The footprint of organised violence held steady through the 2000s and widened from 2012 with the wars in Syria, Iraq and Yemen, then with the Sahel insurgencies, Mexico’s cartel wars, Ethiopia, Ukraine and Sudan. In 2024 and 2025 about 2,000 cells saw at least one event, twice as many as in 2010.
Severity is measured without counting the dead: the share of months in the window with organised violence. The model gives every cell a probability for each of four levels: none, sporadic (under 10% of months), recurrent (10–50%) and sustained (more than half). Switch between the chance of any violence and the chance of sustained violence; hover a cell for all its numbers.
People facing sustained violence is each cell’s population times its probability of sustained violence, summed by country. It weighs where violence is likely to stay against how many people live there.
Afghanistan and Syria lead because violence there has been sustained for years in populated valleys and cities. Pakistan, Ukraine, India, Somalia and Mexico follow. Sudan ranks lower: its current war is intense but recent, and sustained here means more than half of all months in the decade.
Of cells that saw no violence for a decade, 4.5% had some in the following ten years in the back-tests. The model ranks those cells about four times better than chance. For 2026–2035 it flags – quiet cells at 20% or more, mostly on the edges of today’s wars and in densely populated places.
India tops the list: large, densely populated quiet areas lie next to its insurgencies in the centre-east, the northeast and Kashmir. Congo, Iran, Sudan, Ethiopia and Chad follow, where violence of recent years borders provinces that have been quiet.
Use the “Quiet for 10 years” switch on the map to see these cells.
For each test year the models were trained only on windows that had already closed, then scored on what followed. The benchmark assumes each cell stays at the level it had over the previous years. Precision is the area under the precision-recall curve; a random ranking scores the base rate.
Which inputs the boosted-tree model uses to place a cell on the four levels. One year ahead, recent violence is almost everything. Ten years ahead, population, terrain and material value carry more than a third of the signal.
Each row is one test of the materialist account on the grid, with cell and country-year fixed effects and errors clustered by country. A dot to the right of the band supports the hypothesis at 5%; to the left contradicts it.
| Hypothesis | Result |
|---|
The regions with annotated maps in the project, ranked by people facing sustained violence in the decade.
| Region | Cells ≥ 50% | Quiet, 20%+ | People, sustained | Most affected |
|---|
| Claim | Holds |
|---|---|
| Forecast beats persistence at 1, 5 and 10 years | yes |
| Quiet cells ranked ~4× better than chance | yes |
| Theory layers improve the forecast | barely |
| Foreign capital nearby → violence | yes, p 0.03 |
| Resource value → violence | no |
| Great powers avoid direct war | yes |
| Limit | Effect |
|---|---|
| Ownership is a snapshot from about 2020 | kept out of forecasts |
| Asset data dense in Africa only | capital tests weaker elsewhere |
| UCDP misses militias coded under governments | Wagner invisible |
| 20-year horizon untestable | data start 1989 |
Outcome. UCDP Georeferenced Event Dataset 26.1 and Candidate events to September 2026: events of state-based, non-state and one-sided violence, each placed in a 0.5° PRIO-GRID cell. UCDP only records conflicts that pass a threshold of lethality, so violence that never turns lethal stays invisible here.
Inputs. Conflict history of the cell, its neighbours, a 300 km box and its country; GHS-POP population since 1975; WorldClim elevation and ruggedness; satellite-mapped mining area (Maus et al. 2022); PRIO PETRODATA oil and gas fields; IMF PortWatch chokepoint traffic; World Bank commodity prices; GeoEPR ethnic exclusion; SIPRI arms transfers and UN voting distance for rival suppliers; AidData Chinese official finance; UCDP sides for foreign troops. Foreign ownership of about 6,900 mines, fields and plants from Jasansky et al. 2023 and USGS compilations.
Forecasts. Boosted trees (LightGBM) over four levels of how sustained violence is in the window: none, sporadic (under 10% of months), recurrent (10–50%), sustained (over half). No death counts enter the outcome or the reporting. Uninhabited cells that never saw violence are left out and forecast as none.
Limits. The forecast describes where violence is likely given the past, not where it should be prevented first, and not who will fight. Country borders are today’s.