainability assessments.
- The note proposes replacing current framework with a methodology based on risk assessments at three horizons:
  - Near-term: multivariate (logit) model predicting sovereign stress over 1–2 years.
  - Medium term (5 years): (i) a debt fanchart to assess prospects for debt stabilization, (ii) a gross financing needs module for granular rollover risk analysis, and (iii) triggered stress-tests for specific risks (e.g., natural disasters, commodity price shocks, banking stress).
  - Optional long-term tools (beyond 5 years).
- The new framework requires additional data and disclosure in areas including debt coverage, liquid assets, holder and maturity profile of debt, and country-specific risks — entailing additional resource costs to be contained via automation.
- Extensive testing indicates substantially better predictive accuracy than current framework.
- Outputs communicated using mechanical risk signals (high/moderate/low) at near- and medium-term horizons, feeding into judgment-based final assessments at all three horizons.
- Applications:
  - Derive statements about debt stabilization under current policies (mandatory in surveillance cases).
  - Three-zone debt sustainability assessments mandatory in program cases and optional in surveillance cases.
- Renaming the framework to “Sovereign Risk and Debt Sustainability Framework”.

### Background — limitations of the existing 2011–13 framework
- The 2011–13 review introduced:
  - expectation of comprehensive on- and off-balance sheet debt coverage and option to project beyond 5 years,
  - realism tools for growth, primary balance, and inflation,
  - heatmap for vulnerability thresholds (debt, gross financing need, debt profile),
  - debt fancharts to capture distribution of risks,
  - standardized DSA template, write-up, and publication requirements.
- Persistent shortcomings identified:
  - mixed capacity to predict sovereign stress,
  - uneven/inadequate debt coverage,
  - baseline optimism,
  - unclear bottom-line assessments,
  - mechanical outputs generating noise that complicates transparent, judgment-based adjustments,
  - inadequate foundation for probabilistic debt sustainability assessments and three-zone classifications.

### Structure and analytic innovations in the proposed framework
- Architecture summarized as risk tools at three horizons, with supporting modules:
  - Near-term multivariate logit model (1–2 years) for predicting sovereign stress.
  - Medium-term (5 years):
    - Debt fanchart methodology for distribution around baseline and debt stabilization prospects.
    - Gross Financing Needs (GFN) module for rollover and refinancing risk.
    - Triggered stress-tests for country-specific vulnerabilities.
  - Long-term tools to analyze structural risks beyond 5 years.
- Reporting: mechanical signals to improve transparency, followed by staff judgment.
- Framework designed to support probabilistic DSAs required under Fund lending.

### Data, coverage, and realism tools
- Debt coverage: broader and more consistent coverage intended, with General Government (GG) as default.
- Incorporation of public sector liquid financial assets as mitigating factors.
- Realism toolkit expanded to reduce baseline optimism and to cover additional domains (including proposed expansion to exchange rate analysis, subject to Directors’ concerns for pegged regimes).
- Emphasis on better disclosure of debt-related data: liquid assets, holder and maturity profiles, and country-specific risk factors.
- Recognition that many EMs will need capacity development to meet data requirements.

### Predictive accuracy, communication, and implementation
- Extensive testing shows improved predictive accuracy relative to the current framework.
- Mechanical three-tier risk signals (high/moderate/low) at near- and medium-term horizons make results more communicable.
- Operationalization expected in Q4 2021 or Q1 2022, preceded by completion of Guidance Note and template, plus extensive engagement with country authorities and external stakeholders.
- Transition will be carefully managed to ensure consistency; early engagement with selected country teams to test tools in parallel with current framework recommended.
- Automation to contain resource costs; capacity development support and close Board engagement encouraged.
- Implementation to be accompanied by an effective communication strategy with member-country authorities and external stakeholders.

### Use in Fund operations and surveillance
- Sovereign risk analysis preparation generally required in both program and surveillance contexts.
- Program context: full range of medium- and long-term outputs plus overall risk assessment to be included in staff reports (not near-term outputs).
- Surveillance and precautionary arrangements: full Board disclosure but limited public disclosure for 12 months (omitting near-term signal), subject to reassessment.
- Sustainability assessments required for arrangements involving GRA resources (including precautionary) and for PCI; optional in surveillance generally, with case-by-case application in high-risk surveillance.
- In precautionary arrangements, sustainability assessment may include adverse full-drawing scenarios where appropriate.

*Source: MAC DSA REVIEW (excerpt), November 25, 2020.*

### 4.      Against this backdrop, in this paper staff proposes a set of reforms to the framework.

### ppea2021003 - 4.      Against this backdrop, in this paper staff proposes a set of reforms to the framework.

### Proposed reforms (overview)
- Objectives of the proposed reforms:
  - (i) increase the robustness of sovereign risk analysis through broader debt coverage, a longer projection horizon and enhanced realism tools;
  - (ii) improve the framework's capacity to predict sovereign stress through new analytical tools at three different time horizons that both account for countries’ structural characteristics, and rely on continuous metrics rather than discrete, single-variable thresholds; and
  - (iii) enhance transparency in exercising judgment and arriving at (horizon-based) bottom-line assessments.
- Additional aim: support probabilistic debt sustainability assessments, as required by the Fund's lending framework.

### Brief history of the MAC DSA
- The MAC DSA was introduced in 2002 to improve the consistency and discipline of debt sustainability analyses.
- Key milestones:
  - 2003 and 2005: reviews introduced refinements.
  - 2011–13: last review introduced important reforms in response to shortcomings revealed by the Global Financial Crisis (GFC) and euro area sovereign debt crises.
- Shortcomings uncovered by the 2011–13 review included: inconsistent use or discussion of the public DSA; optimistic growth projections in several crisis countries; lack of a bottom-line sustainability assessment; limited scope for country heterogeneity; and ineffective tools to illustrate uncertainty.
- The current MAC DSA framework, launched in 2013, introduced:
  - A requirement for at least one DSA per year for program cases and one per Article IV cycle for non-program cases;
  - A risk-based approach distinguishing high- and low-scrutiny countries;
  - New DSA template elements: (i) an analysis of the realism of baseline projections, (ii) a heat map anchored by noise-to-signal (NTS) based thresholds for debt, GFNs, and debt profile indicators, and (iii) debt fancharts to capture the full distribution of risks around the baseline.

### Definition of debt sustainability (Board-approved)
- Quoted Board-approved definition (paragraph 6):
  - “In general terms, public debt can be regarded as sustainable when the primary balance needed to at least stabilize debt under both the baseline and realistic shock scenarios is economically and politically feasible, such that the level of debt is consistent with an acceptably low rollover risk and with preserving potential growth at a satisfactory level.”
- Interpretation:
  - Includes both solvency and liquidity requirements.
  - Stronger than some academic definitions that focus only on solvency, justified because:
    - Borrowing costs and market access depend on (actual and perceived) solvency, making a clear-cut solvency/liquidity distinction impractical.
    - The IMF’s lending framework uses debt sustainability as an indicator of capacity to repay the Fund; lack of liquidity can impair repayment capacity.

### Application differences: surveillance vs program contexts
- Surveillance context:
  - IMF’s role: alert the member to likelihood of sovereign stress and help steer away from it.
  - Solvency and liquidity risks are equally relevant.
  - Fund signals stress without taking a view on how it will be resolved.
  - Sovereign stress is broader than unsustainable debt.
- Program context:
  - Assessments pay particular attention to (conditional) solvency, but cannot ignore market financing risks after incorporating Fund financing in the baseline.
  - Short-term liquidity risks can be reduced or eliminated by IMF financing in some cases.
  - Liquidity risks matter because:
    - Market access after program completion affects capacity to repay.
    - Countries borrowing from the market during the program can face liquidity risks that undermine program success.

### Clarifying related concepts and renaming proposal
- Three related but distinct concepts (Box 2):
  - Risk of sovereign stress: likelihood of experiencing stress, regardless of resolution method. Broader than “unsustainable”.
  - Unsustainable debt: no set of politically and economically feasible policies can stabilize debt/GDP with acceptably low rollover risk. Implies need for debt restructuring (and/or exceptional bilateral support).
  - Failure of debt to stabilize under baseline policies: debt/GDP does not stabilize under staff’s best prediction by end of projection horizon; may or may not imply unsustainability under the Fund’s definition.
- Staff proposal:
  - Rename MAC DSA to “Sovereign Risk and Debt Sustainability Framework” (SRDSF).
  - Refer to output as “Sovereign Risk and Debt Sustainability Analysis” (SRDSA).
  - Rationale: frameworks are tools to analyze and warn about risk of sovereign stress and to inform debt sustainability assessments, so renaming clarifies purpose.

### Assessment of the current framework against five Board-endorsed objectives
- Paraphrased objectives from the 2011–13 review:
  - (i) adequate coverage and disclosure of debt-related risks,
  - (ii) discriminatory (predictive) capacity,
  - (iii) realistic baselines, including robust representation of uncertainty,
  - (iv) a risk-based approach with greater scrutiny for higher-risk countries,
  - (v) a sharper output allowing effective, transparent, and even-handed judgment.
- Staff evaluation identified areas for improvement (detailed in subsequent subsections).

### Adequate coverage and disclosure of debt-related risks (key findings)
- Coverage gaps:
  - Although 2011–13 review introduced general government debt as the appropriate concept, actual coverage remains narrow in many cases.
  - About two-fifths of EMs still restrict coverage to the central government, with little improvement over time.
  - No requirement to report the instrument and valuation basis for the debt reported in the DSA.
- Debtholder profile and disclosure:
  - Insufficient information on the debtholder profile; no disclosure requirements beyond a resident/non-resident split.
  - Debtholder composition can be critical for assessing rollover risk and safeguards.
- Additional coverage issues needing attention:
  - (i) Better incorporate liquid assets in analytical tools (currently enter mainly as judgment or to help classify low-scrutiny countries).
  - (ii) Report contingent liability risks from narrow institutional coverage, government guarantees, PPPs, and SPVs.
  - (iii) Treat Fund credit intended for boosting reserves (no explicit guidance though legally all Fund credit should be included in public debt).
  - (iv) Account for central bank liabilities, such as FX swaps and liquidity paper (increasingly important).
  - (v) Handle cases where central bank holdings of government debt materially reduce measured debt and GFNs if consolidated.
- Long-term fiscal pressures:
  - Assessments have rarely considered long-term fiscal pressures.
  - Longer horizons are essential for capacity-to-repay assessments in UFR cases and for setting/evaluating debt and GFN targets in restructurings.
  - The 2013 guidance note allowed projections beyond 5 years where relevant, but long-term projections have been limited.

### Discriminatory (predictive) capacity (key findings)
- Background:
  - 2011–13 review introduced thresholds for debt, GFN and five debt profile indicators with a heatmap; thresholds calibrated to minimize errors in predicting crises one year ahead.
  - Different risk thresholds were adopted for AEs and EMs.
- Performance issues:
  - Predictive power of single-variable thresholds has been weak.
  - Empirical result: in 57 percent of all stress episodes that occurred in AEs in 2007–13 and EMs in 2007–18, both the debt and the GFN indicators failed to flash.
  - Debt profile indicators, especially those capturing external vulnerabilities, showed greater discriminatory power relative to debt and GFN.
- Known limitations and heatmap aggregation:
  - The 2011 Board paper noted a “lack of empirical basis for generalized debt thresholds.”
  - When thresholds were introduced in 2013, the final risk assessment was expected to rely on the overall heatmap rather than breaches of individual thresholds; guidance did not specify how to aggregate breaches.
- Additional performance shortcomings:
  - Similar heatmaps for countries with very different risk profiles: AEs and EMs buckets may be insufficient to capture wide variation in debt carrying capacity.
  - Lack of attention to timing and magnitude of risks:
    - Heatmap treats threshold breaches the same regardless of whether they occur early or late in the projection horizon.
    - Treats past breaches the same as projections of future breaches.
    - Does not account for magnitude of the breach, so heatmap may not change despite material changes in risks.

*Source: ppea2021003 - 4.      Against this backdrop, in this paper staff proposes a set of reforms to the framework.*

### 21.      While re-estimation of current thresholds on a larger sample of countries or stress

### ppea2021003 - 21.      While re-estimation of current thresholds on a larger sample of countries or stress

### Limitations of current empirical approach
- Key predictors of sovereign stress are omitted from the present heatmap: institutional quality, history of stress, cyclical imbalances, and global risk appetite.  
- Separate single-variable group-wide thresholds do not account for interactions among variables, which may be relevant to appropriately capture heterogeneity across countries.  
- Continuous models have been found to have better predictive performance than threshold-based approaches when applied to the same set of variables, because the latter only classify countries “in stress/not in stress”, without giving weight to the severity of the breach, and do not account for variable interaction.

### C. Baseline realism and robust modeling of uncertainty
Findings:
- Introduction of visual realism tools in the 2013 framework appears to have helped reduce optimism in baseline projections, but forecast debt trajectories remain optimistic.
- On average, projection errors for debt drivers covered by the realism tools—primary balance and real growth rate—were smaller than for those not covered (exchange rate and interest rate).
- Forecasts for the change in debt/GDP remain more optimistic than outturns, and medium-term debt stabilization is predicted more frequently than it occurs.
- The 2013 macro-fiscal stress-tests are deterministic: assumed shocks and transmission mechanisms are identical across all countries; shocks are simulated individually rather than jointly, making scenarios far from realistic stress episodes.
- Stochastic tools like debt fancharts are underutilized and suffer methodological shortcomings, including:
  - assumptions about the normality of the distribution from which debt driver shocks are derived;
  - failure to provide an accurate picture of risks if the baseline is optimistic and/or risks are tilted to the downside;
  - lack of standardization in deriving the asymmetric fanchart, precluding comparability across countries;
  - inability to account for uncertainty related to debt data revisions (base effects).
- Empirical point: 80 percent of the debt and GFN trajectories under the primary balance, interest rate, and exchange rate stress-tests fell within the 10-90th percentiles of the symmetric fanchart.

Implications / Possible improvements referenced:
- Enhance realism toolkit to cover exchange rate, interest rate, stock-flow adjustments.
- Mainstream use of continuous, stochastic methods to better capture risk distribution.
- Standardize fanchart, robust to optimism.

### D. Classification of countries into low and high scrutiny “buckets”
Findings:
- The division into low and high scrutiny buckets has practical problems:
  - Countries classified as low scrutiny ended up in crisis (examples cited: Bosnia, Georgia); criteria would have precluded deeper analysis for some low debt and GFN AEs that experienced stress post-GFC (examples cited: Ireland, Iceland, Spain).
  - Within high-scrutiny group, analysis is not adequately risk-based; differences within the high-scrutiny bucket are not fully captured, and there is virtually no use of triggered stress-tests (beyond the generalized contingent liability shock) that might elucidate key vulnerabilities.

Recommendations:
- Ensure minimum risk analytics for all countries.
- Introduce triggered stress-tests/modules for key vulnerabilities (e.g., commodity prices, natural disasters, aging population).

### E. Granularity, aggregation, and application of judgment
Findings:
- Limited accuracy of the threshold approach, lack of granularity in capturing risks, and absence of cohesive aggregation of outputs constrain effective, transparent, and evenhanded application of judgment.
- The predictive shortcomings of mechanical tools place an undue burden on judgment; team judgment did not improve upon the performance of the heatmap in either stress or non-stress cases.
- Judgment rarely offsets the lack of tools for country-specific aggravating and mitigating factors (examples: natural disasters in small states; commodity price swings and resource depletion/discovery in commodity exporters; long-run fiscal costs; large government assets). The chapeau in most DSAs lacked discussion of key risks and mitigating factors not already captured by mechanical tools.
- No guidance on when to invoke the option to perform a supplemental DSA using net debt.
- Framework output consists of many disaggregated results (at times conflicting) from the heatmap and other tools; aggregation and interpretation are left to team discretion, making it difficult to understand whether—and why—staff disagrees with the framework’s output.
- The current framework does not require standardized reporting of risks in surveillance cases; Fund programs require a bottom-line statement on debt sustainability, but there is no analogous requirement in surveillance cases. This limits meaningful cross-country comparisons and comparisons over time.
- External stakeholders identified the lack of aggregate mechanical signals, unclear bottom-line assessments, and non-transparent application of judgment as the framework’s most notable weaknesses; noted that the EC’s framework and the LIC-DSF perform better in this respect.

Recommendations:
- Introduce horizon-based summary mechanical risk signals to provide clear starting points for application of judgment.
- Allow final horizon-based assessments to be judgment-based, but require explanation where judgment deviates from mechanical signals.
- Standardize DSA template, write-up, and publication requirements.
- Integrate liquid assets in a more standardized way.
- Consider 10-year horizon for all countries; introduce tools to analyze specific risks beyond 5 years (e.g., aging).

### Structure of the proposed framework (overview and key features)
Objectives of the reform:
- increase robustness of sovereign risk analysis via broader and more consistent debt coverage, a longer projection horizon, enhanced realism tools, and superior analytical methods;
- improve the framework’s discriminatory (predictive) capacity by introducing a horizon-based approach that accounts for country-specific structural characteristics and uses continuous rather than discrete metrics;
- enhance transparency in the bottom-line assessment (for each horizon) and in the exercise of judgment.
- Support probabilistic debt sustainability assessments, as required by the Fund’s lending framework.

Proposed architecture (components and features):
- Coverage:
  - General Government (GG) as default; justification required for narrower coverage; broader coverage (including central bank) in some cases.
  - Disclosure requirements on coverage definitions, debtholder profile, and guidance on certain instruments (like swaps).
- Horizon:
  - 10-year debt and GFN projections for all cases.
  - Risk assessments for near-, medium-, and long-term horizons.
- Realism tools:
  - Cover additional drivers (exchange rate, financing terms on external debt, stock-flow adjustments), and public debt.
  - In-depth tools for potential growth and fiscal multipliers.
- Near-term risk analysis (1–2 years ahead):
  - Multivariate logit model based on actual (i.e. not projected) data.
  - Aggregates information from stress drivers and mitigating factors; accounts for country-specific structural characteristics; provides more granular discrimination than AE/EM bucketing.
- Medium-term risk analysis (up to 5 years ahead):
  - Debt fanchart to probabilistically assess prospects for debt stabilization.
  - GFN module to analyze rollover risks, taking into account creditor composition.
  - Triggered/tailored stress-tests to assess country-specific risks not captured elsewhere.
  - Three indicators for medium-term index: i) probability debt does not stabilize in medium term, ii) fanchart width, iii) debt level at t+5 controlling for debt-carrying capacity (fanchart accounts for deviation of baseline projections from historical trends via skewed shocks). Index based on 3 indicators weighted by predictive power; index split in low, moderate, and high-risk zones (calibrated to 10% missed crisis and false alarm rates).
  - GFN Tool: three indicators—(i) initial bank claims on government, (ii) maximum cumulative change in bank claims over projection period under a generalized stress scenario; (iii) average projected GFN/GDP in baseline. Index based on 3 indicators weighted by predictive power; index split in low, moderate, and high-risk zones (calibrated to 10% missed crisis and false alarm rates).
  - Triggered stress-tests to simulate debt and GFN paths under: (i) contingent liabilities related to narrow coverage, (ii) banking crisis, (iii) natural disasters, (iv) commodity price shocks, and (v) REER shock. Allows for customized stress-tests for idiosyncratic risks.
- Long-term risk analysis (beyond 5 years):
  - 10-year debt and GFN projections, and optional tools for risks from population aging, natural resource discovery/depletion, debt amortizations, and climate change; option to extend debt and GFN projections.
- Fanchart tool:
  - Visual tool based on symmetric shocks (asymmetric shocks used at team’s discretion).
  - Proposed improvements: debt fanchart probabilistically assesses stabilization using three indicators (probability of non-stabilization, width, and level at t+5), and accounts for baseline deviations via skewed shocks.
- Macroeconomic shocks:
  - Effect of shocks to primary balance, real GDP growth, real interest rate, and exchange rate on debt and GFN levels reflected in heat map signals.
- Aggregation and judgment:
  - Outputs produce mechanical near-term and medium-term signals (low; moderate; high) to serve as starting points.
  - Teams provide judgment-based near-term, medium-term, and long-term assessments (low; moderate; high), with requirement to explain deviations from mechanical signals.
  - DSA output: writeup, including overall assessment of risks.

Tabled comparisons summarized (proposed vs. existing framework):
- Proposed: GG default coverage; 10-year horizon; enhanced realism tools; multivariate logistic regression for near-term; probabilistic fancharts integrated; triggered stress-tests; standardized indices split into low/moderate/high calibrated to 10% missed crisis and false alarm rates; judgment required to explain deviations.
- Existing: narrower than GG in some cases; 5-year projections; realism tools limited to growth, inflation and primary balance; deterministic stress-tests; heatmap with no aggregation/overall signal; no standardized bottom-line assessments.

### Final summary of implementation aims
- Use multivariate early warning model(s) with broader set of stress drivers and mitigants.
- Develop tools/metrics for medium-term debt and GFN vulnerabilities that account for differences in institutional capacity and creditor profile.
- Introduce horizon-based summary mechanical risk signals and require explanation when team judgment deviates from mechanical signals.
- Mainstream probabilistic, continuous, and standardized tools (fancharts, GFN module, triggered stress-tests) and extend projection horizon to 10 years where appropriate.

*Source: MAC DSA REVIEW (excerpt from ppea2021003).*

### 30.      Each element of the proposed framework has been rigorously tested and audited.

### ppea2021003 - 30.      Each element of the proposed framework has been rigorously tested and audited.

### Overall predictive performance
- Overall predictive performance, as measured by the sum of missed crisis and false alarm rates, significantly improves relative to the existing framework’s single-variable thresholds.
- Table 3 excerpts (as presented):
  - Logit model (1990-2015) 2: 30%37%10%27%0.88
  - Medium-Term Index (2014-15): 14%38%27%10%0.85
  - Debt fanchart (2010-15 ): 34%43%14%29%0.82
  - GFN module (2014-15): 13%42%33%9%0.83
  - OR condition (2006-16, AE and EM average): 76%79%12%68%
  - AND condition (2006-16, AE and EM average): NaN100%100%0%
- Interpretation provided in text for the current framework:
  - A crisis prediction based on the “OR condition” rarely misses a crisis (just 12%), but sends false alarms 68% of the time, yielding a Total Misspecification Error (TME) of 79%.
  - A crisis prediction based on the “AND condition” never sends a false alarm (all crises are associated with at least one red signal) but misses all crises (no crisis is associated with the entire heatmap being red).

### Design of the near- and medium-term tools
- The design is based on three steps:
  i. Identify relevant stress drivers and/or mitigating factors at each horizon.
  ii. For each horizon, combine relevant indicators into a continuous composite index; predict a stress event if the index exceeds a chosen threshold. For any chosen threshold and sample period, calculate:
    - percentage of missed crises (Type I errors)
    - percentage of false alarms (Type II errors)
  iii. Divide the index into three risk zones (low, moderate, high) based on two cutoffs corresponding to probabilities of missed crises and false alarms.
- Specific cutoffs:
  - "High" risk: index exceeds the upper cut-off, corresponding to a false-alarm probability of 10 percent.
  - "Low" risk: index below the lower cut-off, corresponding to a missed-crisis probability of 10 percent.
  - "Moderate" risk: index between the upper and lower cut-offs.
- Rationale:
  - Steps (i) and (ii) use rigorous statistical procedures to maximize predictive performance.
  - Step (iii) leverages the tools’ capacity to separate stress from non-stress episodes, enabling calibration to relatively low probabilities of missed crises and false alarms (10 percent) without creating a very wide moderate zone.

### Interpretation of “High”, “Medium” and “Low” risk signals (Box 3)
- Thresholds target misclassification rates based on historical data:
  - Threshold separating low and medium: missed crisis rate of 10 percent.
  - Threshold separating medium and high: false alarm rate of 10 percent.
  - Maximum tolerance for potential misclassifications set at 10 percent.
- Ex-post crisis probabilities (average values reported):
  - Near-term tool: P(S | H) = 0.40; P(S | L) = 0.02.
  - Medium-term tool: P(S | H) = 0.43; P(S | L) = 0.04.
- Additional probability and discrimination details:
  - Lowest stress probability associated with a high risk signal is about 0.2 for both tools (index realization right at the cut-off between medium and high).
  - Highest stress probability associated with a low-risk signal is 0.09 for the near-term and <0.1 for the medium-term tool (realization at the cut-off between low and medium).
- Sensitivity and specificity reported:
  - Near-term tool sensitivity = 0.55 (80/146), specificity = 0.73 (1034/1411).
  - Medium-term tool sensitivity = 0.58 (6/11) and specificity = 0.55 (51/88).

### Role of judgment, review, and reporting requirements
- Mechanical tool outputs form starting points; staff judgment is expected to adjust assessments where appropriate, constrained by the review process and required to be transparent in SRDSA write-ups.
- Areas where judgment and customization can be applied:
  - At mechanical-signal generation: incorporate additional variables (e.g., accounting for liquid assets in the GFN module or counting them as part of international reserves in the logit (¶49)); triggered stress tests (¶63); long-term risks (¶70).
  - At staff risk-assessment generation for each horizon: staff can deviate from mechanical signals, especially in medium-term assessments where stress tests may lead to different staff assessments; the general presumption is deviations would not exceed one notch (¶68).
  - In long-term assessments: optional modules and other considerations can inform judgment (¶70).
- Reporting requirement: surveillance teams must report if the debt/GDP ratio stabilizes under the baseline.
- Probabilistic debt sustainability assessment will be required in Fund programs, optional in surveillance cases.

### Audit, validation, and modification
- Following the informal Board discussion on May 29, 2020, all analytical tools proposed were audited by an independent RES/ICD team; this paper reflects the results of that validation.
- Annexes IV and V respond to audit points and discuss framework elements modified in reaction to the audit.

### Implications of COVID-19 for the new framework (Box 4)
- COVID-19 effects:
  - Sharp output contractions and substantial spending needs have led to a surge in fiscal deficits and public debt-to-GDP ratios; timing and extent of recovery are significantly uncertain.
- Preliminary analysis and performance:
  - A preliminary analysis indicates the proposed tools performed significantly better than the current framework in predicting stress during the first four months of 2020, with a high correlation between the evolution of the mechanical risk indices and observed sovereign rating downgrades.
- Adjustments and guidance:
  - SRDSF realism tools allow some modification of inputs for the post-shock period; residual issues should be discussed in SRDSA write-ups.
  - For prospective fiscal adjustment, it may be appropriate to measure underlying primary balances excluding temporary crisis-related support.
  - When assessing output projections, it may be useful to ignore temporary contractions and rebounds associated with short-term supply effects of shutdowns.
  - During recovery, an adjustment to the fanchart methodology is warranted to limit incorrect triggers of the optimism correction mechanism (Annex VI).
- Uncertainty and re-estimation stance:
  - The debt fanchart and GFN tools rely on historical data to calibrate forecast uncertainty and are not designed to reflect temporary rises in uncertainty; staff view is that mechanical adjustment to capture increased uncertainty is not feasible now.
  - Parameters are envisaged to be ‘frozen’ until the next review; a one-off re-estimation could be considered once crisis dynamics and recovery are clearer if it likely improves out-of-sample performance.

### Debt coverage and institutional coverage
- Gross debt remains the core concept in the SRDSF, with an enhanced role for liquid assets where data quality and materiality permit; otherwise, enhanced guidance will apply (Annex II).
- General Government (GG) debt, defined per GFSM 2014, will be the default institutional coverage.
  - GG encapsulates all non-market government-controlled entities and is already the most commonly used institutional coverage in the present MAC DSA.
  - The Fiscal Monitor uses this coverage level; thus, GG is set as the explicit minimum benchmark coverage for MACs going forward, with incentives to report on this basis.

*Source: ppea2021003 (MAC DSA Review, selected sections).*

### 38.      Countries currently reporting on a central government (CG) or an incomplete GG basis

### 38.      Countries currently reporting on a central government (CG) or an incomplete GG basis

### Coverage, justification, and mandatory stress-testing
- Countries reporting on a central government (CG) or an incomplete general government (GG) basis must justify why the narrower coverage is appropriate.
- Countries with narrower coverage than GG will be subject to a mandatory contingent liability stress-test to capture omitted risk exposures.
- The contingent liability stress-test will be calibrated based on countries’ historical nonfinancial contingent liabilities per FAD database.
- The contingent liability stress-test would also be applied to GG or broader coverage countries with omitted fiscal risk exposures outside their chosen coverage.
- A situation where narrower coverage would be appropriate would be when government functions are concentrated at the CG (for example, in small states). In cases where narrower coverage is due to data shortcomings, teams will be expected to outline in staff reports the authorities’ plans to address these shortcomings, including any TA needs.

### Expanding coverage beyond GG: NFPS and CPS considerations
- The framework allows expanding coverage beyond GG to fully capture sovereign risks and potential mitigants.
- Full or partial non-financial public sector (NFPS) coverage could be appropriate if it:
  - captures material fiscal risks from SOEs;
  - aligns with national legislative requirements;
  - anchors policy discussions (including the production of official statistics at this level).
- Inclusion of public banks involved in quasi-fiscal activities could warrant a full or partial consolidated public sector (CPS) concept.

### Central bank consolidation and treatment of central bank liabilities
- The framework proposes consolidations only in cases of central banks with large negative capital positions and/or where the country team considers the central bank to be involved in significant direct monetary financing of the budget and/or quasi-fiscal activities.
- Such consolidation would imply that (i) central bank claims on the government are netted out and (ii) central bank debt liabilities (excluding currency and deposits held by residents) are added.
- In other cases, the framework will incorporate mitigating factors from central bank holdings of government debt into its analysis without consolidation (Annex II).
- Some central bank liabilities could represent material risk and warrant risk-based inclusion even when the central bank is not consolidated with the government. This includes central bank bilateral FX swap liabilities that do not represent normal central bank monetary or liquidity operations, or are not extinguishable by the central bank without actions detrimental to government debt levels (Annex II).

### Fund credit treatment in SRDSA
- Fund credit disbursements will continue to be included as public debt for SRDSA purposes.
- Rationale:
  - Fund credit is legally an obligation of the member country and thus must always be included in a public debt sustainability analysis.
  - Economically, Fund credit can be disbursed either to finance the budget or to raise FX reserves to a safe level. In both cases, Fund money substitutes for other types of sovereign borrowing and is conceptually a mirror image of the counterfactual change in the public debt ratio that would have happened in any case.
- IMF financing can mechanically affect the assessment of debt risks and sustainability through four channels:
  1. the debt terms in the proposed logit model;
  2. the FX reserves term in the logit model;
  3. the debt level and index in the fanchart tool;
  4. creditor composition, which enters the GFN tool.
- The net mechanical effect of IMF financing on assessed risks depends on the counterfactual; program assumptions and conditionality also affect the macroeconomic baseline and should unambiguously improve sustainability via external stabilization, lower borrowing costs and possibly higher growth.

### Contingent liabilities and inclusion in projections
- Inclusion of contingent liabilities in debt projections should reflect the likelihood of their materialization.
- Contingent liabilities are not generally expected to be included in GG debt.
- If teams can anticipate and estimate their materializations (e.g., uncalled government guarantees, legal settlements, bank recapitalization needs), these should be included in the debt projections, with a corresponding adjustment in the contingent liabilities stress-test.
- Government guarantees should be included fully as public debt if there is a high likelihood of their materialization (consistent with GFSM principles).
- Liabilities of government-controlled non-market SOEs or SPVs recognized as part of the GG per GFSM principles should be part of GG debt; if not, they should be added manually to the debt definition used in the SRDSF.
- Illustrative practice: use of an “augmented” debt measure for China which includes the debt of Local Government Financing Vehicles and other government funds that, although legally separate from the government, perform government functions. Similar criteria have been used in other countries (Belgium, Brazil, Russia, United Kingdom).

### Strengthened debt disclosures and profile breakdowns
- The new framework requires strengthened debt disclosures to support evenhandedness and help achieve greater harmonization over time.
- SRDSA would be expected to include metadata on:
  - institutional and instrument coverage;
  - the valuation method (nominal, face, or market value).
- Where available, this will be accompanied by a consolidation table showing:
  - gross debt outstanding by level of government;
  - cross-holdings that are consolidated away;
  - final consolidated debt position (Appendix I).
- Enhanced reporting of debt profile vulnerabilities is proposed. The new framework will include:
  - a breakdown of domestic/foreign law debt and marketable/non-marketable debt;
  - the currency composition of FX debt;
  - additional holder profile information (foreign official, foreign private, domestic central bank, domestic commercial banks, and domestic nonbank) drawing on the Arslanalp-Tsuda database.
- Inclusion of holder information is essential to allow an empirically grounded analysis of rollover risks.
- At the time of a program request, holder profile of debt amortization is particularly important; while not currently compiled, it can be approximated with cooperation from country authorities (Annex VII). Provision of this information will bring MAC DSA reporting closer to LIC DSF practices.

### Enhanced realism tools for baseline projections
- Proposal to expand existing realism tools (covering growth, inflation and primary balance) to encompass all debt drivers—including:
  - exchange rates;
  - financing terms on external borrowing;
  - stock-flow adjustments;
  - public debt itself.
- New tools to assess the realism of assumed fiscal multipliers and potential growth rates in light of evidence indicating systematic bias in output gap estimates (Annex I).
- Comparator buckets will be defined to account for differences in forecast error distributions across commodity and non-commodity exporters, and surveillance vs. program cases.
- Tools are intended to:
  - put the baseline in context;
  - illuminate key assumptions that differ from past and cross-country experience;
  - encourage early use by teams as part of the iterative process of producing the baseline macro framework.
- Where tools flag differences, teams should explain country-specific factors in the SRDSA; where unexplained, a re-examination and possible revision of macro projections may be warranted.
- Adjustments are proposed to ensure the tools remain relevant following the COVID-19 shock.

### Tool for near-term risk analysis: multivariate logit model specification and diagnostics
- Staff proposes a multivariate logistic regression as the workhorse tool for near-term risk analysis to produce a single continuous probability of stress 1–2 years ahead.
- The proposed logit specification:
  - includes a parsimonious set of 10 regressors;
  - organizes regressors in five buckets: institutional quality; stress history; cyclical; debt burden and buffers; and global.
  - regressors were selected after consulting the empirical literature, extensive statistical testing, and internal and external peer reviews (Annex V).
- Economic interpretation of buckets:
  - Institutional quality and stress history capture country heterogeneity and relate to “debt carrying capacity”.
  - Cyclical indicators capture country-specific buildup of vulnerabilities (external position, financial sector, fiscal weakening).
  - Debt burden and buffer indicators capture vulnerabilities associated with debt level/dynamics/structure and the risk-mitigating role of reserves.
  - Global variables proxy investor risk appetite; change in VIX proxies change in global tolerance for risk-taking.
- Model diagnostics and estimated statistics:
  - Number of Observations: 1,579
  - LR chi2: 246.70
  - Pseudo R2: 0.25
- Estimated coefficients (Coeff. and Std. Coeff.) by regressor:
  - Institutional Quality: Coeff. -1.073 *** ; Std. Coeff. -0.377
  - Stress History: Coeff. 0.514 *** ; Std. Coeff. 0.1006
  - Current account balance/GDP: Coeff. -0.024 ** ; Std. Coeff. -0.095
  - REER (3-year change): Coeff. 0.013 ** ; Std. Coeff. 0.070
  - Credit/GDP gap (t-1) (if + ve): Coeff. 0.086 *** ; Std. Coeff. 0.258
  - Δ(Public debt/GDP): Coeff. 0.052 *** ; Std. Coeff. 0.1182
  - Public debt/revenue: Coeff. 0.002 *** ; Std. Coeff. 0.1213
  - FX public debt/GDP: Coeff. 0.024 *** ; Std. Coeff. 0.1601
  - International reserves/GDP: Coeff. -0.034 *** ; Std. Coeff. -0.2348
  - ΔVIX (global): Coeff. 0.015 *** ; Std. Coeff. 0.1373
- Relative importance example from staff note:
  - The standardized coefficient for the FX public debt to GDP is about 1.4 times the magnitude of the coefficient for the change in public debt-to-GDP. This implies that ceteris paribus, a 1 standard deviation higher FX public debt-to-GDP ratio (about 16.8 percent of GDP, see Table AV.5) would have roughly the same effect on the stress probability as a 1.4 standard deviation increase in change in public debt-to-GDP (approximately 7.5 percent of GDP, see Table AV.5).

*Source: ppea2021003 - 38.      Countries currently reporting on a central government (CG) or an incomplete GG basis*

### 48.      The estimated Logit Stress Probability (LSP) will be used to assign countries to low-,

### The estimated Logit Stress Probability (LSP) will be used to assign countries to low-, moderate- and high-risk of sovereign stress

### Near-term mechanical risk signal (Logit LSP)
- Key design:
  - The LSP assigns countries to low-, moderate-, and high-risk of sovereign stress (the “near-term mechanical risk signal”).
  - Low- and high-risk cutoffs are calibrated to keep the rate of missed crises and false alarms at 10 percent, respectively.
  - Stress probability cutoffs: 9 percent (threshold between low and moderate) and 20.5 percent (threshold for high risk).
  - Interpretation:
    - Low risk signal: model estimates probability of near-term sovereign stress < 9 percent.
    - High risk signal: model estimates probability of near-term sovereign stress > 20.5 percent.
  - Model output includes the estimated logit stress probability in each individual case.
- Use in policy messaging:
  - Tool issues a signal of sovereign stress 1–2 years ahead to allow time to adjust policies.
  - Template will show contributions of the five regressor buckets to the year-on-year increase in the LSP to support discussion of policy actions to minimize corresponding risk sources.
  - Estimated probability of stress will be reported against contemporaneous values for peers (teams have discretion over peer buckets).

### Special cases and guidance for near-term tool
- Guidance will address special cases not adequately captured by the standard model (e.g., use of liquid assets and treatment of “safe havens”).
- Model default mutes the share of currency union MACs in stress (a regional spillover variable) but it can be “switched on” if country teams consider spillover risks within the currency union material.

### Medium-term risk analysis — overview
- Medium-term toolkit comprises three modules analyzing solvency and medium-term rollover risks:
  1. Improved debt fanchart (probabilistic assessment and robustness to baseline optimism).
  2. Gross Financing Needs (GFN) module analyzing financing risks given the debtholder profile.
  3. Triggered/tailored stress-tests for specific country vulnerabilities.

### A. Debt fanchart — two-step procedure and use for medium-term stress prediction
- Two-step procedure:
  - Step 1: Compare team’s baseline with a “historical fanchart” (diagnostic for baseline realism).
  - Step 2: Produce final fanchart based on Step 1 results:
    - i. If no optimism issue: generate a “standard” (symmetric) fanchart.
    - ii. If optimism issue: generate a “realism adjusted” (asymmetric) fanchart.
  - During the Covid-19 recovery phase (2021-22), a modified historical fanchart partly dependent on the team’s baseline is proposed to limit incorrect realism corrections.
- Predictive metrics evaluated over 2010–17; three metrics stood out:
  - Probability that debt does not stabilize in the medium-term: probability that projected baseline primary balance at t+5 < balance required to stabilize t+5 debt.
  - Debt level at t+5, controlling for debt-carrying capacity: product of debt level at t+5 and the institutional quality index used in the near-term tool.
  - Fanchart width: distance between the 5th and 95th percentiles of the debt distribution in the final projection year (t+5).
- Aggregation and risk classification:
  - The three metrics are aggregated into a composite Debt Fanchart Index (DFI) weighted by predictive power.
  - DFI classifies countries into low-, medium-, and high-risk zones using thresholds calibrated to a 10 percent missed crisis and false alarm rates.
- Guidance on special cases:
  - Adjustments for countries where public sector holds large financial assets, countries with structural breaks, and countries close to or recently reached a debt restructuring agreement.
- Output:
  - DFI value reported on a continuous scale of past DFI values associated with stress and non-stress episodes; comparisons with peers included.
  - Output designed to show relative contributions of the three metrics to the DFI to inform policy adjustments.

### B. Gross Financing Needs (GFN) Module
- Rationale:
  - Ability to meet financing needs varies with domestic investor base size, investor class stability, and government asset buffers.
- Three-step analysis (see Annex VII for methodology):
  - Projecting GFNs and financing:
    - Start from GFN-to-GDP projections.
    - Teams allocate domestic issuance among central bank, commercial banks, and other private creditors; foreign issuance divided among foreign official and foreign private creditors.
    - A default disaggregation preserving debtholder stock shares would be available to teams.
  - Modeling risk:
    - Generalized stress scenario with adverse shocks in: (i) macro-fiscal variables (similar to combined macro-fiscal shock), (ii) maturities (shortening of maturity in stress), and (iii) access to external debt markets (loss of foreign private investor appetite).
    - Default holder shock: rollover rates of foreign investors drop to 67 percent, and these investors are unwilling to finance any additional new borrowing for a period of two years.
  - Analyzing residual financing:
    - Stress scenario creates a need for financing when the pool of potential creditors has shrunk; residual financing need assumed absorbed by domestic banks unless customized.
    - Customizations available to account for government liquid asset holdings and the domestic non-banking sector’s role.
- Bank absorption metrics (three indicators combined into GFN Financeability Index, GFI):
  - Initial bank claims on the government expressed as a percentage of banking system assets.
  - Maximum cumulative change over the projection period (size and “financeability” of demand on banks).
  - Average level of GFN-to-GDP ratios over the projection horizons (use of an average to distinguish persistent high GFNs).
- Aggregation and classification:
  - Indicators combined into aggregate GFI weighted by explanatory power.
  - GFI divided into three risk zones (high, moderate, low) using thresholds associated with a 10 percent missed crisis and false alarm rates.
- Standardized customizations and guidance address:
  - Use of government financial assets to offset funding pressures.
  - Role of the domestic non-bank financial sector as residual creditor.
  - When non-bank financial intermediaries are a source of government funding risk.
  - Timing of onset of stress and use of granular information.
- Policy translation:
  - For moderate- or high-risk signals, recommendations could focus on fiscal adjustment or debt management (e.g., maturity extension) to contain GFN levels and bank claims.
  - Discussion may include holder structure and ways to finance the same baseline GFN in less risky ways.

### C. Triggered stress-tests for country-specific vulnerabilities
- Purpose:
  - Simulate debt and GFN paths under major specific shocks to capture country heterogeneity and tailor scenario analysis and policy advice.
  - Teams encouraged to customize default parameters based on historical and empirical evidence.
- Role relative to other tools:
  - Complement fanchart (which may not capture future risks absent from country history) and GFN generalized stress scenario (triggered tests focus on specific tail risks).
- Five specific vulnerabilities proposed for triggered tests (default shocks and customization options):
  - Contingent liabilities related to narrow coverage
    - Trigger: Countries with shortfalls between actual and appropriate debt coverage.
    - Default shock/design: Default set to zero if coverage is deemed appropriate; otherwise the average fiscal cost of observed nonfinancial contingent liabilities per FAD database.
    - Customization: Tailor default parameters based on country-specific information.
  - Banking crisis
    - Trigger: Existing mechanical criteria and vulnerable countries per MCM FSI heatmaps.
    - Default shock/design: Fiscal cost implied from loss of 10 percent of financial assets (as in the existing framework), after assuming tier 1 capital.
    - Customization: Tailor fiscal cost based on results of stress-tests implemented in FSAPs.
  - Natural disasters
    - Trigger: Countries at risk of natural disasters per past damages, exposure and vulnerability.
    - Default shock/design: One-off shock to public debt-to-GDP ratio (+7 percent of GDP) and real GDP growth (-5 ppts) representing the worst quintile impact of staff event analysis of past natural disasters.
    - Customization: Tailor parameters to account for recent structural or financial resilience initiatives and whether baseline already assumes some event.
  - Commodity price shock
    - Trigger: Commodity exporters, and vulnerable commodity importers.
    - Default shock/design: Shock fiscal revenues (expenditures) to GDP ratios by -0.5 (+0.9) ppts for each 10 percent decrease (increase) in export (import) prices for commodity exporters (importers). Additionally, shock baseline interest rate and growth.
    - Customization: Tailor parameters based on GRAMs and country-specific information (including revenue collateralization and state-contingent instruments where relevant).
  - REER shock
    - Trigger: Countries with high initial over-valuation, and REER change over t-t+5 insufficient to reduce overvaluation <5%.
    - Default shock/design: A nominal one-off depreciation sufficient to close the country’s overvaluation gap over the projection horizon.
    - Customization: Tailor size or duration of shock depending on country-specific information.
- Notes on commodity price shock mechanics:
  - To shock commodity prices, staff will use a price gap defined as the difference between the baseline commodity price and a +1 (-1) SD for importers (exporters) using RES commodity (fuel and non-fuel) price fancharts.
  - Fiscal impacts reported come from Spatafora and Samake (2012). One may consider shocking expenditures in exporters with large subsidies.

*Source: MAC DSA Review, excerpted content.*

### 66.      The results from the debt fanchart and GFN module are aggregated into one overall

### 66.      The results from the debt fanchart and GFN module are aggregated into one overall 

### Medium-term aggregation and the Medium-Term Index (MTI)
- The debt fanchart and GFN module results are aggregated into one overall medium-term risk signal capturing both solvency and liquidity risks.
- Rationale:
  - The composite signal is more informative and accurate than either measure alone.
  - Neither underlying measure reflects pure solvency or liquidity risk: the fanchart mainly captures solvency but its width reflects past behavior of interest rates (liquidity); the GFN mainly captures liquidity but is influenced by solvency-related variables such as debt level.
- Construction:
  - The Medium-Term Index (MTI) is created as the average of the DFI and GFI.
  - The simple average is used because DFI and GFI have roughly similar predictive capacities of stress (as measured by the respective AUCs) and are thus equally weighted.
  - The MTI is divided into three risk zones (high, moderate, low) using the same approach as for the GFI and DFI, generating a single medium-term mechanical signal.
  - A low-risk signal can be interpreted as indicating a single-digit probability of sovereign stress; a high-risk signal indicates a stress probability in excess of 20 percent (see ¶32 and Box 3).
- Judgmental adjustment via stress-tests:
  - Triggered stress-test results can modify the MTI final assessment when all of the following hold:
    - (i) the corresponding scenario has a high probability in the RAM;
    - (ii) the scenario is judged to not be fully captured by the fanchart; and
    - (iii) the scenario generates a debt trajectory above the 75th percentile of the fanchart.
  - In such cases there is a presumption that the final MT assessment would be one notch worse than the mechanical signal, though the extent of correction remains judgment based.

### Long-term risk tools and 10-year horizon
- Horizon extension:
  - A 10-year horizon is proposed for debt and GFN projections to provide confidence around debt and GFN levels/trajectories after 5 years.
  - The 10-year horizon covers the Fund repayment period and is essential for analyzing debt restructuring cases.
  - Extension is limited to the baseline; uncertainty around baseline projections (fancharts and stress-tests) is not modeled after 5 years.
  - Guidance will explain how teams can extend projections to 10 years in an internally consistent and evenhanded manner.
- Optional long-term modules (qualitative inputs; no mechanical signals):
  - Teams will be required to analyze key risks where identified and report judgment as low, moderate, or high (reporting numerical estimates optional).
  - Covered risk categories:
    - Population aging, including pension and social security liabilities, and health care costs.
    - Scaling up/down of natural resource extraction affecting resource revenues beyond five years.
    - Large debt amortizations beyond the five-year horizon.
  - Parameterization, testing, and interpretation guidelines to be developed in the Guidance Note.
- Climate change long-run reporting:
  - Teams should report long-run public finance consequences of climate change for countries facing existential threats or high vulnerability.
  - Suggested analysis: discuss implications for key macroeconomic variables (growth, public spending) and, where feasible, derive debt ratio implications over a period of 30 years.

### Tools to assess long-term risks (summary of modules in Table 6)
- Pension and social security benefits:
  - Trigger: Countries with a deficit in the social security program, either currently or in the next 10 years; or significant population aging.
  - Output: Future pension fund reserves and/or yearly pension fund liabilities in periods t+6 to t+16, benchmarked against criteria.
- Pressures from health costs:
  - Trigger: Countries with sizable public health expenditures likely to grow rapidly from demographic change or excess cost growth.
  - Output: PV of health expenditures, net of health revenues and/or path for health expenditures in periods t+6 to t+10 under various excess cost growth paths.
- Natural resource volume changes:
  - Trigger: Countries where extraction volumes over t+6 to t+15 deviate by more than one standard deviation from the historical average (last 10 years).
  - Output: Projected change in debt and/or primary balance at t+10 compared to t+5, assuming constant natural resource revenues and non-interest expenditures.
- Large debt amortizations:
  - Trigger: Debt amortizations in t+6 to t+25 above historical average over the last ten years plus one standard deviation.
  - Output: Average yearly deposit over the period t+1 to maturity needed to meet the obligation, expressed in millions and in percent of GDP.
- Climate change:
  - Trigger: Countries with existential or high vulnerability to climate change per exposure, susceptibility and adaptive capacity.
  - Output: Discuss impact on key macro-fiscal variables and where feasible draw out 30-year implications for the debt.

- Note: As of end-2018, two indices identify 21 countries for high risk (listed in the source).

### Deriving sustainability assessments
- Use of proposed tools in Fund-supported programs (including Exceptional Access):
  - Medium-term debt stabilization prospects quantified using the debt fanchart composite index.
  - Assessments of rollover risk informed by the GFN module, taking all components of program financing into account (including prospective Fund disbursements).
  - A crisis prediction model calibrated on past episodes of unsustainable debt will help predict crises associated with unsustainable debt (defaults and restructurings) rather than just sovereign stress.
- Aggregation and three-way mechanical signal:
  - Outputs from the three modules are aggregated into a three-way mechanical signal on debt sustainability: sustainable with high probability; sustainable, but not with high probability; not sustainable.
  - Team judgment and robust review processes will complement mechanical signals to produce a bottom-line assessment.
  - Interpretation guidance:
    - High risks across tools suggest concerns about sustainability.
    - Low risks across tools suggest sustainability with high probability.
    - Middle outcomes suggest sustainability, but not with high probability.
  - Calibration aims to limit false alarms and missed crises and ensures:
    - (1) the signal “not sustainable” is associated with a probability of an unsustainable event of more than 50 percent;
    - (2) the signal “sustainable with high probability” is associated with a probability of an unsustainable event of less than 20 percent.
  - Because methodology is potentially market-sensitive, precise aggregation method and index cutoffs will remain confidential.
- Application in surveillance and restructurings:
  - Surveillance: Sustainability assessments not required; if provided, additional scenarios needed to anchor them. To conclude sustainability in absence of program baseline, staff must show metrics are satisfied under baseline or under an alternative politically and economically feasible scenario. To conclude not sustainable, staff must show metrics are not satisfied even by policies considered “at the frontier” of feasibility.
  - Debt restructurings: Tools can set targets to restore sustainability. Use a longer 10-year horizon. GFN targets are a starting point to verify financing needs are manageable. Post-restructuring debt trajectories derived from new debt structure should attain GFN targets and be analyzed through the fanchart module to ensure high probability of stabilization. If not achieved, adjust debt relief envelope until both modules signal sustainability. Judgment and complementary targets for specific vulnerabilities remain.

### Reporting requirements and disclosure options
- Applicability:
  - Sovereign risk analysis required for all members in both surveillance and program contexts.
  - Sustainability assessments required for arrangements involving GRA resources (including precautionary arrangements) and for the PCI.
  - Non-program countries: sovereign risk analysis needed at the time of the Article IV consultation.
  - For countries with Fund arrangements involving GRA resources: both sovereign risk and debt sustainability analyses are required at program approval and annually thereafter (unless developments warrant more frequent analysis). Exceptional access cases require an updated DSA (with three-zone sustainability assessment) in every program review.
- Reporting in staff reports (program context):
  - Staff reports with a DSA must include a sovereign risk analysis and potentially a three-zone sustainability assessment.
  - Risk of sovereign stress: reports will contain full range of outputs for medium and long-term (not near-term); an overall risk assessment synthesizing across horizons will be reported. In precautionary arrangements, reporting based on program status (see Box 5).
  - Disclosure options for debt sustainability assessments:
    - Option a: No change to current practice—three-zone assessment included in staff reports in exceptional access cases but not in normal access cases.
    - Option b: Disclosure to the Board in both normal and exceptional access cases, but public disclosure only in exceptional access cases. Option b would require a modification to the Transparency Policy (to permit blanket deletion of a particular set of information in normal-access staff reports).
- Reporting in surveillance context:
  - Sustainability assessment publication optional.
  - Two disclosure approaches for sovereign risk analysis:
    - Full disclosure: publish outputs of the risk assessment framework in staff reports; deletion of market-sensitive signals still possible case-by-case under Transparency Policy.
    - Full disclosure to the Board but limited public disclosure (e.g., omit near-term risk signal/assessment); would require Transparency Policy modification.
  - If Board chooses limited public disclosure options, staff proposes reconsidering public disclosure after 12 months.
- SRDSA presentation:
  - The SRDSA included in staff reports will be sharper and reduce need for lengthy write-ups; Appendix I shows a mock SRDSA under a ‘full disclosure’ setting (per source).

### Precautionary arrangements (Box 5 highlights)
- Precautionary arrangements: baseline assumes no drawing; risk assessments for these arrangements informed by baseline.
- Sustainability assessments for precautionary arrangements informed by baseline and, when appropriate, by an adverse or full-draw stress scenario.
- Running sustainability assessment on a full-draw scenario is appropriate in:
  - exceptional access cases;
  - if shocks that may trigger drawing are not adequately captured by medium-term tools; or
  - when review departments have doubts about baseline realism that cannot be resolved.
- Program design implication: a high near-term risk under baseline before/at approval may indicate a precautionary arrangement is inappropriate and a financing arrangement may be more suitable.

*Source: ppea2021003 - 66.      The results from the debt fanchart and GFN module are aggregated into one overall*

### 82.      The implementation of the new framework will be accompanied by close engagement

### ppea2021003 - 82.      The implementation of the new framework will be accompanied by close engagement

### Implementation plan and stakeholder engagement
- Developing materials to support implementation (expected to be available by the second half of 2021): guidance note, software and Excel files underpinning the new framework; settling remaining issues including adaptation to country-specific circumstances, evenhandedness across countries, and public communication.
- Early engagement with a subset of country teams to test new tools and templates, running them in parallel with the current framework to inform the staff guidance note and template design.
- Close engagement with country authorities and debt management offices to illustrate features ahead of rollout; after formal Board approval staff will reach out to country officials, including during the 2021 Spring Meetings.
- Outreach to debt management offices at an early stage to ensure new files are easy to use, intuitive, and free from bugs.
- Engagement with external stakeholders (civil society organizations, think tanks, private creditors) during the 2021 Spring and Annual Meetings.

### New reporting format (Box 6)
- Published DSAs under the proposed SRDSF could be no more than 7–8 pages (shorter than the 10-page average at present).
- SRDSA composition (standardized tables and charts with staff commentaries):
  - i. Upfront table summarizing the risk and sustainability assessments and if and how judgment has been applied, followed by a SRDSA chapeau.
  - ii. Tables summarizing debt coverage disclosures and related information; figure on composition of public debt by currency, holder, legal basis, level of public sector, and maturity.
  - iii. Table reporting debt and GFN projections and debt drivers, with staff commentary highlighting important trends or features.
  - iv. Figure reporting results of the realism tools with staff commentary explaining any red flags.
  - v. A page each for outputs of the near-, medium-, and long-term risk analyses and derivation of mechanical risk signals; cross-country comparisons and stress-test results reported here.

### Transition timing and operational support
- The current framework will continue to be used between Board approval and the rollout of the new framework, expected for Q4 2021 or Q1 2022.
- If data availability challenges arise before rollout, an implementation team from SPR will provide support to country teams; as a last resort some tools may be adjusted to accommodate data limitations.
- In program cases, transition management:
  - For programs approved prior to rollout but expiring after rollout: SPR and area departments will examine differences between frameworks and make necessary adjustments; adjustments through judgment under the current framework are appropriate if the new framework is valid given new analyses or compelling conditions; tool or application issues could be addressed in the guidance note.
  - Programs expiring prior to rollout are not affected; successor programs approved after rollout will be assessed under the new framework and communications with authorities must be managed accordingly.
- Guidance note will be adjusted if transition experiences raise concerns about limitations of the new framework.

### Data requirements, resource implications, and transitional arrangements
- New data requirements introduced by the framework will be contained by automation of data sources, gradual implementation supported by Fund TA, and may be offset by substantially reduced, focused writeup requirements.
- Most needed data is already required to run the current framework; additional data will be linked to existing centralized datasets where possible.
- Review team ran new core elements on nearly every MAC, confirming feasibility.
- A limited set of new debt profile data will need to be collected by country teams in consultation with country authorities (Annex VIII).
- SPR implementation team will support country teams during transition, replicating practices used in LIC DSF adoption.
- Staff will work with authorities, including through TA, to prepare for additional data requirements.
- Where expanding debt coverage takes time, the contingent liability stress-test can analyze risks while statistical capacity is strengthened.
- Transitional solutions will be developed for frontier LICs or recent PRGT graduates in the guidance note (examples: use of parameters calibrated on peer countries, simplifications of standard tools).

### Training and capacity building (Box 7)
- Comprehensive training before and during rollout—internal training to enable staff to explain the framework to country authorities; training materials for country officials and potentially external stakeholders through onsite and online courses and seminars.
- Guidance note development after formal approval; staff will conduct additional outreach to Executive Directors and submit the completed draft for the Board’s information. Under usual process Executive Directors will have 2 weeks to call a meeting to discuss it; thereafter the guidance note will become operational. This step will likely take [6-8] months.
- Seminars at the Spring Meetings to present new tools and sensitize country officials.
- Close contact between authorities and country teams upon rollout to ask technical questions and apply tools to country specifics.
- Training courses, such as those under the Debt Management Facility (DMF), will be developed and delivered as framework goes live; additional training events beyond the DMF may be possible but may need financing.
- All training materials (presentations, illustrative example, user’s manual) would be posted to the Fund’s public website and available to all country officials.
- With General Government as expected standard coverage for public debt in the MAC DSA, some countries will require support to develop needed statistical capacity; this is expected to be gradual and Fund TA will continue to be made available.

### Illustrative example — Ruritania (Appendix)
- Overall final assessment: Moderate; staff assessment reflects mitigating role of a large liquid asset buffer and impact of scaling-up of natural resources.
- Near term:
  - Mechanical signal: High
  - Final assessment: Moderate — mitigating role of a large liquid asset buffer cited.
- Medium term:
  - Mechanical signal: High
  - Final assessment: Moderate — mitigating role of a large liquid asset buffer despite mechanical medium-term signal indicating “high” risk, largely attributable to the debt fanchart’s width. Commodity price stress-test did not generate a debt trajectory above the 75th percentile of the debt fanchart.
  - Reported components: GFN: Moderate; Fanchart: High; Stress-test: Yes (commodity price shock)
- Long term:
  - Mechanical signal: Low
  - Final assessment: Triggered module indicates scaling up/down of natural resources leads to lower debt and higher primary balances long term (after new oil and gas fields come on stream).
- DSA Summary Assessment (selected numerical and factual points):
  - Large realized and projected increase in public debt to finance investments in the gas and oil sectors corresponding to US$20 billion or 25 percent of nominal GDP.
  - Substantial liquid assets equal to 70 percent of total debt were cited as an important mitigating factor.
  - Additional revenue expected from new gas and oil fields coming on stream after t+5 of about US$40 billion at current gas prices.
  - Note: Sustainability Assessment not required as country is not in Fund-supported program; risk of sovereign stress reported only for surveillance cases; sustainability assessment required only for program cases and optional in surveillance.

### Issues for Discussion (selected)
- Continued application of the existing definition of debt sustainability (as previously adopted by the Board)?
- Endorse naming the proposed framework “Sovereign risk and debt sustainability framework”?
- Agreement with enhanced debt coverage and the proposed 10-year horizon?
- Agreement with the proposed realism tools and realism adjustments?
- Agreement with the proposed horizon-based approach?
- Support for the use of standardized tools to provide mechanical risk signals at each horizon, with judgment added for country-specific information?
- Inclusion of sovereign stress signals and assessments in reports for program cases, with near-term sovereign stress signals only in reports for precautionary programs?
- Disclosure of three-zone debt sustainability assessments in program cases — preference between:
  - a. continuing current practice of including three-zone-assessments only in staff reports on exceptional access cases, or
  - b. disclosure to the Board in both normal and exceptional access cases but to the public only in exceptional access cases, requiring a change in the Transparency Policy?
- Publication of sovereign stress signals in surveillance reports — preference between:
  - a. full disclosure to both the Board and the public, or
  - b. full disclosure to the Board but limited disclosure to the public, requiring a change in the Transparency Policy?
- Support for proposed use of new tools to guide debt restructurings?
- Endorsement of the proposed DSA reporting format?
- Support for the proposed timeline for implementation of the framework?

*International Monetary Fund — excerpt from ppea2021003, pages 44–50*

### 5. Debt consolidation across subsectors:

### 5. Debt consolidation across subsectors:

### Debt consolidation coverage and recording
- Color code legend (as presented): chosen coverage; missing from recommended coverage; not applicable.
- Subsector labels: CG = Central Government; GG = General Government; NFPS = Nonfinancial Public Sector; PS = Public Sector.
- Valuation and recording notes:
  - Market value of debt instruments is the value as if they were acquired in market transactions on the balance sheet reporting date. Only traded debt securities have observed market values. 7/
  - Stock of arrears could be used as a proxy in the absence of accrual data on other accounts payable (OAP). 2/
  - Insurance, Pension, and Standardized Guarantee Schemes (IPSGSs) typically include government employee pension liabilities. 3/
  - Includes accrual recording, commitment basis, due for payment, etc. 4/
  - Nominal value at any moment in time is the amount the debtor owes to the creditor. It reflects the value of the instrument at creation and subsequent economic flows (such as transactions, exchange rate, and other valuation changes other than market price changes, and other volume changes). 5/
  - The face value of a debt instrument is the undiscounted amount of principal to be repaid at (or before) maturity. 6/

### Ruritania: public sector debt structure (staff commentary and indicators)
- Staff commentary summary:
  - Public debt at end-2018 was mostly denominated in foreign currency, of long-term maturity and issued under foreign law.
  - Debt is mostly held by foreign banks.
  - Over the medium-term, the share of foreign currency debt is projected to increase.
  - Residual maturity (2017): 16.8 years.
  - Large liquid asset buffer noted (70 percent of debt), referenced in risk assessment adjustments.

- Key indicator highlights shown in figures:
  - (a) Debt by currency (percent of GDP) — historical and projection series by foreign currency vs local currency.
  - (b) Debt by holder (percent) — central bank, domestic banks, other domestic, other foreign, foreign official.
  - (c) Debt by legal basis (2017) (percent of GDP) — bilateral & multi-lateral 1/, domestic law, foreign law (market debt).
  - (d) Evolution of CG Debt by instruments (percent of GDP) — CG marketable debt vs CG non-marketable debt, projections.
  - (e) Debt by maturity (percent of GDP) — ≤1 year, 1-5 years, 5+ years; projections.

1/ Includes SDR allocations, holdings by regional financing arrangements, and currency union central banks, where applicable.

### Baseline scenario (Table A2) — summary of key projected magnitudes and dynamics
- Staff commentary summary:
  - Public debt at end-2017 stood at 53.5 percent of GDP and its projected trajectory is expected to largely follow developments of the oil and gas sectors.
  - Over the next 5 years, debt would accumulate reflecting associated investment expenditures; then debt is expected to decline due to additional revenue from oil and gas projects coming on stream.
  - The project is profitable overall, so cumulative primary balances over the next decade do not impart a net upward contribution to debt. However, higher borrowing costs imply that interest makes the principal positive contribution to debt ratios.
  - GFNs remain relatively stable for the next 10 years.

- Selected rows from Table A2 (Actual / Projections; in percent of GDP unless otherwise indicated):
  - Nominal gross public debt: 53.5 58.7 61.1 62.2 65.1 69.3 73.1 70.0 62.4 54.8 47.2 39.6
  - Change in gross public sector debt: 6.7 5.1 2.5 1.1 3.0 4.1 3.9 -3.1 -7.6 -7.6 -7.6 -1.3
  - Identified debt-creating flows: 7.7 8.3 5.0 3.7 5.2 5.8 6.1 -0.9 -5.4 -5.4 -5.4 1.1
  - Primary deficit: 7.2 5.0 4.9 3.3 4.1 4.7 4.9 -2.1 -6.6 -6.6 -6.6 -0.1
  - Primary (noninterest) revenue and grant: 37.1 37.2 36.3 37.6 36.7 35.6 34.7 41.7 46.2 46.2 46.2 46.2 40.4
  - Primary (noninterest) expenditure: 44.3 42.2 41.2 40.9 40.8 40.3 39.6 39.6 39.6 39.6 39.6 39.6 40.3
  - Automatic debt dynamics 5/: 0.8 3.2 0.1 0.4 1.0 1.1 1.2 1.2 1.2 1.2 1.2 1.2 1.2
  - Of which: real interest rate: 1.8 3.4 3.5 1.7 1.6 2.0 2.3 2.3 2.3 2.3 2.3 2.3 2.4
  - Of which: real GDP growth: -1.0 -0.2 -3.4 -1.3 -0.5 -0.9 -1.1 -1.1 -1.1 -1.1 -1.1 -1.1 -1.2
  - Exchange rate depreciation 7/: 0.0 ....................................
  - Other identified debt-creating flows: -0.3 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
  - Residual, including asset changes 8/: -1.0 -3.2 -2.6 -2.7 -2.2 -1.7 -2.2 -2.2 -2.2 -2.2 -2.2 -2.2 -2.3
  - Gross Financing Need: 14.8 15.2 13.7 12.6 15.5 13.0 13.0 13.0 13.0 13.0 13.0 13.0 13.5

- Memo items (selected):
  - Real GDP growth (in percent): 2.2 0.3 5.9 2.1 0.9 1.4 1.7 1.7 1.7 1.7 1.7 1.7 1.9
  - Inflation (GDP deflator, in percent): 0.9 -2.4 -2.3 1.2 1.8 2.0 2.0 2.0 2.0 2.0 2.0 2.0 1.1
  - Nominal GDP growth (in percent): 12.0 -2.1 3.5 3.3 2.7 3.4 3.8 3.8 3.8 3.8 3.8 3.8 3.0
  - Effective interest rate (in percent) 4/: 4.9 3.9 3.7 4.0 4.4 5.2 5.5 5.5 5.5 5.5 5.5 5.5 4.9

- Footnotes (selected):
  - 1/ Public sector is defined as general government.
  - 4/ Defined as interest payments divided by debt stock (excluding guarantees) at the end of previous year.
  - 5/ Derived as [(r - π(1+g) - g + ae(1+r)]/(1+g+π+gπ)) times previous period debt ratio, with r = interest rate; π = growth rate of GDP deflator; g = real GDP growth rate; a = share of foreign-currency denominated debt; and e = nominal exchange rate depreciation (measured by increase in local currency value of U.S. dollar).
  - 6/ The real interest rate contribution is derived from the numerator in footnote 5 as r - π (1+g) and the real growth contribution as -g.
  - 7/ The exchange rate contribution is derived from the numerator in footnote 5 as ae(1+r).
  - 8/ Includes asset changes and interest revenues (if any). For projections, includes exchange rate changes during the projection period.
  - 9/ Assumes that key variables (real GDP growth, real interest rate, and other identified debt-creating flows) remain at the level of the last projection year.

### Realism of baseline assumptions and near-/medium-/long-term risk analysis
- Realism (Figure A2) — staff commentary:
  - Baseline assumptions of most debt drivers do not point to a systematic bias in the forecast track record, broadly in line with those observed in peer surveillance countries.
  - Forecast track record comparisons and percentile assessments shown for 1 Yr., 3 Yr., and 5 Yr. ahead horizons.
- Near-term risk analysis (Figure A3) — staff commentary and metrics:
  - Mechanical results suggested a “high” probability of near-term sovereign stress.
  - Judgment was applied to override the mechanical signal and upgrade the near-term risk assessment to “moderate” given the large liquid asset buffer (70 percent of debt) not accounted for in the total reserves used in the logit regression.
  - Probability of missed crisis during 2019-20 (if stress is not predicted): 50%
  - Probability of false alarm during 2019-20 (if stress is predicted): 8%
- Medium-term risk analysis (Figure A4) — staff commentary and metrics:
  - The fanchart’s mechanical signal is “high” owing to its relatively wide width and elevated debt level at t+5. The GFN module’s mechanical signal is “moderate”.
  - Macro-fiscal shocks drive the rise in bank claims in the stress scenario; holder shock is limited by contained foreign private financing.
  - The combined medium-term mechanical signal is “high”.
  - Triggered commodity price stress-test indicates a t+5 debt/GDP ratio below the 75th percentile of the fanchart, although RAM indicates a “high” likelihood of occurring.
  - Staff’s final medium-term rating is “moderate” due to large liquid asset buffers that mitigate liquidity pressures and support market confidence and expectation that investment will increase potential growth (and revenue).
  - Probability of missed crisis during 2019-23 (if stress is not predicted): 60%
  - Probability of false alarm during 2019-23 (if stress is predicted): 4%
- Long-term risk analysis (Figure A5) — staff commentary:
  - Scaling-up/down of natural resources is triggered as extraction volumes over t+6 to t+15 are expected to increase more than one standard deviation compared to the last 10 years.
  - Outlook suggests a steady decumulation of debt and improvement in the primary balance.

### Backtesting results for the current framework (Annex I): coverage and predictive capacity
- Coverage findings (paragraphs 1–2 and Table AI.1):
  - Coverage remains an area for further reform.
  - Most advanced economies report at least on a general government basis, with only 9 percent reporting on a central government basis; about two-fifths of EMs still restrict coverage to the central government.
  - Table AI.1 (Debt Coverage Reported in MAC DSAs; percent):
    - EMs (78 countries) / AEs (35 countries):
      - Central government: 34.6 68.6
      - General government: 37.2 80.0
      - Nonfinancial public sector: 11.5 0.0
      - Consolidated public sector: 7.7 5.7
      - Other: 9.0 5.7
  - Risks from narrow coverage are confirmed by distribution of revisions to nominal debt levels by coverage level: revisions (percent deviation) were larger and more upward skewed where coverage was limited to the central government (Table AI.2).
  - Table AI.2 (Country's last DSA — Debt (pct. deviation)):
    - Central government: Mean 11.0 Median 2.1 75th percentile 4.5 Skew 2.8
    - General government: Mean 0.6 Median 1.4 75th percentile 3.0 Skew -2.3
    - Other definition: Mean 1.8 Median 0.0 75th percentile 5.1 Skew 0.7

- Predictive (discriminatory) capacity (paragraphs 3–4 and Box AI.1):
  - The predictive capacity of the threshold approach underlying the current framework has been weak.
  - High rate of missed crises associated with individual thresholds: e.g., around 70 percent for the debt and GFN thresholds in-sample (2007-13).
  - Considering an OR condition (call a crisis if any individual heatmap threshold is breached) reduces missed crises to an average level of 12 percent, but raises false alarms to 68 percent.
  - Predictive power has further worsened over time: missed crises rates for EMs declined slightly out of sample (2014-18), but false alarm rates for debt rose in 2014–18 relative to 2007–13 for most indicators.
  - The framework's limited discriminatory capacity allows countries with very different risk profiles to display very similar heatmaps (examples given: Spain vs Canada in 2014; Angola vs China; U.S. vs Lebanon).
  - Box AI.1 detailed findings (selected metrics from predictive performance tables):
    - Debt missed crises (in-sample / out-of-sample and by group) — example entries:
      - Public debt missed crises: 70.7 75.6 68.8 15.7 ...
      - Gross financing needs missed crises: 69.4 15.3 60.0 21.7 ...
    - Share of missed crises and false alarms using t-1 threshold breaches reported for multiple indicators and country groups; results show notable shares of missed crises and rising false alarms.

*Source: IMF staff (content unit: ppea2021003 - 5. Debt consolidation across subsectors:).*

### 5.      There is little evidence that false alarms resulted from policy action to avert crises in

### ppea2021003 - 5.      There is little evidence that false alarms resulted from policy action to avert crises in

### Policy endogeneity and false alarms
- Staff found little support that measured false alarm rates were biased upward by “policy endogeneity” (authorities’ timely policy actions to avert crisis after a DSA flagged risks).
- Teams rarely predicted explosive debt or GFN paths or made clear pronouncements on unsustainability; staff identified only two possible examples where policy reactions may have occurred: Slovenia and Serbia (Figure AI.2).

### Indicator-based signals versus team judgment
- Summary of performance ahead of stress episodes (2013–17):
  - There were 16 stress episodes during 2013–17.
  - Ahead of these 16 episodes:
    - The debt-to-GDP indicator flagged green in ten cases.
    - The GFN indicator flagged green in five cases and yellow in two cases (Figure AIII.3).
  - Team judgment outcomes:
    - In seven out of ten cases where debt-to-GDP flagged green, team judgment did no better than mechanical signals.
    - Team judgment predicted greater risks than the framework in three cases: Albania 2014, Bosnia and Herzegovina 2016, and Suriname 2016.
  - Mechanical framework correctly predicted six of the 16 stress episodes (both indicators flash red before a stress episode).
    - Teams’ bottom-line assessments flagged a major sustainability problem in only two of those six.
    - In the remaining four, teams gave more sanguine assessments than suggested by the heatmap.
    - In two instances debt was judged sustainable despite both debt and GFN indicators flagging red (Figure AI.3).
  - Aggregate comparison for the 16 cases (Figure AIII.3):
    - Team judgment was about in line with the mechanical signal in seven cases (twice correctly and in five instances incorrectly).
    - Team judgment did worse than the mechanical signal in six cases.
    - Team judgment did better than the mechanical signal in three cases.
  - Based on 2018 stress events, there is little evidence these patterns changed (Figure AI.4).

- Definition and sample notes:
  - The analysis excludes stress events that began before 2018 (e.g., Angola).
  - Argentina, Barbados, Pakistan constitute stress events because of their program requests. Turkey satisfies the inflation criterion and Lebanon is exhibiting high spreads.
  - The “false alarm” DSAs examined satisfy: (i) contain red flags, (ii) are subject to the high-scrutiny reporting requirements, and (iii) where there was no sovereign stress.

### Interpretation of false alarm cases (DSA chapeaus)
- In false alarm cases (high-scrutiny, non-stress DSAs), chapeau analysis shows:
  - Teams mainly acknowledged debt and GFN risks already highlighted in the heatmap.
  - Discussion of relevant mitigating factors was included in less than a quarter of cases.
  - Mitigating factors associated with indicators not included in the heatmap were more likely to be mentioned.
  - References to red flags for debt profile risks were generally uncommon (Table AI.3).

- Table AI.3 (as reported) — interpretation metrics (strings preserved exactly):
  - Panel 1: False alarms (% red flags)
    - To t a l61.258.624.057.738.348.029.6
    - AE80.090.00.069.244.448.90.0
    - EM45.949.025.050.036.847.630.9
  - Panel 2: Risk signal acknowledged (% false alarms)
    - To t a l78.080.033.323.438.941.166.7
    - AE75.088.9...8.925.021.7...
    - EM82.475.033.336.742.950.066.7
  - Panel 3: Mitigating factor for risk signal (% risks acknowledged)
    - To t a l1.623.30.013.614.323.37.1
    - AE2.829.2...25.00.060.0...
    - EM0.019.40.011.116.716.07.1
  - Panel 4: Other mitigating factors (% false alarms)
    - To t a l29.328.033.333.027.826.033.3
    - AE22.925.9...22.20.017.4...
    - EM38.229.233.342.935.730.033.3

### Factors teams considered beyond the standard framework
- Text analysis of DSA chapeaus (242 writeups between 2013 and 2018) shows teams cited a wide array of mitigating factors and risks:
  - Resilience to market risks: assets or buffers, holder profile of debt (stable/captive investor base as mitigating; exposure to sudden stops as risk).
  - Long-term factors: long-term fiscal costs associated with old age benefit and/or health programs.
  - Tail risks: bailout risks from SOEs; banking system health often cited as mitigating against contingent liabilities; natural disaster risks in many small states.
  - Authorities’ intentions: commitment to strong policies as mitigating; doubts about authorities’ capacity to deliver reforms as risk (Figure AI.5).

### Baseline realism and modeling of uncertainty — optimism bias
- Visual realism tools introduced in the 2013 framework appear to have reduced optimism in baseline projections for some debt drivers:
  - On average, projection errors for debt drivers covered by realism tools—primary balance, growth rate—were somewhat smaller than for drivers not covered—e.g., exchange rate and interest rate.
- Aggregate forecast performance and decomposition:
  - The average three-year change in debt/GDP outturn in post-2013 DSAs was about 5 percent of GDP higher than forecast, with an interquartile range of 1–7 percent of GDP.
  - Decomposition of errors indicates higher than expected exchange rate depreciations and interest rates were important contributors to adverse debt surprises (Figure AI.6). These drivers are not covered by the existing realism toolkit.
  - In the latest MAC DSA vintages, 78 percent of country teams projected debt stabilization by year t+5, despite only 34 percent of MACs achieving this since 2011.
- Cross-sample findings on forecast errors:
  - Since 2013, forecast errors were largest for Emerging Markets (EMs), especially commodity producers.
  - Negative debt forecast errors in Advanced Economies (AEs) (e.g., Ireland) were exceptions.
  - Small states exhibited high propensity for adverse debt surprises.
  - Forecast errors were generally smaller for program than for surveillance countries.

- Formal tests for projection bias (Table AI.4 and related regression output, preserved exactly as reported):
  - Regression (dependent variable: 3-year forecast error, current year to t+2; independent variable: constant). Selected reported coefficients and standard errors:
    - Debt/GDP 7.841*** (2.854)
    - Prgm/GFN/GDP 1.381 (1.633)
    - Real growth -0.592 (0.662)
    - Real IR 2.174** (0.917)
    - Prim. Def. 1.481* (0.843)
    - ER (contrib) 2.416*** (0.69)
  - Additional table fragment (preserved formatting and values):
    - AllPrgmHist.Con st PB
    - Debt/GDP7.841***-0.149-1.255-4.42
      (2.854)(6.075)(6.539)(6.317)
    - GFN/GDP1.3810.067-1.49-3.407
      (1.633)(3.048)(3.277)(3.183)
    - Real growth-0.5921.4060.5721.406
      (0.662)(2.006)(2.119)(2.006)
    - Real IR2.174**1.363*1.780**1.485*
      (0.917)(0.76)(0.793)(0.784)
    - Prim.  Def .1.481*-0.5693.6694.899
      (0.843)(1.218)(4.474)(4.819)
    - ER (con trib)2.416***2.003*......
      (0.69)(1.039)......
- Interpretation:
  - A bias is detected for debt/GDP projections on the full sample, but not when the test is performed on program cases only.
  - Decomposing forecast error into debt drivers shows some evidence of bias in real interest rate and real exchange rate forecasts for both the full sample and the program subsample.

### Output gap optimism
- Optimism bias exists for output gap estimates:
  - For high-scrutiny DSAs where an initial output gap was negative but closed by the end of the forecast horizon, many baselines showed above-potential growth in countries experiencing fiscal adjustments and lower inflation.
  - If growth had evolved according to the historical scenario (historic average), the output gap would not have closed by the end of the projection period (Figure AI.7).

### Annex II — Debt coverage specifics
- A. Liquid assets
  - New tools introduce customizations for liquid assets:
    - FX reserves can include readily available liquid assets (e.g., large foreign sovereign wealth funds (SWF)) in the near-term risk module’s ‘FX reserves’ variable.
    - The GFN module allows use of liquid assets in the stress scenario before extra debt is issued to the domestic banking sector.
  - Debt fanchart tool limitation:
    - Customization is not feasible for the debt fanchart due to data limitations, but staff judgment could substitute in specific cases.
    - Example criterion used by staff: for countries with SWF assets in excess of both 100 percent of gross debt and 100 percent of GDP, staff considers a low risk fanchart signal appropriate because tapping large assets could neutralize an explosive debt path.
    - For other countries with significant but sub-threshold assets, mechanical fanchart signal remains based on gross debt; medium-term risk assessment could be adjusted by country teams based on liquidity and availability of assets.
    - Operationalization details to be fleshed out in the Guidance Note.

- B. Central bank consolidation
  - Proposed consolidation approach:
    - Consolidate central bank in public debt only in cases of central banks with large negative capital positions and/or where the country team considers the central bank involved in significant direct monetary financing of the budget and/or quasi-fiscal activities.
    - Rationale: consolidation appropriate to fully capture public debt burden and debt risks.
    - When member country’s own debt reporting focuses on a consolidated concept, consolidation could benefit policy dialogue.

*Source: Excerpt from MAC DSA REVIEW—ANNEXES (pp. 6–12) as provided in the document content unit.*

### 4.      In case of central banks with healthy balance sheets, the framework will incorporate

### ppea2021003 - 4.      In case of central banks with healthy balance sheets, the framework will incorporate

### A. Mitigating role of central bank holdings (no consolidation)
- From a solvency perspective, substantial central bank holdings of government debt can be a mitigating factor when the net worth of the central bank (incorporating the expected value of its future seigniorage profits) is substantially positive—for example, where these holdings reflect a natural expansion of the monetary base.
- From a liquidity perspective, financing risks associated with central bank holdings of government debt are mitigated because central banks can typically be counted on to continue funding the government in periods of stress to the extent that this does not aggravate macro instability.
- Operational treatment:
  - Incorporate future seigniorage revenues into the fiscal projections and account for their impact on government financing risks.
  - The GFN module does not consider central bank purchases of government debt as being at risk of a sudden stop.

### B. Central bank liabilities: risk-based inclusion in public debt for non-consolidated cases
- Staff proposes a risk-based approach to include two types of central bank liabilities in public debt where the central bank is not consolidated with government accounts.

- Liquidity papers (issued solely for monetary policy):
  - Normally excluded from the DSA debt definition provided:
    - (i) no financing to the government can be provided through their issuance;
    - (ii) the government is not de facto responsible for paying debt service thereon; and
    - (iii) the securities do not represent a material fiscal risk (as indicated, for example, by a track record of central bank independence and monetary stability).
  - If one or more of these conditions is not met, liquidity papers would be included in public debt and GFNs for DSA purposes unless their outstanding stock can be deemed de minimis.

- Bilateral FX swap liabilities (CBFXS):
  - Will not be included in the DSA public debt definition so long as:
    - (i) they represent normal central bank monetary or liquidity operations (as opposed to sovereign-to-sovereign medium-term balance of payments support); and
    - (ii) the central bank is expected to be able to extinguish the swap position without actions detrimental to government debt levels (e.g. outright government foreign borrowing to pay off the swap).
  - If either condition is not met, the drawn amount of the FX swap should generally be included in the DSA, unless deemed de minimis.
  - When drawn, swaps for normal central bank liquidity operations are associated with accumulation of a short-term FX claim on banks by the central bank; matching short-term FX asset and liability signals monetary/liquidity nature.

- Special cases and operational clarifications:
  - If central bank issues Treasury securities in the primary market solely for monetary policy, these securities would normally be excluded from the DSA debt definition provided:
    - (i) funds collected as counterpart for the issuance will be kept in a blocked account in the books of the central bank that can only be debited for repayment of the said securities; and
    - (ii) the securities do not represent a material fiscal risk.
  - Further direction on when claims are de minimis will be included in the Guidance Note.

### C. Proposed realism tools (refinements and expansions)
- Full set of realism tools could include:
  - Color-coded table showing track record for forecast of all debt drivers and public debt at one-, three-, and five-year horizons versus a comparator group (scale from green to red; red indicates forecast optimism).
  - Decomposition of past and projected drivers of debt dynamics comparing past 5 years and projection period (next 5 years) to identify large shifts (tool already included in the LIC DSF).
  - Distribution of observed changes in debt-to-GDP ratios over a three-year horizon to compare a country’s projected change; notably, a 3-year debt reduction greater than the 75th percentile equals 3.5 ppts of GDP in the example.
  - Distribution of fiscal adjustments (three-year change in cyclically adjusted primary balance) to compare projected adjustments with country history and peers; a 3-year CAPB adjustment greater than the 75th percentile equals 3 percent of GDP in the example.
  - REER gap figure: initial REER misalignment extrapolated assuming no change in equilibrium REER; an initial gap that exceeds ±5 percent would trigger greater scrutiny.
  - Chart comparing real GDP growth projections with potential growth and output gap; signs of optimism where output gap without fiscal stimulus is positive at end of projection or significant increase in real growth relative to historical average.
  - For countries with output gap projections since 2010, a color-coded table showing track record for revisions of real-time, three- and five-year ahead output gap projections (scale from green to red; red indicates negative bias).
  - Consistency check between fiscal adjustment and growth assumptions (included in LIC DSF): compares impact of planned fiscal adjustment on growth under a range of fiscal multipliers and persistence parameters with baseline growth path.
  - Tool assessing new private borrowing and financing terms in maturity composition and spreads under baseline versus those implied by the Laubach rule; a shift toward long maturities or compression in spreads during debt accumulation flags realism problems.

- Notable technical parameter:
  - Laubach (2009) rule: bond spreads increase linearly by about 4 bps in response to a 1 ppt increase in the projected debt-to-GDP ratio.

### D. Definition and identification of stress events (mechanical criteria and validation)
- Purpose: refined and broad set of criteria to identify stress events for calibration; maintains prior criteria with broadened coverage.

- Mechanical criteria for stress events:
  i. Large IMF programs and exceptional financing from other IFIs and donors:
    - IMF Program size equal or greater than 100 percent of quota AND positive disbursement during the first year of the program. Years after the first are stress years if there are continuing positive disbursements.
    - Other IFI arrangements above 5 percent of GDP, and positive disbursements in years classified as stress.
    - Exceptional donor disbursement above 5 percent of external debt.
  ii. Default episodes:
    - External arrears equal or greater than 5 percent of public external debt AND increasing at least 10 percent in nominal terms (from the BoC-BoE Sovereign Default Database).
    - Domestic defaults: List from Erce and Mallucci (2018).
  iii. Restructuring episodes:
    - List from Das et al. (2012), complemented with Guscina et al. (2017).
  iv. Hyperinflation episodes:
    - Doubling of inflation rate compared to the year before AND inflation rate equal or greater than 25 percent OR inflation above 100 percent.
  v. Market-related indicators:
    - For AE:
      - Spreads (EU countries computed against corresponding German Bund maturity, others against corresponding US Treasury maturity) equal or greater than 1.5 standard deviations above 10-year mean AND above 150bp, OR spreads above 500bp.
    - For EM:
      - 100 percent increase or more in EMBIG spreads compared to the year before AND EMBIG equal or greater than 500bp OR, if EMBIG not available, 100 percent increase in real domestic interest rate compared to the year before AND real domestic interest rates equal or greater than 10 percent.
    - Loss of market access:
      - List from Medas et al. (2018) and Guscina et al. (2017).
  vi. Financial repression:
    - Central Bank claims on Central Government (from IFS) greater than 4 percent of GDP AND annual growth greater than 100 percent.
    - Commercial Banks’ claims on Central Government (from IFS) greater than 9.1 percent of GDP AND growth greater than 100 percent.
    - T-bill rate increase (IFS Database) above 4.5ppts y/y (if rate less than 11 percent) OR above 50 percent y/y (if rate equal or above 11 percent).
    - List selected individually from the Money and Capital Market Department of the IMF, based on TA reports and FSAPs.

- Validation process:
  - Mechanical-list underwent extensive validation using IMF staff reports, articles, working papers, newspapers, and additional databases (Paris club, World Bank, Central Banks, etc.).
  - For restructuring: verified whether debt treatment was preemptive or post-default, and whether part of larger operation or isolated; set stress episode dates accordingly; continuation of stress between default and restructuring considered only if country continued to accumulate external arrears.
  - IMF country teams cross-checked stress events.
  - If two stress episodes are separated by one year, considered the same episode and the intermediate year treated as stress (example: Jamaica 2012 included because Jamaica 2011 and 2013 are stress years).
  - An audit team from the IMF Research Department and the Institute for Capacity Development reviewed the list in July 2020, producing some minor corrections.

*International Monetary Fund — MAC DSA Review — Annexes (content excerpt).*

### 4.      This process allowed to identify 486 stress country-years, corresponding to 139

### 4.      This process allowed to identify 486 stress country-years, corresponding to 139 distinct “stress episodes”.

### Dataset and scope
- 486 stress country-years corresponding to 139 distinct “stress episodes”.
- The sample used to estimate the near-term tool covers the period 1990-2017 and includes most Market Access Countries (MACs).
- Staff identified more than 150 variables (and their transformations) that could be included in four categories of regressors.

### Classification and tables (Tables AIV.1 and AIV.2)
- Table AIV.1: lists stress country-years with color codes for:
  - stress country-years identified by mechanical criteria,
  - single country-years separating two stress episodes identified by mechanical criteria,
  - country-years inserted by applying judgement,
  - country-years added post-audit.
- Table AIV.2: provides details on the country-years that were added exercising judgement.
- Table AIV.1 contains a comprehensive listing of stress country-years across many countries and years (examples include Albania 2014, Argentina 1990–2014 entries, Belize 2006–2017 entries, Brazil 1990–1998 entries, and many others as enumerated in Table AIV.1).

### Main triggers of stress episodes
- Among the stress episodes, defaults (37 percent) and market stress (32 percent) were the most common “triggers”, defined as those occurring more often in the first years of stress episodes.

### Stress country-years included through exercise of judgement (Table AIV.2)
- Argentina: 2006-07; 2010-11 — Limited or no access to international capital markets, the central government heavily relied on the Central Bank balance sheet to finance its deficit (IMF Country Report No. 16/69).
- Armenia: 2014-16 — IMF program for 89.4 percent of quota (US$ 0.1 billion) + financing from Eurasian Fund for Stabilization and Development (US$ 0.3 billion) (IMF Policy Paper “Collaboration between Regional Financing Arrangements and the IMF”, 2017).
- Barbados: 2014-17 — Large accumulation of domestic arrears estimated at 4 percent of GDP in 2015 (IMF Press Release No. 15/342). In 2016 Moody's downgraded Barbados to Caa1.
- Equatorial Guinea: 2015-16 — Large accumulation of domestic arrears (information from IMF country team).
- Lebanon: 2006-07 — Financing needs satisfied through donor conference (US$7.6 billion) (see IMF WP/08/17).
- Malaysia: 1997 — Large capital outflows (52 percent decline in the Stock Exchange composite index), sharp cut in government spending (-17 percent), 35 percent exchange rate depreciation at end-1997 (see IMF Public Information Notice 99/88).
- Namibia: 2016-17 — Persistent under-subscriptions on government securities in auction across all maturities. Shortfall satisfied by the Government Pension Institution Fund through a private placement. (Information from IMF country team).
- St. Lucia: 2013 — Government unable to sell in auction about 2/3 of total (info from IMF country team).

### Near-Term Risk Tool — Model overview
- The near-term risk module consists of a multivariate logit model whose regressors characterize:
  - domestic institutions,
  - stress history,
  - cyclical variables,
  - debt burden,
  - global conditions.
- The annex explains regressor selection, estimation methodology, predictive capacity in- and out-of-sample, robustness checks, and customization options.

### Regressor selection approach
- Staff categorized candidate variables into four types:
  - (a) structural indicators,
  - (b) cyclical indicators (potential early warning indicators, EWI),
  - (c) debt and buffer indicators,
  - (d) global variables.
- Two statistical analyses guided selection:
  1. Identification of individual cyclical indicators with strong early-warning properties using a signal detection approach applied separately to advanced economies (AE) and emerging markets (EM), tested at five different (pointwise) projection horizons. This analysis highlighted common indicators for both groups, including debt dynamics and the current account balance.
  2. A Bayesian logit methodology to select EWIs together with structural and debt burden indicators, accounting for variable interactions and high dimensionality. This methodology produces a ranking of covariates by importance.

### Statistical and computation details
- The Bayesian methodology uses shrinkage priors to induce sparsity; staff adopted a horseshoe prior.
- Computation carried out by Markov Chain Monte Carlo methods (Gibbs Sampler).
- Thinning was used to reduce autocorrelation of MCMC samples.
- Variables were standardized to improve the efficiency of MCMC sampling.
- The Bayesian estimation was implemented in Matlab with the bayesreg package (Makalic and Schmidt, 2016).

### Performance assessment and validation
- Predictive performance of individual indicators assessed using the area under the receiver operating characteristic curve (AUC); significance of AUC derived non-parametrically via bootstrap resampling to calculate point-wise confidence intervals.
- Out-of-sample performance tested using:
  - temporal cutoffs (train on a time period, test on remaining period),
  - cross-validation on country-samples (train on certain group of countries, test on remaining countries).
- The final specification was compared to a benchmark fiscal crisis prediction model based on machine learning; machine learning improved out-of-sample performance but the improvement was limited and offset by reduced transparency and interpretability.

### Choice of final methodology and specification
- Final model: logistic regression (logit) selected for robustness, high statistical forecasting power, and ease of interpretation and reproducibility.
- Trade-off: more sophisticated techniques (Bayesian approaches, machine learning) often outperform logit but are harder to communicate and reproduce.
- Relative to probit, logit’s inverse linearizing transformation is directly interpretable as a log-odds.
- Final specification captures:
  - structural factors: institutional quality and stress history,
  - cyclical factors: current account balance/GDP, 3-year real effective exchange rate appreciation, credit/GDP gap,
  - debt burden/buffers: change in public debt/GDP, public debt/revenue, foreign currency public debt/GDP, FX reserves/GDP,
  - global factors: change in VIX.
- Forecast window: two-year forecast window (t+1, t+2) to accommodate timing uncertainty and allow time for corrective action.

### Consultations, robustness and outliers
- Internal and external consultations were conducted; suggestions were tested and endorsed when supported by statistical evidence.
- Robustness checks:
  - Removing potential outliers did not significantly affect coefficient estimates or predictive performance but reduced statistical significance of some variables.
  - Examination showed extreme observations correspond to countries with severe stress events and were retained.
- External consultation included experts such as C. Reinhart and K. Rogoff (Harvard), E. Duggar (Moody’s), L. Giorgianni (Tudor) and S. Pamies (EC).

*International Monetary Fund — MAC DSA Review, Annexes (excerpt).*

### 6.      The variables included in the

### ppea2021003 - 6.      The variables included in the

### Model variables and interpretation
- The final specification includes widely used variables from the literature: debt dynamics, cyclical changes, global indicators, and structural variables.
- Example of standardized-coefficient interpretation:
  - The standardized coefficient for FX public debt to GDP is about 1.4 times the magnitude of the coefficient for the change in public debt-to-GDP.
  - A 1 standard deviation higher FX public debt-to-GDP ratio (about 16.8 percent of GDP) would have roughly the same effect on the stress probability as a 1.4 standard deviation increase in change in public debt-to-GDP (approximately 7.5 percent of GDP).

### Coding and episode definition
- For the left-hand side (stress/non-stress) variable, cases where two stress episodes were separated only by only one year were considered a single episode.

### Technical audit: robustness checks and main findings
- An independent audit by IMF Research Department and Institute for Capacity Development summarized the following:
  - Robustness to sample selection:
    - Using 15- and 20-year rolling windows, in 85 percent of cases (161 out of 190) estimated coefficients remain within the 2-standard-error bands of the baseline coefficients.
    - When deviations occur beyond the bands, deviations are small and the sign is preserved.
    - Statistical significance: using a 15-year window, significance is lost in 15 percent of cases (18 out of 120) due to shorter sample size.
    - With 20-year windows, statistical significance is preserved at least at a 10 percent level in 99 percent of cases (69 out of 70). Exceptions: coefficients for public debt to revenue and current account balance to GDP lose significance in one subsample.
  - Leave-one-country-out analysis:
    - In all cases the sign of coefficients remained unchanged.
    - In more than 95 percent of cases estimated coefficients remained within the 2-standard-error bands of baseline coefficients.
    - In over 99 percent of cases coefficients remained statistically significant at least at a 10 percent level.
    - Coefficients that become insignificant when a particular country is removed: current account balance to GDP and change in the REER were the least robust.
  - Out-of-sample validation (2016-17 to predict 2017-19):
    - Out-sample AUC of 0.9737 when current account and REER change variables are included.
    - Out-sample AUC of 0.9698 when those two variables are excluded.
    - Conclusion: difference is minor; supports inclusion of the two variables in the final model.

### Fixed effects (FE): consideration and decision
- Audit recommended investigating fixed effects; MAC DSA team considered but rejected FE for conceptual and statistical reasons:
  - Conceptual/statistical concerns:
    - FE estimated over 1990-2015 would penalize countries that improved debt carrying capacity post-2015 (reforms, structural transformations, debt restructuring/relief).
    - Slow-moving structural variables are used instead to capture evolution in debt-carrying capacity.
    - Country fixed effects are politically sensitive and difficult to communicate, implying unchangeable inherent vulnerability.
  - Sample change issue:
    - Introducing FE more than halves the sample size (675 observations for 52 countries versus 1,675 for pooled logit), because the fixed effect can only be computed for countries that experienced stress in the estimation period; many advanced economies drop out.
    - This makes FE estimates unavailable for many advanced countries.
  - Robustness check with FE:
    - Significance and sign of coefficients remain broadly stable when FE are estimated, except:
      - Coefficient of “stress history” switches from positive to negative.
      - Coefficients of FX public debt/GDP and International reserves/GDP lose statistical significance.
    - Likely explanation: estimation with FE is performed only on countries that experienced stress and slow-moving variables are captured by FE.

### Standard errors and heteroscedasticity
- Audit recommended standard errors corrected for heteroscedasticity and within-country correlation.
- MAC DSA team adopted robust standard errors.
  - All coefficients remain statistically significant except for the current account variable; nevertheless current account is retained in the regression for reasons explained above.

### Capturing country heterogeneity and institutional quality
- Staff examined slow-moving structural variables to capture country heterogeneity (“debt carrying capacity”).
- Estimation results indicate:
  - The WGI-based variable (quality of institution index) and stress history have strong predictive power and best statistical properties among candidates.
  - The WGI-based variable significantly outperforms alternatives in statistical significance, coefficient magnitude, and robustness.
  - WGI are perception-based but are a broad summary measure based on several hundred individual variables from 31 separate data sources.
  - Two of the six WGIs are used in the quality of institution index: Government Effectiveness and Regulatory Quality.
- Operational note:
  - If teams judge the WGI-based institutional quality variable to be a poor proxy for a country, teams may adjust their judgement in the final risk assessment.

### Capturing regional spillovers
- VIX sometimes poorly proxies regional contagion.
- Staff tested regional spillover proxies (share of AE/EM in stress, share by region, share with strong trade linkages) with no statistical significance.
- The share of currency union (CU) members in stress was highly significant.
  - Examples consistent with this finding: euro area sovereign debt crises and CEMAC stress after the 2014-15 oil price drop.
- Default model setting mutes the CU variable, but it can be switched on if country teams consider spillover risks within a CU material.

### In-sample predictive performance
- Overall in-sample performance is very good and a significant improvement over the existing heatmap framework:
  - AUC (Area Under the Curve) = 0.88.
  - Minimum total misspecification error (TME) = 37 percent.
  - The 37 percent TME corresponds to a 9 percent probability of stress (vertical blue line).
  - The minimum TME of 37 percent reflects:
    - Missed crisis rate = 10 percent.
    - False alarm rate = 27 percent.
- Comparison with existing heatmap OR rule:
  - Existing framework missed crisis rate similar (9 percent for EMs and 14 percent for AEs) but much higher false alarm rates (63 and 72 for EMs and AEs, respectively), implying TMEs of 72 percent and 86 percent, respectively.
- Country-level performance:
  - In-sample performance is very good in individual countries.
  - Predictive performance higher in countries exposed to regional spillovers when the logit includes the share of CU MACs in stress.
  - Performance weaker for stress driven by political instability (e.g., MCD countries 2010-12 Arab Spring; Ukraine 2014), underscoring the role of judgement in final assessments.
- Consistency checks:
  - When using sufficiently long training periods, performance of logit and alternative models was broadly comparable.
  - Estimated risk rankings (based on 2018 data) from the two models revealed a correlation of 0.83.

### Pseudo-out-of-sample tests and comparison with machine-learning model
- Pseudo-out-of-sample tests used two training samples:
  - i) 2000 to 2015 (test sample 1990–99).
  - ii) 1990 to 2012 (test sample 2013–15).
  - Results: performance in missed crises, false alarms, and minimum TME is robust under both cutoffs.
- Comparison with Vulnerability Exercise Fiscal Module (VEFM), a machine-learning “Random Forest” model:
  - VEFM delivers even better out-of-sample predictive performance than the logit, particularly when trained on shorter sample periods.
  - Examples:
    - Trained over 1990-2012: AUC for logit = 0.88 versus AUC for VEFM = 0.90 (VEFM trained over 1980–2012).
    - Trained over 1990–2005: AUC for logit = 0.73 versus AUC for VEFM = 0.82 (VEFM estimated over 1980–2005).
  - Trade-offs:
    - Logit advantages: greater transparency and easier economic interpretability.
    - VEFM disadvantages for policy discussion: Random Forest uses above 100 variables including interaction effects that are less straightforward to explain.
  - Implementation decision:
    - The logit will be the main workhorse for near-term risk analysis due to interpretability.
    - VEFM will be made available to teams to provide complementary insights for judgement-based final assessments.

*Source: Fund staff calculations and MAC DSA team audit summarized in the provided text.*

### 12.      The data was checked carefully for outliers. Large regressor values (for example the very

### ppea2021003 - 12.      The data was checked carefully for outliers. Large regressor values (for example the very

### Data validation and outliers
- Large regressor values (for example the very large surplus in the CA of Gulf countries, or the very large credit-to -GDP gap in countries that experienced a financial crisis) were all cross-checked and validated in the data.
- Staff ran the specification with the top and bottom 1 percentile removed (263 observations). The results of this analysis confirmed the magnitude and signs of estimated coefficients.

### Performance of the proposed mechanical signals (Logit Stress Probability — LSP)
- The logit stress probability (LSP) predicted by the model is divided into three risk zones (high, moderate, low) based on the probability of missed crises and false alarms.
- Low- and high-risk cutoffs are calibrated to keep the rate of missed crises and false alarms at 10 percent, respectively.
- Corresponding stress probability cutoffs:
  - 9 percent at the threshold between the low and moderate risk signal
  - 20.5 percent at the threshold between the moderate and high risk signal
- Average stress probability by fitted signal (historical sample):
  - “High”: 40 percent
  - “Moderate”: 16 percent
  - “Low”: 2 percent
- Plausibility/backtest check:
  - Risk signals ahead of selected well-known stress episodes are reported in Table AV.4. With one exception—Jordan 2012—the model flagged risks in advance as moderate- or high-risk signals.
  - Note: A risk signal for some cases reflects inclusion of the currency union variable; exclusion can change a “high” to “moderate” (see Table AV.4 footnote).

### Customization of the logit tool in special cases (guidance points)
- Commodity exporters:
  - Where GDP is more volatile, large increases in the credit-to -GDP gap could be due to GDP shrinking rather than to credit to the private sector increasing, introducing noise in the signal. In those cases, it could be warranted to use the credit to non-oil to GDP ratio to compute the gap.
- Countries with large foreign assets in a SWF:
  - A customized approach would allow inclusion of the share of those assets that are liquid and readily available in case of stress in the model’s ‘FX reserves’ variable.
  - Guidance will discuss how to handle situations where a clean accounting of liquid assets is not available.
- Countries less vulnerable to changes in global risk appetite (e.g., safe havens or countries with very low near-term external financing needs):
  - Guidance will be provided to deal with situations where VIX movements alone are seen to drive a change in the mechanical risk signal for such countries.
  - Note: The impact will not be nil, since changes in the VIX can also signal expected real economic activity effects via trade and foreign direct investment channels.

### Logit regressors — summary statistics (Table AV.5 highlights)
- Table AV.5: Logit Regressors’ Summary Statistics (this excludes variable values observed during stress episodes)
  - 1557 observations
- Selected summary percentiles/values for the first regressor column (Institutional Quality):
  - min: -1.60
  - p1: -1.38
  - p10: -0.45
  - p25: 0.01
  - p50: 0.62
  - p75: 1.21
  - p90: 1.75
  - p99: 2.03
  - max: 2.25
  - sd: 0.84
  - mean: 0.64

### Annex VI — Technical notes on the debt fanchart (overview)
- Purpose:
  - Describe the two-step procedure to generate a new debt fanchart, derive three metrics from it, define an overall index, and present backtesting.
- High-level aim:
  - Replace current fancharts and standardized macro-fiscal stress tests with a procedure that applies a realism check and imposes a “realism-adjustment” when risks to debt projections appear heavily skewed.
  - Construct fancharts based on historical shocks of debt drivers, yielding generally asymmetric fancharts.

### A. Fanchart methodology (two-step procedure)
- Step 1 — Historical fanchart and realism diagnosis:
  - Generate a “historical fanchart” by drawing stochastic realizations of debt drivers from their joint empirical distribution to capture correlations across debt drivers.
  - Inter-temporal dependence captured via a “block-bootstrap” approach using consecutive two-year blocks drawn from the 1990-2018 sample.
  - Process to populate the fan chart is repeated 10,000 times.
  - If the team’s baseline debt path falls below the 20th percentile of the historical debt fanchart, the baseline is assessed as unlikely to represent an adequate balance of risks and further scrutiny is required.
- Step 2 — Final fanchart construction:
  - i. If the team’s baseline does not fall below the 20th percentile of the historical fanchart:
    - Produce a “standard” fanchart. The baseline’s forward-looking information determines direction; width and skew are determined by the historical fanchart.
  - ii. If the team’s baseline continues to fall below the 20th percentile after scrutiny:
    - Compare the deviation between the team’s baseline and the level implied by historical trends with the historical cross-country distribution of this metric for relevant peers to determine the country’s percentile.
    - Construct a “realism adjusted” fanchart by adding skewed shocks to the underlying debt drivers, moving the distribution right until the team’s baseline falls at the same percentile in the fanchart distribution as it does in the cross-country distribution.
    - Peer groups: Advanced Economies, EM commodity exporters, EM non-commodity exporters.
  - Minimum setting for the adjustment percentile is the 10th percentile to avoid the team’s baseline falling outside the fanchart.
- Special-case exit clause:
  - An escape clause from the asymmetric fanchart can be applied in rare situations (e.g., restructuring cases) with clear guidance and explanation required.
- Covid-19 recovery modification (2021-22):
  - A modified historical fanchart uses the team’s baseline debt projections for the first two years as central tendency, then the standard historical data-generating process afterwards, to limit incorrect triggers of the optimism correction mechanism during atypical recovery declines in debt-to-GDP.
  - Staff test: Using a modified historical scenario that uses baseline debt projections for 2021-22, the realism correction would be applied in only 9 cases. Using the standard historical scenario it would be applied in 22 cases.
- Key simplifying assumptions retained:
  - (i) No feedback between the debt drivers and the level of debt.
  - (ii) Interest rates on domestic- and foreign-currency debt face the same shock distribution (calibrated on past behavior of average effective interest rates).
  - (iii) Foreign currency debt shares are non-stochastic (fixed at baseline projections).
  - Implication: The fanchart understates true uncertainty (fancharts will be too narrow) and upper percentiles may miss explosive debt paths driven by snowball effects from rising spreads.

### B. Fanchart metrics and predictive performance
- Candidate metrics considered (analyzed using 2010–15 fans) fall into four categories:
  - (i) Probability of debt stabilization over the projection horizon.
  - (ii) Probability of long-term debt stabilization.
  - (iii) Uncertainty around the debt projection.
  - (iv) Projected level of debt.
- Three metrics selected for intuitive appeal and encouraging performance (ROC analysis over the backtesting period):
  - Probability that the debt does not stabilize in the medium-term.
  - Fanchart width.
  - Debt level at t+5 interacted with an index of institutional quality (to capture debt-carrying capacity).

*Source: Fund staff estimates.*

### 9.      While each of these metrics can be used to predict sovereign stress individually, their

### 9.      While each of these metrics can be used to predict sovereign stress individually, their

### Composite Debt Fanchart Index (DFI): construction and predictive performance
- The three fanchart metrics are aggregated into a composite Debt Fanchart Index (or “DFI”) that weights each metric by its predictive power and is used to classify countries into risk groups.
- The aggregate index distribution differs markedly for crisis and non-crisis episodes (2010–15), indicating strong discriminatory capacity.
- Quantitative performance metrics:
  - AUC of the aggregate index: 0.82
  - Minimum TME: 43 percent, corresponding to an index value of 0.32 (vertical blue line)
- Risk-zone calibration:
  - Low- and high-risk thresholds are calibrated so the high-risk threshold is associated with a 10 percent missed crisis rate and the low-risk threshold with a 10 percent false alarm rate.

### Posterior stress probabilities and interpretation of risk zones
- The DFI level is not itself a direct estimate of the probability of a “stress” event; posterior probabilities of stress at each DFI level are estimated empirically from sample shares of countries within DFI ranges that experienced stress.
- Empirical estimation approach:
  - DFI divided into 20 “bins” (intervals); posterior stress probabilities computed for each bin based on observed stress incidence (2010–15).
  - Limited number of “stress” observations implies imprecise estimates, particularly at higher DFI values where fewer observations exist.
- Key empirical posterior probability findings:
  - Posterior probability of stress conditional on a GFI “high risk” signal: at least 40 percent.
  - Posterior probability of stress conditional on a GFI “low risk” signal: at most 10 percent.
  - Average posterior probability of stress by proposed risk zone:
    - “high risk” signal: 44 percent
    - “moderate risk” signal: 23 percent
    - “low risk” signal: 3 percent

### Customization of the Fanchart Tool in Special Cases — guidance and adjustments
- Guidance will specify adjustments to the standard methodology in special cases:
  - Public sector large financial assets (e.g., stabilization or sovereign wealth funds, SWFs):
    - Government solvency typically stronger than suggested by standard gross debt fancharts because assets can neutralize explosive debt paths.
    - Staff does not view automatic incorporation of these effects in fanchart construction as feasible due to data limitations; guidance could ensure appropriate accounting in mechanical risk signals.
    - Staff identified seven MACs (Brunei, Kuwait, Norway, Qatar, Saudi Arabia, Singapore and the UAE) where SWF assets exceed both 100 percent of gross debt and 100 percent of GDP; assigning a low risk fanchart signal to these cases would seem reasonable.
    - For other countries with significant but smaller assets, team judgment should assess liquidity and availability of the assets; mechanical fanchart could remain based on gross debt, with overall medium-term risk assessment adjusted by country teams as appropriate.
  - Structural breaks and “realism-adjustment” exceptions:
    - Cases where the fanchart’s ‘realism-adjustment’ would not be appropriate and metrics need to be based on the “standard” (symmetric) fan.
    - Examples (rare): a past major crisis not expected to repeat; major natural resource discovery or depletion relative to the past; regime changes like accession to a currency union.
    - Guidance will outline an escape clause to the “asymmetric fanchart” for these cases.
  - Countries close to reaching a debt restructuring agreement:
    - Do not apply the realism correction; build the fanchart around the team’s baseline by default.
    - Post-restructuring effective nominal interest rate volatility likely lower; volatility could be scaled down by a factor corresponding to the ratio of new and past debt issuances.

### Past restructurings and whether fanchart adjustment is warranted
- Staff conclusion: no fanchart adjustment warranted for countries with past debt restructuring experiences.
- Empirical check: dropping restructuring years from historical data typically did not materially narrow 2020 fancharts; in many cases width did not decline or slightly increased, except in the case of Greece where restructuring years were associated with a large recession and width fell.
- Rationale: restructuring years often associated with positive primary balance shocks (fiscal adjustment) that can offset negative growth shocks and dampen debt dynamics; removing these years can increase measured volatility.
- Table AVI.1 sample results (impact of dropping restructuring years on fanchart width):
  - Antigua and Barbuda: Restructuring years 2010-11 — Standard width 67.2; Width after dropping 68.2; Difference 1.0
  - Barbados: Restructuring years 2018-19 — Standard width 48.8; Width after dropping 47.7; Difference -1.2
  - Belize: Restructuring years 2012-13 — Standard width 31.4; Width after dropping 32.4; Difference 1.0
  - Greece: Restructuring years 2011-12 — Standard width 87.2; Width after dropping 73.1; Difference -14.0
  - Jamaica: Restructuring years 2013-14 — Standard width 31.5; Width after dropping 32.8; Difference 1.3
  - St. Kitts and Nevis: Restructuring years 2011-12 — Standard width 47.0; Width after dropping 48.8; Difference 1.8
  - Ukraine: Restructuring years 2015-16 — Standard width 61.4; Width after dropping 64.7; Difference 3.4
  - Average: Standard width 53.5; Width after dropping 52.5; Difference -1.0
  - Median: Standard width 48.8; Width after dropping 48.8; Difference -0.1

### Annex VII — GFN Module: data requirements
- The GFN module introduces several new data requirements; reliance on standardized cross-country databases (e.g., Fiscal Monitor, International Financial Statistics, a centralized debt holder profile database) is intended to limit new staff effort.
- Key new or refined inputs:
  - Estimates of amortization by debtholder are a key ingredient and often require refinement (Box AVII.1).
  - Holder-profile of stock of debt (Arslanalp-Tsuda methodology) combined with maturity profile information in country DSA files used to produce working estimates of holder profile of debt amortizations for almost all MACs.
  - Country teams encouraged to enter more accurate holder-specific amortization information before framework goes live in early 2021, with cooperation from country authorities.
- Practical pointers for gathering holder-profile amortization data (Box AVII.1):
  - External debt holders:
    - Total external amortization is compiled by country authorities and available in IMF-World Bank databases.
    - Private external amortization can be calculated as the difference between total external amortization and amortization on non-marketable obligations to official creditors.
    - COFER database could be enhanced to identify maturity profile of marketable debt held by foreign central banks.
  - Domestic holders:
    - Amortizations to the domestic central bank should be readily available.
    - Amortization due to domestic commercial banks could be collated from banking surveys containing maturity profile of banks’ government securities holdings or from a country’s securities registry.
    - Amortization due to domestic nonbank sector obtained as the residual.
  - Note: holder profile data cannot be pinned down exactly because marketable debt can change hands; even approximate holder profile data is critical for sovereign risk analysis.

### GFN Module: financing assumptions and holder shock implementation
- Financing assumptions expanded beyond differentiation between domestic and external financing:
  - Users asked to allocate domestic issuance among central bank, commercial banks, and other resident sectors, and divide external debt issuance among official and private creditors.
  - Implied data burden limited because: (i) holder profiles for certain instruments obvious; (ii) BOP and monetary sector projections inform assumptions; (iii) teams can assume holdings remain equal to share of existing debt where allocations are unclear.
  - Module introduces a minor new requirement to indicate whether an instrument was marketable or not.
- Additional data inputs of relevance in key cases:
  - Government asset buffers (e.g., SWFs) — DSA templates could be populated automatically from centralized databases like the Fiscal Monitor.
  - Non-bank financial institutions — may require national balance sheet/flow of funds accounts data for countries where sovereign relies on non-bank financing.
- Stress scenario composition (macro-fiscal and financing shocks):
  - Macro-fiscal shocks broadly similar to existing MAC DSA stress tests and include:
    - A one standard deviation (computed over the last 10 years) reduction in the real GDP growth rate for two years.
    - For countries outside currency unions and with their own legal tender: a one-year exchange rate shock equal to the largest annual depreciation observed in the last 10 years.
    - For currency union members and dollarized economies: a deflator shock equal to half of the largest one-year change in inflation rates.
  - Knock-on effects on inflation:
    - Exchange rate shock pass-through: for a 1 percent depreciation, inflation rises by 25 basis points for EMs and 3 basis points for AEs.
    - Growth shock pass-through: growth shock reduces inflation by 25 basis points for each 1 percentage point reduction in real GDP growth.
  - Primary balance effects (two years):
    - Revenue/GDP ratio fixed at baseline (elasticity of 1).
    - Expenditures fixed at baseline nominal levels (elasticity of zero).
  - Calibrations consistent with current MAC DSA; caps placed on inflation and fiscal balance to avoid counterintuitive results.
- Financing shocks:
  - Shortening of maturities increases GFNs via higher amortization payments.
  - Scenario financing assumptions allocate issuance toward shorter-term instruments, following average maturities of bond issuances in recent crisis events, with about half of issuance concentrated in T-bills (Figure AVII.1).
- Holder shock (external debtholder rollover shock) implementation:
  - After allocating debt issuance across 5 creditor groups (central bank, domestic commercial banks, other domestic creditors, foreign official, foreign private), impose holder shock.
  - Shock simulates a loss in foreign appetite: foreign private rollovers drop to a 67 percent (i.e., rollovers fall to 67 percent) and investors unwilling to finance any new borrowing requirement over a two-year period.
  - First line of defense: government asset buffers; if depleted, domestic banking sector absorbs residual financing needs (Box AVII.2).
- Robustness of risk signals:
  - Risk signals are not sensitive to exact definitions of the stress scenario or holder shock because thresholds for the mechanical signal are calibrated based on probabilities of missed crises and false alarms for a given test definition; more severe shock definitions would yield higher thresholds.

### Behavior of the domestic banking sector in sovereign stress episodes (Box AVII.2)
- Empirical patterns:
  - Domestic banking systems tend to increase exposure to government in sovereign stress and serve as residual financing source.
  - Ability of banks to increase government debt holdings constrained by existing exposures.
  - Empirically, bank claims on government seldom rise above 20 percent of banking system assets and tend to rise less in stress events when starting exposures are high.
  - Empirical findings consistent with Arslanalp and Tsuda (2014), who used 15 percent of assets as a risk threshold.
- Supporting empirical charts and regressions reported (sources: International Financial Statistics and Fund staff calculations), including:
  - Relationship between change in claims on government during crisis and starting level of bank claims on government (regression: y = -0.2121x + 3.2227; R² = 0.1558).

*Source: Fund staff calculations and text excerpts from the MAC DSA REVIEW—ANNEXES (pp. 50–57).*

### 8.      Staff simulated the GFN module using past DSA templates. This involved running the

### 8. Staff simulated the GFN module using past DSA templates

### Simulation dataset and setup
- Staff ran the GFN module using macroeconomic, fiscal, and financing assumptions obtainable from MAC DSA templates prepared over 2014-15 and submitted to the MAC DSA archive.
- These were augmented with debt holder profile data from the last observed year in the corresponding MAC DSA template and information on banking system assets obtained from the IFS.
- Altogether, this process provided 125 observations (corresponding to about 60-70 country DSAs per year).
- Predictive-performance analysis focused on GFN signals between 2014 and 2015 because stress outcomes for more recent periods cannot be observed for the full medium term (5-year) prediction period.

### Risk indicators examined and individual performance
- Staff examined several potential risk metrics and concluded an index composed of three indicators showed the best performance:
  - GFN levels:
    - GFN levels have significant explanatory power in predicting crises (although not in the non-linear fashion implicitly assumed by threshold-based signals).
    - ROC curve analysis on the average GFN projections in past DSAs submitted to the MAC DSA database suggests an in-sample AUC of 0.81.
  - The volume of financing needed from domestic banks:
    - Staff tested the change in the ratio of bank claims on the government to banking system assets under both the baseline and stress scenarios.
    - The baseline change had no explanatory power.
    - The change in bank claims on the government in the stress scenario showed an AUC of 0.79.
  - The level of initial bank claims on the government (in percent of assets):
    - ROC curve analysis confirmed that higher initial bank exposures reduce capacity to increase holdings and have predictive power.

### GFN Financeability Index (GFI) — aggregation and performance
- Aggregation:
  - Staff combined the three indicators into an aggregate GFN Financeability Index (GFI), weighted by their explanatory power (i.e., their AUC).
- Performance metrics:
  - The overall index has an AUC of 0.83, an improvement over each indicator in isolation.
  - Back-testing over archived DSAs 2014-15:
    - Composite index AUC: 0.83.
    - Composite index TME: 42 percent.
    - By comparison, existing MAC DSA GFN thresholds are associated with average missed crisis and false alarm rates of 68 and 15 percent, respectively.
  - The implied debt path under the stress scenario lies above the median but below the 95th percentile, five years out, in 89 percent of cases—suggesting the scenario is severe but not an extreme tail risk.
- Risk-zone classification:
  - The GFI is conducive to a three-risk zone classification: high, moderate, low.

### Posterior stress probabilities and MTI
- Posterior probabilities from GFI:
  - Posterior probability of stress conditional on a GFI “high risk” signal: at least 24 percent.
  - Average posterior probability of stress by GFI risk zone:
    - High risk: 42.1 percent
    - Moderate risk: 4.1 percent
    - Low risk: 3.5 percent
  - Estimates derived from shares of countries within GFI ranges that went on to experience stress; limited number of stress observations implies imprecision, especially at higher GFI values.
- Posterior probabilities for the Medium-Term Risk Index (MTI):
  - MTI derived by averaging the fanchart and GFN indices.
  - Posterior probability of stress conditional on an MTI “high risk” signal: at least 20 percent.
  - Average posterior probability of stress by MTI risk zone:
    - High risk: 43 percent
    - Moderate risk: 9 percent
    - Low risk: 4 percent

### Customization of the GFN tool in special cases
- Standardized customizations and guidance will allow teams to account for factors affecting government liquidity availability:
  - (i) Use of government assets to offset funding pressures:
    - Teams can customize the size of available liquid asset buffers that can be used to meet financing needs generated by the holder shock before allocating new claims to banks.
  - (ii) Role of the domestic non-bank financial sector as a residual creditor:
    - Where the domestic non-bank sector is larger than the domestic banking system, funding needs in the stress scenario can be absorbed by the broader financial sector, reducing demand on banks.
    - Relevant for reserve-currency issuers with large domestic nonbank sectors.
  - (iii) Non-bank financial intermediaries as a source of government funding risk:
    - Teams should identify portions of the sector’s financing subject to a sudden stop and treat them like private foreign creditors under the holder shock where appropriate.
  - (iv) Timing of the onset of stress:
    - Start of the rollover shock, macro-fiscal, and maturity-shortening shocks are adjustable to align stress onset with specific events (e.g., elections) or accommodate lack of significant external private debt maturing early in the projection period.
  - (v) Use of more granular information:
    - When teams have detailed information (e.g., FSAP analysis, banking regulations, intra-public arrangements, capacity for liability management, capital flow measures), guidance specifies how to integrate this information with the standard approach.
- Notes on applicability:
  - Liquid asset buffers are particularly important in major EM commodity producers with large sovereign wealth funds (examples listed in source) and in advanced economies with sizable financial assets (examples listed in source).
  - Where the non-bank financial sector is large, its inclusion in initial government exposures/asset ratios for the GFI would likely reduce that ratio, capturing the benefit of a deep financial system.

### Resource implications and data requirements for the new MAC DSA
- Transition and maintenance:
  - Transitioning to the new framework requires time and resources but should not be costlier to maintain once operational.
- Data requirements and centralization:
  - Most data requirements carry over from the existing MAC DSA framework.
  - Fresh data requirements arise in four areas: debt holders, 10-year projections, inputs for stress tests and long-term modules, and debt disclosures.
  - The 10-year projection horizon is a new requirement but can be satisfied through careful extrapolation after the normal 5-year horizon rather than a full financial program.
  - The new template will be pre-populated with default parameter settings and centrally warehoused data to facilitate implementation.
- Transition support:
  - SPR will provide intensive support to area departments through an implementation team to assist in transferring databases and customizing templates, drawing on LIC DSF rollout experience.
- Special transitional arrangements:
  - SPR will facilitate transitional arrangements for PRGT-graduating members and other frontier countries; early identification of potential graduates/new users should provide time for training and technical support.
- Additional data note:
  - Some debt disclosures may be a new data requirement and may require support from country authorities, though they provide critical information on debt risks.

*Source: Fund staff calculations.*

---


_Source: https://www.imf.org/-/media/files/publications/pp/2021/english/ppea2021003.pdf_