## wpiea2026010-source-pdf

## Source details

**Canonical URL:** [wpiea2026010-source-pdf](https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026010-source-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2026/english/wpiea2026010-source-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2026/english/wpiea2026010-source-pdf.pdf.json)

---

### Introduction — context and motivation
- Net international issuances by EMDE sovereigns surged pre-pandemic, producing a threefold increase in the nominal value of outstanding international securities (Figure 1).
- The average cost of debt has remained elevated and increased further since 2019.
- The share of issuances with spreads above 500 bps:
  - reached a record of 20 percent in 2020;
  - has remained above 25 percent thereafter.
- The weighted EMBIG spread, calculated using net issuances, has declined due to high spread issuances focusing on refinancing operations, reducing net issuance volumes for countries with high spreads.
- Implication: constraints to market access may have become more binding recently.

### Theoretical framework — stylized moral-hazard model of credit rationing
- Model setup and assumptions:
  - Two-period model (t = 0, 1) with two states θ ∈ {G, B}.
  - Borrower risk-averse with utility u(c1); lender risk-neutral.
  - Project return Rθ with RG > RB; baseline assumption RB = 0.
  - Borrower finances investment K̃ with endowment A and debt D at cost (1 + r).
  - Effort ei ∈ {eH, eL} with cost B > 0 for eH and π(eH) > π(eL).
  - Signal si ∈ {H, L} with probabilities qH, qL such that qH > qL.
- Key constraints and equilibrium conditions:
  - Net present value condition and lender participation constraint (PC).
  - Incentive compatibility constraint (ICC) ensures high effort is optimal.
  - First-order conditions show rH tightens ICC while rL relaxes it.
- Propositions and comparative statics:
  - Proposition 1 (Non-binding ICC): If 훥훥휋휋 (푅푅_H − (퐾̃ − 퐴) (1 + 푟푟**)) > 퐵퐵 휋휋(푒푒_H), borrowers set a unique interest rate r** = 1/π(eH).
    - An increase in borrower net worth A relaxes the ICC.
    - Higher returns, lower B, or larger Δπ relax the ICC for a given spread.
  - Proposition 2 (Binding ICC): If collateral constraint binds, borrowers offer signal-dependent interest rates with r_L** > r_H**, and reverse-macro contingent instruments (higher rates in bad states) can loosen ICC and restore market access.
- Model interpretations linking to observable variables:
  - Gross financing needs (GFN) and official debt can be interpreted as reducing effective net worth A.
  - Borrowers may default in the bad state independent of interest rate (model assumes u(0) = 0).

### Empirical approach — data, machine learning, and identification limits
- Sample and scope:
  - Quarterly data for all EMDEs with available EMBIG spreads from 2000Q1 until 2023Q2.
  - Dataset spans 94 quarters and 46 countries.
- Dependent variable and forecasting horizon:
  - Binary variable at quarterly frequency = 1 if country issued in either quarter t+3 or t+4, and 0 otherwise.
  - Baseline forecasting horizon: three to four quarters in advance.
  - Class balance: dependent variable equal to one in 58 percent of the observations.
- Modeling strategies:
  - Primary: random forest (Breiman, 2001) with Shapley value interpretability and Partial Dependence Plots (PDPs).
  - Comparison: Logit models with three regularization techniques (lasso, ridge, no regularization).
- Training/testing and validation:
  - Training set: 2001Q1-2021Q4.
  - Testing set: 2022Q1-2023Q2 (implying a 90/10 percent split).
  - Cross-validation: expanding-window time-split with one-period gap to avoid look-ahead; ROC-AUC averaged across four validation sets used for tuning.
  - Missing values imputed using K-nearest neighbors; many variables transformed to "% of GDP", yoy growth, five-quarter averages, variances.
- Identification caveat:
  - Study does not fully identify supply versus demand shocks; results reflect predictive associations rather than causal relationships.
  - Advantage: random forest conditions on demand and supply proxies and their non-linear interactions to estimate counterfactual predictive associations.

### Key empirical findings on drivers of issuance and feature importance
- Overall model performance and interpretability:
  - Random forest ROC-AUC = 0.84.
  - Best-performing logit ROC-AUC = 0.77.
  - Feature importance assessed via mean absolute Shapley values; PDPs used to show directional effects.
- Top features (consistent across models):
  - General government international outstanding obligations (total and short-term), year-over-year change in GG international obligations (net issuances), total short-term external debt.
  - International reserves.
  - Fiscal balance forecasts (notably 3 years ahead) and projected public debt ratios.
  - EMBIG spreads.
  - Nominal GDP.
  - Exchange rate regime and governance indicators (rule of law, control of corruption).
  - IMF credit usage and outstanding official creditor debt.
- Balance-sheet related findings:
  - Larger stocks of outstanding obligations are associated with a higher predicted probability of issuance, but effects plateau at very high debt levels.
  - Changes in outstanding international securities: issuance most likely after modest increases; very large increases reduce issuance likelihood.
  - Higher reserves tend to be associated with less frequent issuance.
- Forward-looking and fiscal features:
  - Projected fiscal balances that are positive three years ahead significantly increase issuance likelihood relative to projected deficits.
  - Projected high debt shares of GDP lower issuance probability.
- External sector and institutional features:
  - More positive trade balances lower issuance likelihood.
  - Fixed exchange rate regimes are associated with lower issuance probability; NEER shows non-linearity with appreciations linked to higher issuance likelihood.
  - Governance effects are nonlinear and more relevant in high-spread situations.
- Stability over time:
  - Feature importance broadly stable across sub-periods (2002–2007, 2008–2014, 2015–2021).
  - Exchange rate regime and IMF financing more influential in earlier periods; forward-looking fiscal indicators and EMBIG spreads gained prominence recently.

### EMBIG spreads, spread subgroups, and nonlinearities
- EMBIG spread relationship:
  - PDP shows inverted U-shaped relationship: probability rises with EMBIG at low levels, peaks around 200 basis points, then declines; sharpest drop around 600 bps.
  - Interpretation: low-spread issuance driven by demand; high-spread issuance constrained by supply.
- Spread group cut-offs:
  - Low spreads: below the 40th percentile (approx. 200 basis points).
  - Medium spreads: between the 40th and 80th percentiles (approx. 200–500 basis points).
  - High spreads: above the 80th percentile (500 basis points).
- Subgroup-specific patterns:
  - Common top predictors across groups include outstanding international obligations, international reserves, and sovereign spreads (9 of top 15 common).
  - Low spreads: long-term external debt increases issuance likelihood; outstanding obligations have little effect.
  - Medium/high spreads: issuance probability initially rises then declines with obligations; very high short-term international debt reduces issuance probability.
  - High spreads: long-term external debt negatively associated with issuance; gross international reserves are the most important predictor; exchange rate regime ranks second.
  - High-spread countries more sensitive to governance improvements; timing shifts with high-spread issuers tending to issue later in the year (third and fourth quarters).

### High-spread issuances — sample, patterns, and unconventional features
- Sample for high-spread issuances:
  - Focus on small, frontier EMDEs with nominal GDP below US$300 billion since 2012Q1 at spreads above 437 bps (sample mean in restricted subsample).
  - 84 issuances identified across 16 economies.
  - EMBI spreads in sample range from 437bps to 3206 bps.
  - Data sources for features: Bloomberg, Bond Radar, Cbonds, Perfect Information (LEG), and news articles one week prior and after issuance date (news via Factiva); Bloomberg via Sovereign Debt Metrics.
- Geographic, size, and timing patterns:
  - Most high-spread issuances occur in Latin American countries.
  - On average, these issuances are the largest (gross issuance in percent of GDP) in Latin America.
  - Number of high-spread issuances increased over time; roughly one-third occurred between 2019 and 2022.
  - During 2019–2022 the average size of high-spread issuances was larger compared to earlier years.
- Restructuring and IMF context:
  - Most high-spread issuances occur in the absence of debt restructuring or default, though they tend to be smaller than issuances during restructuring episodes.
  - Countries are approximately five times more likely to have an ongoing IMF program when issuing at high spreads compared to the full sample.
- Prevalence and characteristics of unconventional features:
  - At spreads > 440 bps, 36 percent of issuances include special features such as guarantees, contingencies, or collateral (mostly sinkable bonds).
  - At spreads > 600 bps, the share of unconventional issuances rises to 42 percent.
  - Unconventional issuances are typically larger than conventional issuances at high spreads.
  - Explanations for conventional issuance at high spreads include speculative investor demand, credible policy reforms/institutional improvements, and legal/institutional constraints limiting structuring options.

### Model performance and operational thresholds
- Discrimination:
  - Random forest ROC-AUC = 0.84.
  - Best-performing logit ROC-AUC = 0.77.
- Decision thresholds and operating points:
  - Random forest F1 Score peaks at threshold 0.52; accuracy also peaks around 0.52.
  - Logit F1 Score maximized at threshold 0.51.
  - Tradeoffs: raising threshold increases precision and specificity but reduces sensitivity (recall).
- Predicted probability distributions:
  - Visualization used to assess counts of actual positives and negatives across predicted probability bins and to select thresholds maximizing F1.

### Estimation, training, and robustness checks
- RFE and hyperparameters:
  - Recursive Feature Elimination improved mean ROC-AUC up to a plateau around 50 features; baseline minimum features set to 80 and RFE stopped at 50 when cross-validated mean AUC stabilized.
  - Random Forest hyperparameters: Number of trees = 500; Number of variables per split = 5; Min observations in terminal node = 1; Max tree depth = 15; Resampling method = Bootstrap.
  - Logit hyperparameters: Regularization type α = 0 (Ridge L2); Regularization strength λ = 0.06.
- Feature importance agreement:
  - Top determinants consistent across random forest and logit: GG international outstanding obligations (% of GDP), fiscal balance forecasts (3 years ahead), nominal GDP, exchange rate regime, EMBIG spreads, net debt metrics, international reserves, IMF credit usage.
  - Random forest highlights variables with interactions/nonlinearities (debt composition, reserves), while logit emphasizes categorical and forward-looking indicators.
- Temporal stability:
  - Core variables remain influential across 2002–2007, 2008–2014, 2015–2021.
  - Recent increases in importance of projected debt ratios and EMBIG spreads noted.

### Implications for theory and policy (links to model intuition)
- Empirical evidence supports theoretical prediction that instruments providing lender insurance or commitment (guarantees, collateral, reverse macro-contingent features) can restore market access when traditional bonds would be rationed.
- The higher prevalence and larger size of unconventional issuances at higher spreads are consistent with the model’s mechanism where state-contingent or enhanced instruments facilitate access under elevated risk.
- Instrument choice between conventional and enhanced debt depends on spread levels, investor supply conditions, policy credibility, institutional constraints, and balance-sheet considerations.

*Source: wpiea2026010-source-pdf — IMF Working Paper content (Annexes II–IV and main text).*

### References  ________________________________________________________________________ 30

### wpiea2026010-source-pdf - References  ________________________________________________________________________ 30

### Figures and Annexes (document map)
- Figures listed: 1 through 12, including specific analyses such as "Emerging Market Sovereign Issuances under International Law", "Frequency of Issuance at Different Levels of Spreads", "Random Forest Shapley Feature Importance", "Partial Dependence Plots", "ROC Curve and Actual Positive vs Actual Negative Predictions", "High Spread Issuances", and "High Spread Issuances with Unconventional Features".
- Annexes listed: I. Proofs Omitted in the Main Text; II. Data; III. Estimation and Training Process; IV. Estimation Results.

### Introduction — Context and motivation
- Net international issuances by EMDE sovereigns surged pre-pandemic, producing a threefold increase in the nominal value of outstanding international securities (Figure 1).
- The average cost of debt has remained elevated and increased further since 2019.
- The share of issuances with spreads above 500 bps:
  - reached a record of 20 percent in 2020;
  - has remained above 25 percent thereafter.
- The weighted EMBIG spread, calculated using net issuances, has declined due to high spread issuances focusing on refinancing operations, reducing net issuance volumes for countries with high spreads.
- Implication: constraints to market access may have become more binding recently.

### Theoretical framework — Stylized moral-hazard model of credit rationing
- Model setup:
  - Two-period model (t = 0, 1) with two states θ ∈ {G, B}.
  - Borrower (sovereign) risk-averse with utility u(c1); lender risk-neutral.
  - Project return Rθ with RG > RB; baseline assumption RB = 0 (zero recovery in bad state).
  - Borrower finances investment K̃ with endowment A and debt D at cost (1 + r).
  - Effort ei ∈ {eH, eL}; cost B > 0 for eH and π(eH) > π(eL).
  - Signal si ∈ {H, L} (e.g., credit rating) with probabilities qH, qL such that qH > qL.
- Key constraints and conditions:
  - Net present value condition: π(eH)RH + (1−π(eH))RL − B − K̃ ≥ 0 and π(eL)RH + (1−π(eL))RL − K̃ < 0.
  - Lender participation constraint (PC) equates expected repayments to D (equation (1) in text).
  - Incentive compatibility constraint (ICC) ensures high effort is optimal for the borrower (rearranged ICC shown as equation (1) in text).
  - First-order conditions for equilibrium with respect to rH and rL are provided (equations (2) and (3) in text), showing how rH tightens ICC while rL relaxes it.
- Mechanisms and implications:
  - Credit rationing depends on the cost of borrowing (linked to default likelihood), borrower net worth (A), effort incentives (B), and signal quality (qH, qL).
  - Spreads alone are insufficient to summarize market access.
  - Instruments that provide lender insurance—guarantees, collateral—or reverse macro-contingent instruments (higher rates upon policy slippage) can restore access by serving as commitment devices or mitigating moral hazard.
- Empirical-relevant example cited:
  - El Salvador used a commitment mechanism in April 2024 where the coupon rate would rise from 0.25 percent to 4.0 percent starting in October 2025, unless (i) an IMF arrangement is approved with regular reviews, or (ii) the credit ratings from at least two agencies improve from its initial level.

### Empirical approach — Data, machine learning, and identification limits
- Dataset includes all EMDEs with available EMBIG spreads and a comprehensive set of indicators from prior literature.
- Two empirical strategies:
  - Non-parametric machine learning model: random forest to forecast issuances and identify key factors, capturing non-linear effects and interactions.
  - Traditional logit model for comparison to assess the relevance of non-linear feature interactions.
- Identification caveat:
  - The study does not fully identify supply versus demand shocks; results reflect predictive associations rather than causal relationships.
- Advantage of the approach:
  - The random forest conditions on variables capturing both demand and supply factors and their non-linear interactions, allowing estimation of the effect of a demand factor holding a supply factor constant on issuance likelihood.

### Key empirical findings on high-spread issuances
- Focus: issuances at spreads higher than 440 bps (and alternative threshold at 600 basis points).
- Unconventional features prevalence:
  - At spreads > 440 bps, 36 percent of issuances include special features such as guarantees, contingencies, or collateral.
  - At spreads > 600 basis points, the share of unconventional issuances rises to 42 percent.
- Size and characteristics:
  - Unconventional issuances are typically larger than conventional issuances (as noted in the summary).
- Interpretation:
  - These empirical patterns are consistent with the model’s intuition that instruments providing lender insurance or commitment can enable market access when traditional bonds cannot be sold.

### Paper structure (roadmap)
- Section 2: Literature review.
- Section 3: Theoretical model of credit rationing.
- Section 4: Empirical strategy (random forest).
- Section 5: Overview of empirical results.
- Section 6: Brief discussion of model fit.
- Section 7: Characteristics of high-spread issuances.
- Section 8: Concluding remarks.

*Source: https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026010-source-pdf.pdf*

### 1.    The participation constraint (1)

### 1.    The participation constraint (1)

### Theoretical results: first-order conditions, participation and incentive constraints
- Case: incentive compatibility constraint (ICC) non-binding (μ_ICC = 0).
  - Proposition 1 (Non-binding ICC): If the return of effort is sufficiently high such that the following condition holds:
    - 훥훥휋휋 (푅푅_H − (퐾퐾_� − 퐴퐴) (1 + 푟푟**)) > 퐵퐵 휋휋(푒푒_H),
  - Then borrowers do not choose signal-dependent interest rates and set the unique interest rate equal to the minimum value that satisfies the PC, 푟푟** = 1/휋휋(푒푒_H).
  - Comparative statics highlighted:
    - An increase in borrower net worth A relaxes the ICC by increasing the borrower's stake in the project.
    - An increase in returns (e.g., higher expected growth/higher expected fiscal revenues), a decline in the cost of effort B, or an increase in Δπ (stronger effect of effort on the probability of a good state) all relax the ICC for a given spread.
- Interpretation of balance-sheet and financing variables within the model:
  - Gross financing needs (GFN) and official debt are not explicit model variables but can be interpreted as factors that effectively reduce borrower net worth A.
    - High GFN → akin to lower net worth A (imminent liquidity pressures and repayment obligations).
    - Official debt seniority can effectively lower perceived net worth A (crowding-out) but official credit may improve signal quality (crowding-in).
  - Note: if borrowers have existing debt at the start of period zero, A may assume a negative value.
- Case: ICC binding (μ_ICC > 0).
  - Proposition 2 (Binding ICC): If the collateral constraint is binding, borrowers offer signal-dependent interest rates and pay a higher interest rate after the bad signal, i.e., r_L** > r_H**.
  - Equilibrium condition when probability of default is sufficiently high:
    - 훥훥휋휋 (푅푅_H − 퐾퐾 − 퐴퐴 / 휋휋(푒푒_H)) ≥ 퐵퐵 휋휋(푒푒_H).
  - Economic intuition:
    - Committing to higher interest rates after the bad signal lowers borrower return after the bad signal, making shirking (choosing low effort e_L) less attractive, which loosens ICC.
    - This "reverse"-macro contingent instrument (higher rates in bad states) can allow re-access to markets when credit constrained.
- Model simplifying assumptions noted:
  - Borrowers always default in the bad state of the world, independent of the level of the interest rate.
  - To ensure well-definedness in the bad state, assume u(0) = 0.

### Empirical framework: linking theory to observable indicators
- Purpose: assess relevance of theoretical concepts (borrower net worth A, exerted effort, signal quality) via observable indicators.
- Indicators mapped to theoretical concepts:
  - Balance sheet strength: debt ratios, FX reserves (proxy for A).
  - Expected performance: projections of fiscal balance, growth projections (proxy for returns / effort).
  - IMF program indicators and institutional indicators: proxy for borrower effort and signal quality.

### Empirical methodology: definition of market access and model specification
- Definition of market access:
  - Two broad measures: price-based and quantity-based.
    - Price-based: builds on Stiglitz and Weiss (1981) asymmetric information—higher interest rates reduce borrower stake and can induce credit rationing.
      - Empirical observation: frequency of issuances declines as spreads increase; sharpest decline when spreads are between 300 and 800 basis points.
      - Some countries continue to access markets under relatively high spreads even without debt restructuring/default.
      - Risk-free rates (proxied by U.S. interest rates) have minimal role in explaining issuance decisions in the data shown.
    - Quantity-based: frequency of actual issuances or primary gross issuances.
      - Authors focus on frequency of actual issuances for three reasons:
        1. Some countries issue under relatively high spreads (important for frontier economies).
        2. Allows assessment of balance sheet strength or macro-fiscal outlook in addition to spreads.
        3. Allows assessment of potential non-linear patterns in feature importance for varying spread levels.
- Dependent variable:
  - Binary variable at quarterly frequency = 1 if country issued in either quarter t+3 or t+4, and 0 otherwise.
  - Baseline forecasting horizon: three to four quarters in advance to match real-time applicability given data lags.
  - Class balance: dependent variable equal to one in 58 percent of the observations.
- Modeling approach:
  - Non-parametric machine learning: random forest (Breiman, 2001).
    - Ensemble of decision trees trained on random subsets of data and random subsets of features per split.
    - For classification, final prediction based on majority vote across trees.
  - For comparison: Logit models with three regularization techniques (lasso, ridge, no regularization).
  - Imputation and transformations:
    - Missing values imputed using K-nearest neighbors’ imputation.
    - Variable transformations include percentages of GDP, growth rates, average growth rates, variances of growth rates.
  - Training/testing split:
    - Training set: 2001Q1-2021Q4.
    - Testing set: 2022Q1-2023Q2.
    - Implies a 90/10 percent split between training and testing sample.
  - Cross-validation and hyperparameter tuning:
    - Expanding-window time-split cross-validation with a one-period gap between training and validation to avoid look-ahead bias.
    - Training partitioned into four folds with expanding in-sample training subsets and fixed validation windows covering the four quarters following the training window plus a one-quarter gap.
    - Performance metric: ROC-AUC averaged across four validation sets.
  - Feature selection:
    - Recursive feature elimination (RFE) applied; baseline minimum number of features set to 80.
    - RFE stopped at 50 features because model’s cross-validated mean AUC stabilizes around this point.
  - Final estimation:
    - Model re-estimated on the full training set using selected hyperparameters and reduced feature set; performance evaluated on out-of-sample test data.

### Data description
- Sample:
  - Quarterly data for all EMDEs with available EMBIG spreads from 2000Q1 until 2023Q2.
  - Dataset spans 94 quarters and 46 countries.
- Issuance data:
  - Based on BIS debt security statistics covering sovereign issuances under international law (can be denominated in various foreign currencies, not only USD; excludes local currencies).
- Explanatory variables:
  - 99 base variables (see Annex II for details) including WEO forward-looking variables, IMF IFS, World Bank IDS, Bloomberg.
  - Variables grouped into:
    - Macroeconomic variables: GDP, fiscal balances, public debt (gross and net), real GDP growth, inflation, exchange rates, trade and current account balances, gross financing needs.
    - Global factors: VIX, U.S. interest rates, global growth trends.
    - Primary and secondary market data: gross/net bond issuances, CDS, EMBIG spreads.
- Note on debt measures:
  - Both gross and net public debt considered relevant for emerging market sovereign spreads.

### Empirical results: model outputs and interpretability
- Modeling outputs and interpretability tools:
  - Random forest feature importance assessed using Shapley values.
  - Partial Dependence Plots (PDP) used to show how individual features influence predictions.
- Reporting:
  - Mean absolute Shapley values for the top 15 base features summarized in Figure 5.
  - PDPs for the top 15 features shown in Figures 6 and 7 to illustrate direction and marginal effect of features on predicted issuance probability.
- Notes on interpretation:
  - Shapley values reflect absolute importance but not direction; PDPs complement with directional effects by showing predicted outcome variation with a single feature averaging over other covariates.
  - PDPs smooth out influence of other covariates by averaging, isolating the relationship between the feature and prediction.

*Source: wpiea2026010-source-pdf - 1.    The participation constraint (1).*

### Annex IV also provides a summary of the top features in the Logit model.

### Annex IV also provides a summary of the top features in the Logit model.

### Balance sheet–related features
- Four of the five most important features relate to external debt: total and short-term general government (GG) international outstanding obligations (short-term defined as residual maturity of up to one year), the year-over-year change in GG international obligations (i.e., net issuances), and total short-term external debt.
- Partial dependence plots (PDPs) indicate:
  - Larger stocks of outstanding obligations are associated with a higher predicted probability of issuance.
  - Relationships are nonlinear: effects are strongest at moderate levels of outstanding debt and plateau as debt levels become elevated, suggesting supply-side constraints or limits to market access at high debt burdens.
  - Changes in GG outstanding international securities: countries are most likely to issue following modest increases in outstanding debt, while larger increases reduce issuance likelihood.
- Gross international reserves (another balance sheet feature):
  - Higher reserve levels tend to be associated with less frequent issuance, likely because reserves substitute for external financing.
  - The PDP averages suggest a dominant demand effect (countries with lower reserves need to borrow more), though a supply-side channel may also be present.
- Importance of long-term external debt highlights that debt composition and maturity profile matter for market access.

### Forward-looking features
- WEO projections for the fiscal balance and for public debt are relevant predictors, underscoring expectations about future fiscal trajectories.
- Projections that debt will be high as a share of GDP tend to lower the probability of issuance.
- Projected fiscal balance shows a sharp change around the zero threshold:
  - Countries with projected fiscal balances that are positive three years ahead are significantly more likely to issue than countries with expected deficits, implying fiscal discipline enhances market access.

### External sector–related features
- Trade balance: more positive trade balances are associated with a lower likelihood of issuing on international markets, reflecting reduced need for external financing.
- Exchange rate regime: issuance is less likely for fixed exchange rate regimes, possibly due to limited options to capture FX resources to repay external debt.
- Nominal Effective Exchange Rate (NEER): substantial non-linearity around 0:
  - Currency appreciations are associated with a higher likelihood of issuance.
  - Effect is less clear for depreciations.

### Official debt and IMF financing
- Official creditor debt and IMF credit are important with complex, nonlinear impacts:
  - Higher outstanding amounts of official creditor debt are associated with a lower likelihood of market issuance, with the relationship flattening at higher debt levels (diminishing marginal effects).
  - Probability of issuance increases when official outstanding debt has previously declined; positive changes in official debt correspond to a significantly lower but relatively flat issuance probability.
  - Interest rate on new official credit: lower official rates are associated with a higher likelihood of market issuance.
- PDP for IMF credit is not included among top 25 individual features.
- Patterns reflect both demand- and supply-side dynamics: shifting between official support and market financing and potential crowding-out due to seniority of official debt.

### Structural and institutional factors
- Nominal GDP is among the most important predictors, consistent with larger economies being more likely to access international markets.
- Governance—particularly rule of law—is positively associated with sovereign debt issuance:
  - Relationship is nonlinear, with strongest marginal effects between the 15th and 40th percentiles of the rule of law indicator.
  - Similar nonlinearities for regulatory quality and political stability.
  - Economic magnitudes are relatively modest (small differences in predicted probabilities).

### Stability of feature importance over time
- Feature importance is broadly stable across sub-periods (2002-2007, 2008–2014, and 2015–2021).
- Exceptions:
  - Indicators of IMF financing and exchange rate regimes are more influential in earlier periods.
  - Forward-looking fiscal indicators (e.g., projected debt ratios) gain prominence in more recent years.

### EMBIG spreads and supply vs. demand factors
- EMBIG spreads are a key factor but not the single most important; they interact with other features.
- The PDP shows an inverted U-shaped relationship with issuance probability:
  - Probability increases with EMBIG spread at low levels, peaking around 200 basis points (bps).
  - Beyond 200 bps the probability begins to decline, with the sharpest drop around 600 bps.
- Interpretation:
  - At low spreads, issuance driven by demand (financing needs).
  - At high spreads, issuance constrained by supply (investor willingness).

### Low, medium, and high spread subgroup findings (cut-offs)
- Spread classification:
  - Low spreads: below the 40th percentile (approx. 200 basis points).
  - Medium spreads: between the 40th and 80th percentiles (approx. 200–500 basis points).
  - High spreads: above the 80th percentile (500 basis points).
- Commonalities across spread groups:
  - Outstanding sovereign international obligations (long- and short-term), international reserves, and sovereign spreads are among top predictors in all groups.
  - 9 of the top 15 predictors are common across groups.
- Key differences and subgroup-specific patterns:
  - Outstanding debt and maturity structure:
    - Low spreads: increases in outstanding international obligations have little effect on issuance probability; higher levels of long-term external debt increase issuance likelihood.
    - Medium/high spreads: issuance probability initially rises then declines as obligations increase further; very high levels of short-term international debt securities reduce issuance probability.
    - For medium/high spreads, long-term external debt is negatively associated with issuance probability.
  - Official debt and IMF credit:
    - High-spread countries less likely to issue with larger outstanding official debt (crowding-out).
    - Higher IMF credit outstanding is associated with a greater likelihood to issue in high-spread countries (possible signaling/catalytic effect).
    - Committed but undisbursed official credit appears among top predictors in the high-spread group.
  - Global conditions:
    - High-spread issuances more likely during strong global economic growth (supply-side constraint evidence).
    - Low-spread countries more likely to issue during weaker global growth (counter-cyclical, fiscal-driven issuance).
  - Fiscal projections:
    - For high-spread group, projected fiscal balances are more prominent; PDP still shows strong negative relationship between projected debt levels and issuance likelihood.
  - Reserves:
    - Gross international reserves are the most important predictor for the high-spread group.
    - Issuance probability declines notably as reserves rise for high/medium spreads; relationship is flatter for low-spread countries.
  - Exchange rate regime:
    - More important for medium and high spreads; ranks as second most important predictor for the high-spread group.
  - Governance:
    - Market access is more sensitive to governance indicators in high-spread situations (improvements in control of corruption and rule of law increase issuance probability).
  - Timing:
    - High-spread countries tend to issue later in the year (third and fourth quarters); low/medium spread countries issue earlier in the year.

*Source: IMF Working Paper content (Annex IV summary).*

### 6.  Model Performance

### 6.  Model Performance

### Model discrimination performance
- The random forest model achieves an ROC-AUC of 0.84.
- The best performing logit model achieves an ROC-AUC of 0.77.
- The ROC-AUC metric evaluates discrimination across all classification thresholds, accounting for sensitivity (true positive rate) and specificity (false positive rate).
- A score of 0.5 indicates no discriminative power (equivalent to random guessing), while a score of 1.0 reflects perfect classification.
- The ROC curves of the random forest model and the logit model are plotted in the left panel of Figure 10.

### Predicted probability distributions and threshold tradeoffs
- The right panel of Figure 10 displays the distribution of actual positives and negatives across predicted probabilities:
  - x-axis: model’s predicted probabilities.
  - y-axis: number of observations.
  - Separate bars indicate counts of actual positives and negatives within each probability bin.
- The vertical dashed line marks the threshold that maximizes the F1 score.
- Tradeoffs when adjusting the threshold:
  - A high threshold reduces the number of false positive predictions (improving precision).
  - A high threshold risks missing actual positives (lowering recall).

### High-spread issuances — sample construction and scope
- Sample focus: issuances of small, frontier EMDEs (economies with nominal GDP below US$300 billion) since 2012Q1 at spreads above 437 bps (the sample mean in this restricted subsample).
- Sample statistics:
  - 84 issuances identified.
  - 16 different economies represented.
  - EMBI spreads range from 437bps to 3206 bps.
- Data sources used to identify unconventional features: Bloomberg, Bond Radar, Cbonds, Perfect Information (LEG), and news articles one week prior and after the issuing date (news information based on Factiva; Bloomberg data accessed via the Sovereign Debt Metrics).

### Patterns in high-spread issuances (summary of Figure 11)
- Geographic and size patterns:
  - Most high-spread issuances occur in Latin American countries, regardless of the specific spread threshold used.
  - On average, these issuances are the largest—measured as gross issuance in percent of GDP—in Latin America.
- Time trends:
  - The number of high-spread issuances has increased over time.
  - Roughly one-third of all such cases in the sample occurred between 2019 and 2022.
  - During 2019–2022 the average size of high-spread issuances was larger compared to earlier years.
- Restructuring and program context:
  - Most high-spread issuances take place in the absence of debt restructuring or default, although they tend to be smaller than those issued during restructuring episodes.
  - While the majority occur without an active IMF arrangement, countries are approximately five times more likely to have an ongoing IMF program when issuing at high spreads compared to the full sample.

### Unconventional features and their prevalence (summary of Figure 12)
- A significant share of issuances include special features such as guarantees or collateral:
  - 36 percent of issuances include special features, mostly in the form of sinkable bonds.
  - When increasing the threshold to 600bps (grey bars), the share of unconventional issuances increases to 42 percent.
- Size differences:
  - Unconventional issuances tend to be larger compared to conventional issuances at high spreads.
- Explanations for continued conventional issuance at high spreads:
  - Speculative investor demand or favorable market sentiment may allow borrowers to access markets without additional safeguards.
  - Credible policy reforms or institutional improvements may signal lower default risk, reducing the need for enhancements.
  - Legal or institutional constraints may limit the ability to structure complex instruments, forcing reliance on conventional debt.

### Implications for theory and policy (links to model intuition)
- Findings align with theoretical intuition that unconventional features (contingencies or guarantees) can restore market access under high spreads.
- The empirical evidence that unconventional issuances are more common and larger at higher spreads is consistent with model predictions that state-contingent or enhanced instruments facilitate access when risk is elevated.
- The choice between conventional and enhanced instruments depends on broader market, policy, and institutional conditions, in addition to spread levels.

*Source: wpiea2026010-source-pdf - 6.  Model Performance*

### Annex II.   Data

### Annex II.   Data

### Features included in the RFE process (variable types, frequencies, sources, and transformations)
- Variable categories:
  - Macroeconomic aggregates (e.g., Nominal GDP, GDP current price per capita, Real GDP growth forecasts).
  - External debt and debt-service metrics (e.g., International sovereign debt securities: gross issuance, net issuance, amount outstanding; Total external debt; External debt stock, public and publicly guaranteed).
  - Sovereign risk and market indicators (e.g., EMBIG, J.P.Morgan EMBI Global Index; US 10-year treasury yield; VIX; Credit default swap).
  - Policy and institutional indicators (e.g., Use of IMF credit, Ongoing IMF supported arrangement, Exchange rate regime, Investment grade, S&P, dummies for regions and Low-income country).
  - Fiscal and public debt forecasts and forward indicators (e.g., General government fiscal balance, cyclical adjusted balance, net debt, gross debt and 1-/3-/5-year ahead forecasts).
  - Official creditor flows and loan characteristics (WB IDS items: Amount of Debt Outstanding, Disbursements, Interest Payments, Net Flows, Average time to final repayment, Average interest rate, Degree of concessionality, Overdue payments, Committed but not disbursed).
  - Financial and balance of payments items (e.g., Net international investment position, Official reserve assets, Current account, Exports plus Imports).
  - Financial sector and credit (e.g., Credit to the private sector, percent of GDP; Foreign direct investment).
  - Price and exchange rate indices (e.g., Consumer price index, Nominal effective exchange rate, Real effective exchange rate).
- Frequencies and transforms applied (examples preserved verbatim from source):
  - Quarterly variables frequently transformed as: Level, yoy growth, five-quarter average yoy growth, 5-quarter growth variance.
  - Many debt and flow series also transformed to "% of GDP".
  - Annual governance indices from WB WGI provided as "Index" and "Level, yoy growth, five-quarter average yoy growth, 5-quarter growth variance" for percentile ranks.
  - Dummies and categorical variables: "Dummy", "Categorial".
  - Specific quarterly and annual sources listed include: WEO, BIS Debt Statistics, Bloomberg, WB IDS, IMF IFS, MONA database, IMF AREAER.
- Notable explicit examples (value-preserving):
  - Nominal GDP, USD billion — Quarterly — WEO — Transformations included: Level, yoy growth, five-quarter average yoy growth, 5-quarter growth variance.
  - EMBIG, J.P.Morgan EMBI Global, Index, quarterly average — Quarterly — Bloomberg — Transformations included: Level, yoy growth, five-quarter average yoy growth, 5-quarter growth variance.
  - Use of IMF credit, USD billion — Quarterly — MONA database — "% of GDP".
  - General government fiscal balance, end of period, % of GDP — Annual — WEO — "% of GDP, yoy growth, five-quarter average yoy growth, 5-quarter growth variance".
  - Total population, Billion — Annual — WEO — "Level".
  - Many WB IDS items reported as "USD billions" and "% of GDP" and, for some, "yoy growth, five-quarter average yoy growth, 5-quarter growth variance".
- RFE process design note (from table header): The RFE process considered many lagged/base-variable aggregations (e.g., 3rd lag to 6th lag aggregated for Shapley analyses noted later).

### Countries included in the sample and number of issuances (selected entries from Table 2)
- Bulgaria 59
- Mexico 56
- Poland 53
- Indonesia 52
- Brazil 44
- Lebanon 43
- Türkiye 43
- Colombia 42
- Philippines 42
- Hungary 34
- Egypt 32
- Panama 30
- Kazakhstan 29
- Uruguay 29
- Latvia 28
- Peru 28
- China 27
- Ecuador 25
- Chile 21
- Ghana 21
- Ukraine 21
- Argentina 17
- South Africa 17
- Russia 16
- Pakistan 14
- El Salvador 13
- Czech Republic 12
- Slovak Republic 12
- Jordan 12
- Guatemala 10
- Mexico entries above and remaining countries continue as listed in the source table (full table preserved in source).

### Transformations and normalization conventions used across variables
- Frequent transformation set used across many series: Level; yoy growth; five-quarter average yoy growth; 5-quarter growth variance.
- Fiscal and debt series frequently converted to "% of GDP".
- Governance indices include both "Index" and "percentile rank" representations.
- Dummy variables used for categorical regime/region indicators (e.g., Low-income country, Asia Pacific, Euro area, IMF supported arrangement).

### Data sources cited (as listed in the annex)
- WEO
- BIS Debt Statistics
- Bloomberg
- WB IDS
- IMF IFS
- MONA database
- IMF AREAER
- Asonuma & Trebesch (Period of Restructuring or default, dummy)

---

### Annex III.  Estimation and Training Process

### Recursive Feature Elimination (RFE) insights
- Relationship observed: as number of features decreases, mean ROC-AUC on validation sets increases for both the random forest and the logit model—indicating improved performance from removing less informative or redundant variables.
- Variability in AUC along the RFE curve reflects randomness in training and selection.
- Figure III.1 description preserved: "RFE, Number of Features vs Mean ROC-AUC, Random Forest vs Logit" — mean ROC-AUC plotted across varying numbers of features for random forest (left panel) and logit (right panel).

---

### Annex IV.  Estimation Results

### Best-performing hyperparameters (Table 2 reproduced in content)
- Random Forest:
  - Number of trees: 500
  - Number of variables considered at each split: 5
  - Minimum number of observations in a terminal node: 1
  - Maximum tree depth: 15
  - Resampling method: Bootstrap
- Logit Model:
  - Regularization type (α): 0
  - Regularization strength (λ): 0.06
  - Interpretation: α = 0 implies pure Ridge-type (L2) regularization; λ = 0.06 is the selected penalty strength.

### Decision-threshold behavior and selected operating points (Figure IV.1)
- Random Forest:
  - As the decision threshold increases: precision and specificity increase; sensitivity declines.
  - F1 Score peaks at threshold 0.52 (marked by dashed line).
  - Accuracy also peaks around threshold 0.52.
- Logit model:
  - Similar metric patterns to random forest.
  - F1 Score maximized at threshold 0.51.

### Feature importance and model agreement (Logit Shapley and Random Forest comparisons)
- Top determinants identified consistently across models (preserved terminology from source):
  - General government international outstanding obligations (% of GDP)
  - Fiscal balance forecasts (notably 3 years ahead)
  - Nominal GDP
  - Exchange rate regime
  - EMBIG spreads
  - Net debt metrics
  - International reserves and debt composition (highlighted more by random forest)
  - IMF credit usage and other indicators of prior market access
- Model-structure driven differences:
  - Logit emphasizes high-level categorical and forward-looking indicators (e.g., Euro Area dummy, U.S. 10-year Treasury yield, real GDP growth forecasts).
  - Random forest emphasizes variables with interaction potential or non-linearity (e.g., debt composition, international reserves, IMF credit usage).
- Overall conclusion: strong alignment in top features across both models reinforces robustness—core drivers include previous market access, fiscal space, external conditions, and investor risk perceptions.
- Figure IV.2 note preserved: "Mean absolute shapely values of top-15 predictors in the logit model, aggregated by base variables (3rd lag to 6th lag)."

### Feature importance over time (subsample stability and shifts)
- Sample split into three sub-periods: 2002–2007, 2008–2014, and 2015–2021.
- Stability:
  - Feature importance largely stable across subperiods; core variables remain influential (international reserves, outstanding international obligations, short-term external debt, EMBIG spreads).
- Shifts observed:
  - Exchange rate regime dummy and IMF usage indicator ranked higher in earlier subperiods and dropped in importance in the most recent period.
  - Forward-looking indicators (projected debt ratios over 3- or 5-year horizons) gained prominence in the most recent period.
  - EMBIG spreads increased in relative importance in the latest subsample.
- Interpretation offered in source: shifts may reflect institutional/structural changes after the global financial crisis and evolving investor focus toward long-term sustainability.

*Source: Market Access and High Spread Issuances, Working Paper No. WP/2026/010 (Annex II–IV).*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026010-source-pdf.pdf_