| Full Path | output/data/panel_data.csv |
| Origin | BIS Data Portal |
| URL | https://data.bis.org/static/bulk/WS_CREDIT_GAP_csv_flat.zip |
| Generated By | pipeline.py -> BISDataFetcher.get_panel_data() |
| Methodology | One-sided HP filter with lambda=400,000 |
| Reference | Drehmann & Tsatsaronis (2014), BIS Quarterly Review |
| Rows | 3,719 |
| Full Path | output/analysis/signal_extraction.csv |
| Origin | Calculated from panel_data.csv + crisis_database.py |
| Generated By | pipeline.py -> CrisisPredictionModel.evaluate_thresholds() |
| Methodology | Binary classification: gap >= threshold signals crisis within 12 quarters |
| Full Path | output/extended_results.json |
| Origin | ML model training on panel_data + crisis labels |
| Generated By | run_extended_analysis.py -> MLCrisisPrediction |
| Methodology | 70/30 chronological split, 12 engineered features |
| Full Path | src/bis_research/data/crisis_database.py |
| Origin | Laeven & Valencia (2020) + 2023 FDIC/FINMA data |
| Coverage | 1970-2023, 149 systemic banking crises |
| Reference | IMF Economic Review, FSB 2023, BIS BCBS 2023 |
- A01: All 19 countries grid
- A02: G7 major economies
- A03: Credit gap heatmap
- A04: HP filter decomposition
- A05: Gap distribution by decade
- A08: Volatility ranking
- B01: ROC curve
- B02: Threshold tradeoff
- B03: Noise-to-signal ratio
- B04: Precision-recall curve
- B05: Confusion matrix
- B06: Lead time sensitivity
- B07: Crisis timing analysis
- C01: AUC comparison
- C02: ROC overlay
- C03: Metrics comparison
- C05-C08: Feature importance (4 models)
- D01: Fixed effects
- D02: Within vs between variation
- E01: Descriptive statistics
- E04: Crisis frequency
- E06: Key statistics infographic
For detailed input/calculation/output documentation with embedded charts, see the full methodology page.