Research & Methodology
Rigorous methodology is the foundation of credible analysis. Tools that cannot be explained should not inform policy.
Research Philosophy
Explainability First
Why it matters: Policymakers must understand why an indicator signals risk, not just that it does. Trust requires transparency.
- SHAP (SHapley Additive exPlanations) for feature attribution
- Partial dependence plots for relationship visualization
- Decision path analysis for individual predictions
- Plain-language summaries with technical outputs
Production Quality
Why it matters: Academic prototypes are insufficient for policy applications. Models must perform under stress.
- Continuous integration/continuous deployment (CI/CD)
- Comprehensive test suites (unit, integration, backtesting)
- Version-controlled code and data
- Secure deployment for sensitive applications
Open Science
Why it matters: Science advances through scrutiny and replication. Transparency enables error correction.
- Full methodology documentation for all public tools
- Reproducible code in version-controlled repositories
- Model cards documenting assumptions and limitations
- External methodology audits
Policy Relevance
Why it matters: Technical sophistication means nothing if outputs don't inform decisions. Our goal is impact, not publications.
- Indicators designed around policy-relevant questions
- Visualization and communication as core deliverables
- Regular engagement with policy practitioners
- Feedback loops from users to methodology development
Methodological Framework
Data Infrastructure
Sources
| Category | Sources |
|---|---|
| Macroeconomic | IMF, World Bank, national statistical offices, central banks |
| Financial Markets | Bloomberg, Refinitiv, central bank publications |
| Banking Sector | Regulatory filings, central bank reports, BIS statistics |
| High-Frequency | Daily/weekly market data, real-time spreads |
Data Quality
- Automated validation checks for outliers and anomalies
- Revision tracking and impact monitoring
- Documented imputation strategies
- Full lineage tracking from source to indicator
Early-Warning Methodology
Crisis Dating
- Systematic review of historical episodes
- Multiple indicator thresholds
- Harmonized dating across countries
- Sensitivity analysis to dating choices
Feature Engineering
- Temporal: Levels, changes, acceleration, trends
- Cross-sectional: Peer comparisons, global factors
- Interaction: Cross-domain linkages
Model Selection
- Gradient boosting (XGBoost, LightGBM)
- Regularized regression (Elastic Net)
- Model averaging for robustness
Gap-Aware Cross-Validation
Standard cross-validation fails with financial time series. Our approach:
Challenges Addressed
- Temporal dependence in data
- Crisis clustering effects
- Missing data handling
- Real-time vs. revised data differences
Our Solution
- Strictly temporal train/test splits
- Embargo periods between training and testing
- Gap-aware imputation with uncertainty
- Real-time data reconstruction
Interpretability Stack
SHAP-Based Explanations
Feature contributions to individual predictions, global importance rankings, interaction effects, and consistency with game-theoretic fairness principles.
Model Cards
Standardized documentation covering model details, intended use, limitations, performance metrics, training data, ethical considerations, and usage recommendations.
Datasheets
Comprehensive data documentation covering motivation, composition, collection methods, preprocessing, distribution policies, and maintenance schedules.
Model Governance
Version Control
- Semantic versioning (major.minor.patch)
- Change logs documenting all modifications
- Ability to reproduce any historical version
- Clear deprecation policies
Monitoring
- Performance tracking over time
- Drift detection for input distributions
- Automated alert systems
- Clear retraining triggers
Audit Trail
- Training data snapshots
- Hyperparameter configurations
- Validation results
- Post-deployment performance
Research Outputs
Working Papers
Methodological innovations, empirical findings, validation studies, and policy applications.
Methodology White Papers
Detailed technical documentation of our EWS, liquidity indicators, and interpretability approaches.
Data Documentation
Comprehensive documentation of sources, construction, quality controls, and update schedules.
External Engagement
Methodology Audits
We commission periodic external reviews: independent assessment, replication of results, recommendations for improvement, and published audit summaries.
Academic Collaboration
We engage with the academic community to incorporate advances, subject our work to peer scrutiny, contribute to open-source tools, and participate in conferences.
Standards Development
We contribute to emerging standards for AI/ML documentation, financial stability indicators, open data practices, and responsible AI in policy.