To make a highly interconnected economic system readable without hiding data revisions, benchmarks, small OOS samples or the limits of causal claims.
3 · What it produces
Several complementary views rather than a single number.
The Global Economy Pulse combines cycle and stress through robust statistics; the graph exposes hypothesised channels among macro variables, sectors, commodities, currencies and bottlenecks. Forecasts and scenarios remain separate and auditable.
4 · How it works
From time-t data to a governed output.
1WDI/FRED/OECD and vintages
2Release lag + feature engineering
3Robust z-score + Global Pulse
5GMM regime + graph/scenarios
6Bootstrap, conformal and audit
Targets become available only after their release lag; robust z-scores use prior history. Graph propagation applies sign, elasticity, force, uncertainty/damping and a cap on incoming absolute weights to limit explosive amplification.
5 · Data used
| Source | Information | Use in calculation | Control |
|---|---|---|---|
| World Bank WDI | global targets and annual indicators | release lags, lags/differences/rolling and point-in-time targets | quarantine and anti-leakage audit |
| FRED | commodities, markets, stress and cycle | means, changes, volatility and Pulse components | network retry/budget and missing-data controls |
| OECD CLI | leading cycle signal | gap from 100 and three-month momentum | series availability and quality |
| Structural graph | nodes/edges with sign, lag, elasticity and uncertainty | propagated features and prudential counterfactual scenarios | versioning, provenance and ablation |
Point-in-time and robust
Lags 0/1/2, differences, moving averages, missingness, release lags and rolling median/MAD robust z-scores based on prior history.
Heterogeneous ensemble
Ridge, Elastic Net, Huber, PCA+Ridge, gradient boosting and random forest; robust median ensemble and prequential selection based only on past folds.
Causal three-state GMM
Gaussian Mixture over Pulse/macro groups trained only on prior history; scenario engine kept separate from forecasting and the graph treated as a hypothesis layer.
Benchmarks before complexity
Expanding backtest with purge gap, naive benchmark, bootstrap skill, conformal intervals, OOS R²/correlation and stability diagnostics.
Workflow technical dossier
Methods, validation and implementation stack.
Runtime and packages
- NumPy and pandas for panels, features and audits.
- scikit-learn for regressors/classifiers, PCA, GMM and pipelines.
- NetworkX for the graph; SciPy for statistical methods; Matplotlib for charts.
- SQLite for state/vintages; requests/PyYAML for sources and registries; ReportLab/openpyxl for reports.
Validation
- Prequential selection: each fold selects using errors from earlier folds only.
- Paired bootstrap against benchmarks and 90% conformal intervals.
- Graph ablation and target-information audit for leakage and incremental-value checks.
Governance
- Fail-fast preflight, contract checkpoints and sealed review bundle.
- Research ex-post skill is separated from operational policy.
- A fresh live full run of the current version remains distinct from governance replay validation.
RPA and persistence
- Weekly GitHub Actions, SQLite state restore/persistence and forecast reconciliation.
- Artifacts, digest, report and gates before publication.
Declared limitations
This is an experimental system, not an official projection or financial advice. Graph weights are expert hypotheses until empirically calibrated/stabilised; revisable macro data and structural breaks can change results.
See recent outputs
The Publications page can show multiple images from the same post or preview, with gallery navigation.