WATCHTOWER / MODEL METHODS

Global Risk Overview

Version global-risk-overview-v1 · Country and regional evidence-pressure indices, source context, and the existing Systemic Risk forecast record in one workspace.

What the audit found

The previous Risk Overview assembled 18 shared adapters into 13 anchors, 8 domains and 3 horizons: 312 cells. The current normalized catalogue assigns risk-tab points to other workspaces too; the live Sources & unique-data audit lists every point and its destinations.

Distinctive capabilities were the regional matrix, local watchlists/playbooks/briefing archive, and Systemic Risk’s typed inputs, explicit forecast definitions, frozen evidence and outcome history. These remain accessible. The original matrix is now an explicitly labelled legacy workflow; saved records are not retroactively re-scored.

The old matrix assigned baseline severity 28–37 without evidence, gave unrelated regions weight 0.22, propagated global signals into every region, and increased confidence for repeated contributors. The new map removes those mechanisms and preserves original source dates. These are verified methodological corrections; they are not a measured percentage improvement in predictive accuracy.

Twenty domains, qualified country geography

All19distinct domains from the old matrix and Systemic Risk remain: military, political, crime/terrorism, finance, funding, banking, markets, sovereign, corporate, recession, outage, energy, cyber, supply, shipping, health, food/water, displacement/borders, and space. Natural hazards is an additional20thdomain.

The250-jurisdiction reference directory does not imply full data for250countries. Source-native economy identifiers or original point coordinates inside the existing Natural Earth reference polygons establish geography for eligible measurements. Text and named-place matches establish reporting subject context, not verified incident location. Publisher headquarters and feed-country fallback pins are excluded. International fund references and unlocated observations stay global; US fund and macro references stay US-scoped.

Regions use explicit directory subregions/continents or disclosed operational-country memberships. Their scores are equal-country averages over eligible member-country values, with coverage and90thpercentile hotspot level. They are not GDP/population/exposure weights, causal spillovers, impact polygons or statements about disputed sovereignty.

Source admission and timestamps

The model reviews the public normalized point catalogue and fixed shared Watchtower country, incident, hazard, provider-status, health, macro and Systemic Risk adapters. Registered raw price measures, qualified fiscal observations, original source notices and relevant dated reporting are considered. Original source models remain inspectable in the embedded Systemic workbench.

Private/licensed owner inputs, demographic affiliation/ethnicity, source-health/inventory counts, ordinary filing volume, authored scenario probabilities, synthetic market baselines and unregistered numeric leaves are excluded from scoring. Public expenditure, taxes and CBDC adoption are not treated as inherently adverse. The country CDS fallback marked “baseline” cannot enter the new index.

Observation/publication/period and source capture remain separate. Future, unknown, stale, expired or resolved inputs are withheld. Annual fiscal/crime fields have at most3calendar years of lag; current-year government debt estimates retain their estimate/projection and scope/consolidation qualifiers and a lower source weight. Price-derived observations have a5-day maximum age to accommodate reference sessions; dated macro periods have a150-day bound. Public source captures must satisfy their separate freshness limits.

1/7/30days selects reporting/evidence lookback, not a forecast horizon. Native structural, financial and active-notice windows remain explicitly different. A fresh HTTP response or recomputation cannot refresh an old observation.

Transparent transformations and aggregation

Raw observations use bounded piecewise-linear transforms. Each driver displays its input unit, period, curve, quality weight, source URL, dependency group and lineage. The complete knot registry is available inside Scoring methods and the read-only API.

Example	Selected raw input → index anchors	Qualification
Government gross debt/GDP	40→10;80→35;120→60;180→85;250→100	Structural burden context; currency, assets, maturity and institutions also matter.
Interest/revenue	5→10;10→30;20→65;35→100	Dated published government ratio.
Monthly price loss	3→15;7→35;15→70;30→100	Reference-fund price change, not default probability.
Earthquake magnitude	4.5→20;5.5→45;6.5→70;7.5→90	Physical magnitude; no inferred damage/exposure.
Provider impact	none→0;minor→25;major→65;critical→90	Active source-reported incident, named service scope only.

These initial research knots are explicit assumptions, not empirically calibrated probabilities. Missing observations are never replaced with a default score. Real numeric zero remains valid.

News uses a disclosed event-language classifier on short original source metadata. Generic topic coverage, explicit opinion/history, scheduled/speculative/disputed wording and undated records are withheld. Peaceful protest mentions are labelled activity context, not inferred violence or illegitimacy. Multilingual coverage and automatic country/topic associations remain incomplete.

Exact canonical-source and title/day duplicates are collapsed. Publisher-balanced means prevent high-output outlets from gaining weight through volume. Reporting pressure has a2-day half-life and a60/100cap. Within the observation channel, related measurements share a dependency group. When both channels are present, observations carry75%and reporting25%; otherwise the available channel is disclosed.

The composite balances five families: security/governance, finance/economy, essential systems, health/food/displacement, environment/space. Duplicate lineage is counted once within each family and discounted if shared across families. At least two families are required for a composite. Missing families are omitted and coverage shown, so countries with different evidence coverage are not directly equivalent. Domain changes are reported only across matching available domain sets; changing coverage is separately flagged.

Updates, snapshots and limits

Visible authenticated pages read ordinary shared-source snapshots approximately once per minute. The service coalesces source reads for120seconds and model snapshots for60seconds, uses at most4concurrent fixed adapters, and never requests source-force refreshes. Existing collectors keep their schedules and access pauses. Inputs are bounded to12000projections; omissions and source failures are visible. Exports use one coherent model snapshot; changed native snapshots return409and require refresh. No new forecast issuance, private watch store or source-state migration is added.

Map changes are session comparisons, not a reconstructed country-score historical series. Original legacy history retains its old model meaning. The new composite has no backfilled validation outcomes.

Validation and the Systemic Risk merge

Qualification covers missing/zero/future/stale/expired/private records, country isolation, deduplication, publisher-volume resistance, monotonic transforms, original clocks, source changes, region arithmetic, coherent pagination, real public captures and account/replay clearing. Synthetic tests are distinct from observed outcomes.

The embedded Systemic Risk workbench retains current conditions, outlooks, impact/dependencies, evidence, historical tests, prospective track record, exact inputs and methods. Its58fund-price targets, original forecast ledger and existing Forecast Lab view remain unchanged. Current-published-vintage historical reconstruction is distinct from prospective skill; minimum sample counts alone never establish accuracy. Regional weather targets concern sampled warning messages, not all-world physical hazards.

The new country composite’s predictive accuracy is not established. Establishing it requires explicit country outcomes, timestamped labels, point-in-time source history, temporal holdouts and baseline comparisons. Existing fund-price forecast evaluation is not repurposed as evidence of country-composite accuracy.

Original Systemic Risk methods · Current model contract · Current source inventory

Attribution

Country references: mledoze/countries, ODbL1.0. Reference boundaries: Natural Earth, public domain. IMF/OECD/Eurostat/World Bank/WHO/USGS/FRED and publisher/provider records retain their original links, periods and qualifications. Short reporting metadata is linked to its original publisher; it is not a redistribution of full articles.