A quantitative model of geopolitical risk

*A research plan for an AI system that scores which countries a crisis actually affects, with the evidence behind every number.*

**Summary:** The best-known geopolitical risk index counts alarming words in newspapers. This plan goes further: small models extract sourced claims, specialist agents judge them by domain, and a multiplicative scoring layer weighs which countries depend on the systems under strain. It is a plan and has not been tested yet.

**Status:** Research plan. **Year:** 2026. **Tags:** Risk modelling, Agent systems, Geopolitics

---

The idea The widely used index from Caldara and Iacoviello counts words like war and terrorism in newspapers. It made the subject measurable, but it can’t say who is exposed. A conflict near a shipping lane matters most to the countries that rely on that lane.

The model ```math \mathrm{risk}(c,a,h) = \underbrace{\mathrm{exposure}(c,a)}_{\text{who depends}}\cdot\underbrace{\mathrm{stress}(a,h)}_{\text{under strain}}\cdot\underbrace{\mathrm{event}(e)}_{\text{what happened}}\cdot\underbrace{\mathrm{prop}(c)}_{\text{network}}\cdot\underbrace{\mathrm{conf}(\mathcal E)}_{\text{evidence}} ``` Country c, load-bearing system a (shipping lanes, energy, food, credit), horizon h. The product is large only when a country depends on a system, the system is under stress and the evidence is strong.

Small, cheap models pull sourced claims out of news in any language. Specialist agents judge those claims within one area, such as energy, food or shipping. Each factor is to be calibrated against historical event panels, and every score keeps a trail anyone can audit.