Sovereign Default Modelling

Problem statement

Sovereign default modelling is highly sensitive to data leakage across countries and time. Evaluation discipline matters more than headline model performance.

Engineering objective

1. Defined country hold-out and temporal train and test splits.
2. Scoped validation-based threshold tuning and train-only resampling.
3. Selected evaluation metrics suited to imbalanced default prediction.
4. Integrated World Bank enrichment with reproducible output paths.

Architecture and approach

A reproducible machine-learning pipeline for sovereign default risk with explicit hold-out evaluation, hybrid modelling, and reporting focused on methodology rather than production credit rating.

• Country hold-out evaluation and temporal split design.
• XGBoost and hybrid model experimentation with F1, recall and PR-AUC reporting.
• Train-only SMOTE and validation-based threshold tuning.
• Reproducible evaluation outputs and enrichment steps.

How it works

Python, XGBoost, scikit-learn ecosystem.

World Bank data enrichment.

Reproducible evaluation scripts and outputs.

Controls and constraints

• Research dataset handling only; not positioned as production credit infrastructure.
• Evaluation splits documented to reduce temporal and geographic leakage.

What the system includes

• Leakage-aware training and evaluation pipeline.
• Model comparison and threshold tuning workflow.
• Reproducible result exports.

Product and system views

A leakage-aware machine-learning pipeline for evaluating sovereign default risk across countries and time.

Evaluation and results

• Demonstrates evaluation discipline for financial ML research.
• Shows how to document methodology and limitations honestly.

Validation and testing

• Hold-out country evaluation and temporal test split.
• Validation-based threshold tuning with train-only resampling.

Limitations

• Not a production credit-rating or investment system.
• Results depend on dataset scope and split design documented in the repository.