Methodology
How the index is computed
Three data channels feed the index itself: the World Bank (stability, economy), Uppsala University's UCDP (armed conflict) and the GDELT news background. Two more sources, OFAC sanctions and INFORM disaster risk, are live as separate layers, deliberately not folded into the score. The layers are computed and shown separately; the verdict only looks at where they point together.
What's live right now
The live HRI is computed from armed-conflict events of the last ~12 months (UCDP/GED, Uppsala University, CC BY 4.0), government stability (World Bank WGI) and inflation. Conflict intensity follows UCDP convention plus a party-to-war supplement. NPI (news) is live from GDELT for most countries. Disasters (INFORM), governance depth (WGI) and sanctions (OFAC) are live too, but shown as SEPARATE layers. We do not fold them into HRI, because those weights are not backtested. Travel advisories are still pending. We mark this honestly rather than implying full coverage.
Real risks
A weighted sum of indicators from open data. The most reliable part of the score. Component weights (live ones highlighted):
What's in the news
The mood of news about the country: how tense or calm the background is relative to usual. About perception in the media, not the probability of an event.
Scenario session: the decision for your specific case
Custom ResearchNot part of the base index. In the Custom Research package, Arxterra together with the School of Prognostic Modeling runs a facilitated scenario session for your specific case: a group works through your market-entry scenarios at each horizon, maps team readiness and blind spots, and the result is delivered as a SWOT-style protocol alongside the data brief.
Final verdict
We don't merge layers into one number. We look at where the data layers point together:
Risky: hard risk is high (HRI >= 56) Calm: hard risk is low (HRI <= 33) Moderate: mid-range risk, no clear direction
Five horizons
Every country is assessed at five horizons: today, 3 months, 1 year, 3 years and 10 years. The central estimate does not move between them, because our backtest showed that assuming no change beats extrapolating a trend (see Track record below). What the horizon changes is the calibrated uncertainty band and the historical probability that the verdict category changes.
The remainder up to 100% is uncertainty; a scenario session (Custom Research) helps close it for your specific case.
Track record
We backtested our own horizon model with a rolling-origin design (no look-ahead): at each historical year we projected forward using only data available then, and compared with what actually happened. Here is what we found, including the part that went against us.
| Horizon | Observations | MAE, no-change | Verdict accuracy | Old model, skill |
|---|---|---|---|---|
| 1 year | 4,428 | 2.066 | 95.5% | -1.403 |
| 3 years | 4,016 | 2.971 | 93.7% | -2.388 |
What the backtest killed
Our earlier model extrapolated a trend into each horizon. It was about twice as bad as simply assuming nothing changes, and its skill score against the base rate was strongly NEGATIVE: −1.40 at one year and −2.39 at three. It was not merely useless; it actively degraded the verdict. So we removed it. The central estimate at every horizon is now today's level, the honest champion of that test. (The winner's own skill reads 0.000 by construction, since it is being compared with itself. The number worth looking at is the old model's.)
What the horizon does change is honest: a calibrated uncertainty band derived from measured error quantiles, and the historical probability that a market changes verdict category.
Probability a market changes category in a year
| Today | Stays | Worsens | Improves | Cases |
|---|---|---|---|---|
| Calm | 98.0% | 2.0% | 0.0% | 3,095 |
| Moderate | 88.9% | 3.6% | 7.5% | 962 |
| Risky | 91.9% | 0.0% | 8.1% | 408 |
Base rates from the same historical series, automatically calibrated, and the number we lead the paid brief with. Backtest limitation: historical series cover the stability and inflation components; the conflict component is held at its current value, because multi-year UCDP history isn't in the pipeline yet. Raw output: /data/backtest.json.
Caveats
Arxterra is for information only and is not investment, legal, tax or any other advice, an offer or a recommendation. Scores are computed from the listed sources; we do not guarantee their accuracy or completeness. You make decisions yourself. Conflict data: UCDP/GED, Uppsala University (CC BY 4.0); governance & economy: World Bank; news: The GDELT Project.
Data sources & attribution
- Armed conflict: UCDP/GED, Uppsala University (CC BY 4.0). Davies, S., Pettersson, T. & Öberg, M. UCDP Georeferenced Event Dataset. Data are aggregated and rescaled by Arxterra.
- Governance & economy: World Bank Open Data: Worldwide Governance Indicators, WDI (CC BY 4.0). Modified by Arxterra.
- News tone & headlines: The GDELT Project.
- Sanctions: OFAC Specially Designated Nationals list, US Department of the Treasury (public domain, US Government work). Used for the sanctions overlay and the comprehensive-embargo verdict cap.
- Governance depth (rule of law, regulatory quality, control of corruption): World Bank Worldwide Governance Indicators (CC BY 4.0). Shown separately, not folded into HRI.
- Map geometry: Natural Earth (public domain). Scenario sessions: School of Prognostic Modeling, by agreement.