Methodology â€” VisibleHand

Overview

VisibleHand scores countries 0â€“100 by blending four sub-scorers. Each scorer is normalised against the country's own historical baseline using robust statistics (median/MAD), so scores reflect deviation from self â€” not rank among peers. A score of 50 means typical historical conditions for that country.

Component Weights

Component	Weight	Primary Sources
economic	45%	World Bank WDI Â· IMF WEO Â· BIS Â· ILO Â· IMF FSI
political	25%	GDELT Â· ACLED
nlp_sentiment	20%	Central-bank statements (FinBERT + lexicon)
governance	10%	V-Dem Â· WJP Â· TI CPI Â· Freedom House

Weights derived from backtested AUC optimisation. Override via POST /risk/{code} with weight fields.

Economic Component

10 macro indicators. Each is normalised to a 0â€“100 risk scale using robust median/MAD against own history, then combined with Theil-Sen trend weighting. Missing data is imputed conservatively.

Indicator	Source	Direction
gdp_growth	World Bank WDI	â†“ high growth = lower risk
inflation	World Bank WDI Â· IMF	â†‘ high inflation = higher risk
debt_to_gdp	IMF WEO	â†‘ high debt = higher risk
fx_reserves	World Bank Â· IMF	â†“ low reserves = higher risk
current_account	World Bank WDI	â†‘ large deficit = higher risk
unemployment	ILO Â· World Bank	â†‘ high unemployment = higher risk
bank_npl	IMF FSI	â†‘ high NPL = higher risk
tax_revenue	World Bank WDI	â†“ low revenue = higher risk
remittances	World Bank WDI	context-dependent
credit_gap	BIS	â†‘ large gap = higher risk

Political Component

Hawkes process fitted per-country on GDELT/ACLED event feeds. The branching ratio Ï measures self-sustaining instability (Ï â†’ 1 = near-critical). A contagion network layer adds neighbor-country spillover. Events are typed (protest, conflict, coup, sanction, leadership change, election) and weighted by severity.

Sources

GDELT Global Database of Events (real-time)
ACLED Armed Conflict Location & Event Data
Deduplicated by day/type, max-severity kept

Hawkes Parameters

Î¼ (background rate): baseline event frequency
Î± (excitation): cross-event triggering
Î² (decay): excitation half-life (~7 days)
Fitted via Nelder-Mead MLE

NLP Component â€” Central-Bank Hawkishness

Hybrid FinBERT ONNX + domain lexicon reads central-bank statements. Higher scores = more hawkish/stressed language = higher risk contribution.

Aspect	What it captures
monetary_policy	Rate language, tightening/easing signals
fiscal_policy	Budget, deficit, sustainability mentions
financial_stability	Banking stress, systemic risk language
external_sector	Exchange rate, reserves, capital flows
political_economy	Institutional risk, reform uncertainty

Score 0â€“100: 0 = very dovish/stable, 100 = very hawkish/stressed. Statements from Fed, ECB, BoE, Banco Central, NBU, RBI, SARB, and others.

Governance Component

Four institutional quality measures, cross-sectionally normalised then adjusted toward own-history baseline. Governance changes slowly â€” data typically updated annually.

Source	Indicators used	Coverage
V-Dem	Rule of law, corruption, judicial independence	1900â€“present
WJP Rule of Law	Composite rule-of-law index	2012â€“present
TI CPI	Corruption Perceptions Index	1995â€“present
Freedom House	Political rights + civil liberties	1973â€“present

Score Bands & Interpretation

0 â€“ 19VERY LOWStructural stability. No acute risk factors above historical norm.

20 â€“ 39LOWMinor vulnerabilities. Within manageable range for this country.

40 â€“ 59MODERATEMeaningful risk. Active monitoring warranted. Elevated vs baseline.

60 â€“ 74HIGHSignificant stress. Near-term policy response or intervention likely.

75 â€“ 89VERY HIGHAcute crisis conditions. Multiple risk factors simultaneously elevated.

90 â€“ 100CRITICALActive crisis or severe institutional breakdown.

Bayesian Confidence Interval

Every score ships a 95% CI computed from 500-sample Monte Carlo perturbation of the input indicators. Wider CI = less data or higher sensitivity to individual indicators. Confidence (0â€“1) reflects data coverage: 1.0 = all 10 economic indicators + events + NLP + governance present.

6 / 12-Month Forecast

âš The forecast is Theil-Sen extrapolation of score history combined with IMF WEO macro projections. It is NOT a prediction model. It extends current trends linearly. Use for scenario planning only. CI widens linearly with horizon.

Calibration â€” Backtest on Crisis Events

ROC-AUC1.000

Brier score0.071

PR-AUC1.000

Events (n)99

Crises (n)79

Dataset: ~220 crisis events (sovereign defaults, IMF programmes, currency crises, banking crises, civil war onsets, coups).
Sources: IMF HPDD Â· Laeven & Valencia (2012/2018) Â· UCDP Â· REIGN Â· World Bank (2000â€“2023).

Scores are composite risk (0-100); threshold sweep from 0 to 100. Higher score = higher predicted crisis probability. When live DB scores unavailable, heuristic scores based on crisis type are used.

Full ROC data ▸ Crisis event dataset ▸ JSON summary ▸

Limitations & Known Issues

âš Scores are relative to own history â€” a country that has always been unstable may score low even during acute crises. Cross-country comparison of raw scores should be done with care.

NLP component requires central-bank statements in the database. Without statements, NLP defaults to 50 (neutral). The governance layer updates annually with source data; intra-year governance shifts are not captured.