AI Driven Fiscal Risk Assessment in the Eurozone: A Machine Learning Approach to Public Debt Vulnerability

Comments

Response

1.     Lack of Novelty and Conceptual Contribution

As an applied paper, it does not present a particularly compelling or original narrative. The authors apply well-known and, in some cases, outdated models (e.g., basic logistic regression) without a clear justification for their selection. There is no innovation in either methodology or framing, and the paper does not meaningfully advance our understanding of fiscal risk modeling with AI/ML tools.

We agree that logistic regression and SVM are widely used. However, our contribution lies not in the novelty of individual algorithms but in the operational integration of these models into a fiscal surveillance framework grounded in fiscal theory. We have revised the introduction and discussion to emphasize this integration, highlighting the use of SHAP values and fiscal rules to enhance interpretability and policy relevance. We also clarify the added value of comparing traditional and ensemble methods for sovereign risk classification in a harmonized panel.

2.     Model Justification and Parameter Tuning

The paper fails to explain why these specific models were chosen over others, especially more modern or interpretable alternatives. For instance, ensemble models such as LightGBM or interpretable ML techniques like GAMs are not discussed. In particular, the SVM model's poor performance raises red flags. Given the simplicity and cleanliness of the dataset, I suspect the authors did not adequately tune hyperparameters for SVM. Without a thorough search for optimal settings (e.g., kernel, C, γ), the reported underperformance lacks credibility.

We have updated the methodology section to explain the rationale for choosing each model. While logistic regression provides a transparent baseline, XGBoost captures nonlinearity and interactions. SVM was included to test robustness under nonlinear kernels. We now clarify that hyperparameter tuning was performed via grid search for all models, and specific ranges for C, γ (SVM), and boosting iterations (XGBoost) are reported. We acknowledge that LightGBM and GAMs are viable alternatives and have mentioned these in the revised limitations section, inviting future comparisons.

3.     Insufficient Data Transparency

The paper does not clearly report the number of training/test data points, nor does it provide basic descriptive statistics about the dataset (e.g., class distribution, missing data rates, number of stress episodes). Without this context, it is difficult to judge whether the models were trained on a sufficiently large and representative sample. This lack of transparency undermines the credibility of the findings.

We have revised the data section to include descriptive statistics. SMOTE was used to balance classes. The paper also states that the dataset includes 600 observations after cleaning and imputation.

4.     Incorrect Cross-Validation Method for Time Series Data

Perhaps the most serious issue is the misapplication of random 10-fold cross-validation on panel time-series data. This is a well-known methodological flaw, as randomly splitting time-series data can lead to information leakage and artificially inflated performance metrics. A proper time-aware split—such as training on earlier years and testing on later ones—is required to reflect real-world forecasting conditions. This methodological error is particularly problematic given the policy implications of the study.

We acknowledge this important oversight. As you correctly noted, random k-fold CV can lead to information leakage in temporal data. We have replaced it with a rolling-origin cross-validation approach (i.e., training on earlier years, testing on subsequent years) and updated all performance metrics accordingly. This correction led to modest reductions in AUC and F1 scores for XGBoost but improved generalizability. We also cite relevant work (e.g., Petropoulos et al., 2023) that critiques inappropriate cross-validation in temporal ML contexts.

Reviewer 2 comments

Response

1.      The linear equation for the primary balance is specified as: ???? = ?? + ???????−1 + ???? + ???. However, the vector ??? is not defined. I assume it refers to additional macro-fiscal control

variables, but clarification is needed. If it represents newly accumulated debt, this should be explicitly stated (Section 2, Page 3)

We agree that the vector  was not clearly defined. It refers to the set of macro-fiscal control variables included in the empirical specification—specifically, real GDP growth, inflation, and long-term interest rates. We have revised the text to clarify this and avoid ambiguity.

2.      In Table 1, the authors describe ???????_??? as lagged government debt. However, in the logistic regression specification presented in the subsequent methodology section, the debt variable appears to be unlagged, ??????. Given that early results rely on lagged debt, consistency is important. Should this variable not also be ??????−1?

Thank you for pointing out the inconsistency. The variable used in the regression is in fact lagged government debt (govdebt_lag= We have corrected the notation in the methodology section to ensure consistency with Table 1 and the underlying data structure.

3.      The manuscript states: “The predictive function is. The general prediction function is:” This appears to be a typographical or editing oversight. One of these sentences should be removed (Line 197).

We acknowledge this editing error. The redundant sentence has been removed in the revised manuscript.

4.      Table 2 presents result from the logistic regression model, where the authors note that both government balance and lagged government debt are statistically significant predictors. However, the reported coefficient for ???????_??? is zero, which implies no marginal effect on the log-odds of fiscal stress. In that case, is it correct to interpret the logistic regression model as effectively having only one significant explanatory variable—government balance?

4.(Table 2)

We appreciate the reviewer’s close attention. The coefficient for lagged debt was small but statistically significant at the 1% level. However, the presentation rounded it to zero due to formatting. We have updated Table 2 to report coefficients with sufficient precision to reflect their statistical contribution.

5.      In the discussion, the authors state: “The statistically significant coefficients for government balance and lagged debt affirm long-standing theoretical claims about the primacy of fiscal variables in determining sovereign risk.” Given the issue raised in point 4 above, this statement may require revision. If the coefficient for lagged debt is effectively zero, then its contribution to the model's predictive power—and theoretical affirmation—may be overstated. (Line 386)

We have revised the statement to reflect the more nuanced finding. While lagged debt is statistically significant, its marginal impact is smaller compared to government balance. The revised text now states: “The statistically significant coefficient for government balance, and the smaller but still significant effect of lagged debt, support fiscal theory’s emphasis on primary balances and debt accumulation as key drivers of sovereign risk.”

6.      Table 5 compares classification performance across Logistic Regression, XGBoost, and SVM. However, results for the Random Forest model are conspicuously absent, despite its mention earlier as one of the models applied. If results exist, they should be included; otherwise, the omission should be acknowledged and explained. (Table 5)

Thank you for this observation. This was a mistake, because we did not intend to apply the random forest model in this paper.

7.      While the authors aim to model and predict fiscal stress episodes, the manuscript does not provide information on the credit quality of the sovereigns studied. It would be valuable to explore whether the model’s predictions align with sovereign credit ratings from Moody’s, Fitch, or S&P. Since these agencies rate most EU countries, a comparison could provide an external validation of the model's predictive credibility and offer insight into how fiscal stress indicators map onto actual sovereign ratings.

We agree that aligning model outputs with sovereign credit ratings could strengthen the validation. While ratings were not included in the original scope, we have added a short exploratory analysis comparing model-predicted fiscal stress episodes with contemporaneous sovereign credit ratings (S&P). This comparison, now included in the discussion section, suggests a moderate correspondence, particularly in high-stress cases. We flag this as a promising direction for future work.