The scale of the phenomenon the researchers confront is striking. The database shows an average of roughly 95 default episodes per country over the six-decade window, with a heavy concentration in low- and middle-income economies. Sudan tops the list with 342 recorded events, followed by Liberia with 310, the Democratic Republic of Congo with 273, the Republic of Congo with 264, and Somalia with 242. Far from being a series of isolated shocks, sovereign distress emerges as a structurally persistent feature of the international financial system: in 68 percent of the years studied, the total number of default events ranged between 155 and 438. As of 2023, the global stock of sovereign debt in default stood at 523 billion US dollars, about 0.5 percent of world public debt, involving 92 countries.

Why does the distinction between creditor types matter so much? The authors argue that defaulting on the International Monetary Fund, on Paris Club bilateral lenders, on Chinese official creditors, on private bondholders, or on domestic agents implies markedly different adjustment paths, contagion risks, and recovery dynamics. A default to private external creditors, which occurred in 19.28 percent of country-year observations in the sample, unfolds through market-based restructuring and reputational costs. Defaults to multilateral institutions such as the Inter-American Development Bank, by contrast, are extraordinarily rare at just 0.15 percent of observations, precisely because such lending is often reserved for countries already in crisis and is rarely repudiated. Treating all of these outcomes as one undifferentiated event, the study contends, discards most of the economically meaningful information.

Methodologically, the framework follows the CRISP-DM data-mining standard and combines the Bank of Canada–Bank of England default indicators with the World Bank’s World Development Indicators, a harmonized panel spanning macroeconomic, financial, institutional, social, and external dimensions. The researchers deliberately kept preprocessing minimal: missing values were imputed using within-country, within-variable means to preserve cross-country heterogeneity, heavy-tailed indicators were scaled robustly, and no discretionary feature engineering was performed before training. Because default incidence varies enormously across creditor categories, class imbalance emerged as a first-order problem. The team tested no balancing, random oversampling, SMOTE, and ADASYN using a fixed Random Forest as a benchmark learner, and found that balancing roughly doubled or tripled the F1-score for moderately imbalanced categories. SMOTE, which synthesizes new minority-class observations in feature space rather than duplicating existing ones, was adopted as the baseline.

To convert the original multilabel structure into a single multiclass outcome, the researchers applied a rarity-prioritization rule: when several defaults co-occur in the same country-year, the observation is assigned to the creditor category with the lowest unconditional default frequency. This transparent, data-driven tie-breaker prevents scarce but institutionally salient categories, such as defaults to multilateral lenders, from being absorbed by more common ones. The dataset was then split temporally, with 1960 through 2019 used for training and 2020 through 2021 held out as an out-of-time test set, a design that provides a more honest estimate of how the model generalizes under shifting economic conditions. Feature selection proceeded in two stages, filtering near-constant variables and then ranking the remainder by mutual information, retaining 1,211 of 1,777 predictors that cumulatively captured 90 percent of the available information.

The modeling results contain a cautionary tale for the AI community. A fully connected neural network, whose architecture was tuned through the Hyperband multi-fidelity optimization algorithm, minimized its validation loss during training yet collapsed almost completely on the out-of-time test set, achieving an overall accuracy of just 0.0021 with near-zero precision, recall, and F1-scores across nearly all classes. The optimized network had effectively learned a degenerate mapping between features and labels. In sharp contrast, an XGBoost gradient-boosted ensemble, which builds its decision function from shallow trees trained sequentially to correct residual errors, delivered an overall accuracy of 0.7742, a macro-averaged F1-score of 0.539, and a weighted F1 of 0.807. The narrow gap between macro precision and recall suggests the model distributes its predictive capacity fairly evenly across creditor-specific outcomes rather than overpredicting a dominant class.

Interpretability was central to the design. Using SHAP, or Shapley Additive Explanations, the researchers decomposed every prediction into feature-level contributions, revealing two headline patterns. First, macroeconomic and external-sector variables dominate: inflation measures such as the GDP deflator and consumer price index, trade indicators, international reserves, and net financial flows consistently rank among the most influential predictors. Second, and perhaps more surprising, feature importance is highly creditor-specific. Across the thirteen classes, the top-ten predictor lists include 105 distinct variables with limited overlap, meaning the model learns genuinely different decision rules for different creditor types. Higher inflation generally shifts probability toward riskier outcomes for market-sensitive creditor classes, while lower reserves and weaker net flows raise risk through external-liquidity constraints and rollover pressures.

Some of the theoretically significant signals are unexpected. The single most important positive contributors in the global model are basic infrastructure access measures, specifically the share of the population with access to electricity, total and urban. The authors interpret this as evidence that structural development gaps and state capacity operate as proximate, high-magnitude predictors of default, refining classical intertemporal solvency frameworks that emphasize primary balances and discounting. Governance indicators, including Government Effectiveness and Voice and Accountability, enter with sizable negative contributions, meaning stronger institutions are associated with lower predicted default probability at magnitudes comparable to the top macroeconomic variables. Social-vulnerability measures, such as the share of female contributing family workers and child-nutrition indicators, also rank highly, consistent with political-economy accounts in which a weak social contract and constrained tax capacity undermine the surpluses needed for debt sustainability.

The practical implications are concrete. Reserve adequacy emerges as an actionable margin: total reserves relative to external debt and reserves excluding gold carry negative signed SHAP values, associating higher reserves with lower predicted risk. The authors suggest that creditors and rating agencies may benefit from weighting institutional measures more heavily in short-horizon risk models, and that the creditor-specific heterogeneity of the SHAP rankings motivates tailored early-warning dashboards rather than a single undifferentiated sovereign-risk score. The team even sketches an operational path, proposing that the serialized model be deployed as an API-based microservice with SHAP explanations delivered alongside risk scores for auditability. The full pipeline is publicly available on GitHub, with pinned random seeds allowing anyone to replicate, audit, or extend the analysis.

The researchers are careful about limits. SHAP values capture conditional associations within the fitted model, not causal effects, so the framework is best deployed for surveillance and triage rather than counterfactual evaluation. Results also depend on the imbalance-correction strategy and on the rarity-prioritization rule used to construct the multiclass outcome, and several influential variables act as proxies for latent constructs whose measurement quality varies across countries. The authors call for follow-up work that converts the strongest SHAP-identified predictors into quasi-experimental or instrumental-variable designs, pairing machine-learning discovery with econometric identification. Even with those caveats, the study marks a meaningful shift in how sovereign distress can be anticipated: not as a single alarm bell, but as a differentiated, creditor-specific risk process whose warning signs span inflation, reserves, governance, and the texture of development itself.

Subject of Research: Machine learning prediction of sovereign default events by creditor type

Article Title: A machine learning perspective on sovereign default events

Article References: Avendaño, M. S. A., González, J. H. S., Pabón, E. O., & Arévalo, N. A. N. (2026). A machine learning perspective on sovereign default events. Discover Artificial Intelligence, 6(1), Article 1401. https://doi.org/10.1007/s44163-026-01721-2

Image Credits: AI Generated

DOI: 10.1007/s44163-026-01721-2

Keywords: sovereign default, machine learning, XGBoost, SHAP, class imbalance, SMOTE, sovereign debt, early warning systems, interpretability, World Development Indicators, Bank of Canada, Bank of England