12.4 Measuring geopolitical risk: the GPR index

12.4.1Construction of the index#

Construction (Caldara and Iacoviello, 2022, American Economic Review 112(4):1194–1225 [V]; methodology page matteoiacoviello.com/gpr_files/gpr_methodology.htm [V]):

1. Corpus. The recent index uses about ten major English-language newspapers, from 1985. Secondary descriptions give the count as 10 or 11, and the full list could not be confirmed this session. The historical index uses 3 newspapers from 1900.
2. Dictionary. An article counts if it contains terms in one of eight categories¹:
   • categories 1–5 are threats: war threats, peace threats, military build-ups, nuclear threats, terror threats;
   • categories 6–8 are acts: beginning of war, escalation of war, terror acts.
3. Index. For month (or day) $t$,
   $$GPR_t = 100\times\frac{s_t}{\bar s_{1985-2019}},\qquad s_t=\frac{\#\{\text{articles in }t\text{ matching the dictionary}\}}{\#\{\text{articles in }t\}}. \tag{12.14}$$
   The recent index is rebased so that its 1985–2019 average equals 100 [V]. The historical index is rebased over 1900–2019 [V].
4. Sub-indices. GPRT (threats) uses only categories 1–5 and GPRA (acts) only 6–8. Country-specific indices count articles mentioning a given country (matteoiacoviello.com/gpr_country.htm).

12.4.2Findings and the AI-GPR index#

Findings From the AER abstract [V]: "Higher geopolitical risk foreshadows lower investment and employment and is associated with higher disaster probability and larger downside risks." Also: "The adverse consequences of the GPR index are driven by both the threat and the realization of adverse geopolitical events." At industry and firm level, investment falls more in sectors more exposed to geopolitical risk. An earlier 2018 working-paper version said the effects were "mostly driven by the threat". The published version says both. Cite the published version.

AI-GPR (Iacoviello and Tong, 2026) "The AI-GPR Index: Measuring Geopolitical Risk using Artificial Intelligence" (Federal Reserve Board working paper, posted March 2026²).

• Articles that pass a keyword filter are scored for geopolitical-risk intensity on a 0–1 scale by a large language model (GPT-4o-mini), instead of counted.
• The corpus is about 5 million articles from three US newspapers (NYT, Washington Post, Chicago Tribune), daily from 1960, normalised to mean 100 over 1985–2019.
• The authors report that AI-GPR has a stronger negative association with stock returns than the keyword index, and use it to identify GPR-driven oil-supply disruptions³.

12.4.3GPR in a returns regression#

Using GPR in a returns regression A natural specification is

with $z_t$ controls such as VIX changes, oil returns and US yield changes. Pitfalls:

1. GPR is persistent and spiky. Use changes, or the innovation from an AR model, not levels.
2. GPR is built from newspapers published after the trading day, so align dates carefully: GPR on day $t$ reflects news through day $t-1$'s editions.
3. Reverse causality is weak (markets do not cause wars) but not zero: market crashes get geopolitical coverage.
4. The index is US- and English-language-centric, which matters for South Asian events. The country GPR for India and Pakistan is the better regressor for May 2025.

M10's gold lab and this module's lab (Problem 12.14) test whether threats move gold and Brent more than acts.

The Economic Policy Uncertainty index (Baker, Bloom and Davis; policyuncertainty.com) is the analogous newspaper measure for policy uncertainty. The April 2025 tariff shock is an EPU event more than a GPR event.