Worum geht es

Ein Forschungsmodell schätzt für 189 Länder die Wahrscheinlichkeit bewaffneter Konflikte über ein bis zwölf Monate, trainiert auf offenen Daten (UCDP, Weltbank, V-Dem, GDELT). Klicken Sie ein Land an, um Horizonte, Verlauf und Profil zu sehen. Achten Sie auf den Unterschied zwischen hohem AUC und tatsächlich kalibrierten Wahrscheinlichkeiten.

What this is about

A research model estimates the probability of armed conflict for 189 countries over one to twelve months, trained on open data (UCDP, World Bank, V-Dem, GDELT). Click a country to see horizons, history and profile. Note the difference between a high AUC and genuinely calibrated probabilities.

Where does the next conflict
erupt? A model estimates it — 95% AUC.

A machine-learning early-warning system for armed conflict — LightGBM trained on 65 years of UCDP data, World Bank indicators, V-Dem democracy indices and GDELT news event counts. Mean out-of-sample AUC 0.95 across all horizons.

–

Countries monitored

–

Countries in active conflict

–

Conflicts worldwide (UCDP)

–

Mean AUC · out-of-sample

12-month conflict risk worldwide · click a country for details

low

high

Most intense active conflicts · trailing 12 months

Outbreak risk · currently peaceful countries · 12 months

State-based conflicts worldwide · since 1990

Battle-related deaths worldwide · since 1990

Forecast horizons

Horizon	P(active)	P(onset, cal.)

Model risk history · P(active / onset within 12m)

What this is: a research prototype built to study how far open data and gradient-boosted trees get you in conflict early warning — and how to calibrate the probabilities honestly. What it is not: a live system or policy advice. The page shows a frozen model state (see “forecast as of”); nothing is recomputed online, no personal data is processed. Country-level structural variables are shared by all cells of a country; the model cannot foresee de-novo conflicts in regions without any recorded history. Onset probabilities for the 6- and 12-month horizons are isotonic-calibrated.

Data: UCDP/PRIO · World Bank · V-Dem · GDELT · Berkeley Earth Model: LightGBM · walk-forward CV · isotonic calibration · ai·maturity Lab

Where does the next conflicterupt? A model estimates it — 95% AUC.