- Laeven-Valencia database covers systemic banking crises from 1970-2017
- Manual extension to 2023 adds US (SVB, Signature, First Republic) and Switzerland (Credit Suisse)
- No standardized crisis database exists for 2018-2022 period
- Crisis severity is not captured (binary classification only)
- 19 countries analyzed - primarily developed economies + major emerging markets
- Many developing countries excluded due to data gaps
- Sector limited to 'P' (Private non-financial) - households and non-financial corps not analyzed separately
- Some countries have incomplete time series (gaps in early years)
- BIS data lag: ~6 month publication delay for credit-to-GDP data
- Real-time implementation would use provisional/estimated data
- Quarterly frequency limits early detection (monthly data not available)
- Lambda = 400,000 is BIS convention, not econometrically derived
- Originally designed for business cycles (lambda=1,600 for quarterly), adapted for credit cycles
- One-sided filter avoids future information but has endpoint bias
- Alternative trend extraction methods (e.g., Hamilton filter) not explored
- Pre-crisis window = 12 quarters (3 years) is researcher choice
- Optimal threshold (4 pp) derived in-sample, may not generalize
- Binary signal (gap >= threshold) ignores gap magnitude
- Country-specific thresholds not explored
- 70/30 chronological split - no temporal cross-validation
- Test set concentrated in recent period (includes GFC)
- No expanding window or walk-forward validation
- Only 4.4% pre-crisis observations (164 out of 3,719)
- Standard accuracy metrics are misleading
- No resampling techniques (SMOTE, undersampling) applied
- Precision is low by design (many false positives)
- 12 features are researcher choices - not exhaustive
- Features: credit_gap, 4 lags, 2 momentum, 2 volatility, mean, std, z-score
- No external variables (interest rates, property prices, debt service ratio)
- Feature selection based on domain knowledge, not automated
- Logistic (AUC=0.694) >> Tree models (AUC~0.5)
- Suggests near-linear decision boundary in feature space
- Complex models may overfit sparse signal
- Hyperparameter tuning was limited for tree models
- No external validation - all results are in-sample or single test set
- 2023 crisis not included in training/testing
- Country-specific performance not reported
- Confidence intervals not computed for all metrics
- Real-time evaluation: Test using only data available at each point in time
- Country-specific models: Compare pooled vs individual country thresholds
- Alternative crisis definitions: Severity-weighted, multi-class (mild/severe)
- Additional EWS indicators: Property prices, debt service ratio, bank leverage
- Out-of-sample testing: Evaluate on 2023 banking crisis events
- Ensemble approaches: Combine signal extraction with ML predictions
- Interpretability: SHAP values for tree models, calibration curves
- Temporal dynamics: Regime-switching models, time-varying thresholds
- Chart PDFs generated at fixed resolution (may pixelate when zoomed)
- Panel data pagination is client-side (full data loaded in browser)
- No automated data refresh (manual pipeline execution required)
- Literature search limited to OpenAlex (no Scopus/WoS integration)