METHODOLOGY GOVERNANCE
Five principles governing every model we build or publish — the sovereign resilience index and the institutional risk engine alike. Each one is stated with the evidence that it holds, because a principle without an artifact behind it is a slogan.
The method is a document with a version number. It changes at planned intervals, not between releases, and prior versions are preserved.
WhyA result you cannot date is a result you cannot check. Versioning turns “the model changed” from a rumour into a record.
Methodology v1.3. Weights, bounds and thresholds are frozen per version; the Core scope and its four pillars are specified in Annex D.
Seven covariance versions archived to date, each labelled with its estimator, window and date. The live matrix is replaced only by explicit activation.
Reference bounds, thresholds and calibrations are fixed for the life of a version, and re-derived only with full back-revision of history.
WhyIf the scale moves whenever the data moves, a score cannot be compared across time. Freezing separates “the world changed” from “we changed the ruler”.
Reference bounds are the 5th/95th percentiles of the pooled 2000–2024 panel, frozen at v1. Classification cuts come from Jenks natural breaks over 413 scored country-years and are version-frozen.
Every calibration is stored as an immutable version with its method, window and parameters. Superseded versions remain available for rollback.
Every published number can be regenerated from stated inputs by someone outside the firm, using published formulas.
WhyReproducibility is the only claim in risk that can actually be tested. Everything else is assertion.
The 2025Q4 ranking was computed twice, independently, from the same official sources — matching to the decimal on all 21 scored sovereigns. Each release ships an audit file naming the source, series code, vintage and confidence tier of every input.
Model changes are recorded in an audit log with actor and timestamp. Implementation validation — re-computing the engine's outputs from an independent implementation — is the documented next step.
The tests a model is subjected to are published with their results, including the ones that disagree with us.
WhySelecting which diagnostics to show is the most effective way to mislead without stating anything false. The remedy is to publish them all.
Publication is blocked unless the ranking survives PCA re-weighting (ρ ≥ 0.90) and a ±5pp weight perturbation (ρ ≥ 0.95). The first release failed the gate and was withheld pending review. Entropy (ρ 0.808) and equal-pillar (ρ 0.688) rankings disagree with the published weights and are printed on the face of the index.
Backtests that fail are kept in the version archive rather than deleted. The risk screen shows the active version, its backtest verdict, and warns when a verdict predates the matrix it is attached to.
What a model cannot see is stated on the artifact itself, not buried in an appendix or discovered by the reader.
WhyEvery model has a boundary. Naming it is what allows someone to use the model correctly; hiding it guarantees eventual misuse.
Core carries no external-balance or market-pricing pillar, so sovereigns with exceptional external assets are understated — stated on the poster. Argentina is withheld for 2017–20 because the IMF published no CPI after censuring Argentine statistics. Singapore is withheld on insufficient fiscal source coverage.
Value-at-Risk is parametric Gaussian and understates tail events by construction. Factor attribution awaits re-derivation of betas against revised factor definitions; the engine is validated for portfolio-level risk, not yet for attribution.