Early Warning System for Banking Crises
Logistic Regression outperforms complex ML models for crisis prediction (AUC=0.694 vs 0.50-0.56 for tree-based models).
Reason: Signal sparsity (4% pre-crisis observations), near-linear decision boundary, effective feature engineering.
View all CSV data: panel observations, signal extraction results, ML model comparison, feature importance, literature.
Complete documentation of 25 Python scripts: input data, calculations, output charts, data provenance.
Data limitations, methodology caveats, model constraints, and future research directions.
| Source | Dataset | Coverage | Reference |
|---|---|---|---|
| BIS Data Portal | WS_CREDIT_GAP | 40+ countries, 1970-2025 | Drehmann & Tsatsaronis (2014) |
| IMF | Laeven-Valencia Crisis Database | 147 crises, 1970-2017 | Laeven & Valencia (2020) |
| FDIC/FSB | 2023 Banking Crisis | US + Switzerland, 2023 | FSB (2023), BIS BCBS |
| OpenAlex | Academic Literature | 200+ papers | API integration |