A decade’s worth of conflict data is forcing researchers to confront an uncomfortable truth: the wars we think we understand might defy prediction entirely.
Eddie Lee at Vienna’s Complexity Science Hub describes a problem that has frustrated conflict researchers for decades. “Experts have incredible intuition about conflicts and regions that they know well,” he says. But here’s the catch. There are tens of thousands of conflict events happening worldwide every year. “No single expert can comprehend all this complexity.”
This gap between human judgment and sheer volume is where Niraj Kushwaha and his colleagues saw an opening. Instead of asking experts to sort conflicts into categories, why not let the data do the sorting? They assembled nearly three decades of conflict records from across Africa, added in information about climate, economics, geography, population density, infrastructure, and demographics, then deployed machine learning to see what emerged.
What came back was startlingly simple. Three types of conflict, appearing again and again across different scales of analysis. Major unrest sprawls across densely populated, well-connected regions. Think Boko Haram in Nigeria or the Central African Republic’s civil war. These persist for years and spill across borders. Local conflicts stay contained within single countries, typically lasting months rather than years: the Seleka and anti-Balaka clashes, clan violence in specific corners of Somalia. Then there are the sporadic bursts, brief flares in remote or underdeveloped areas, like the spillover of Al-Shabaab activity into certain parts of the Horn of Africa.
“Our algorithmic method learns what conflict types should be by letting the data speak,” Lee explains. “And the result is surprisingly simple.”
The team documented consistent patterns in how these three archetypes cluster across geographic and economic space. Raw population density and infrastructure development distinguish the major unrest category most clearly. Economics and geography carve out further subdivisions. “These three conflict types emerged naturally from the data, again and again,” Kushwaha notes, “even when we changed the spatial and temporal scale of analysis or data coverage.”
This is the part where everything should come together neatly. Better classification of conflicts should lead to better predictions of their severity, right? Knowing what type of violence you’re facing should help you forecast how many people will die, how long it will last, how much territory it will cover.
It should. But it doesn’t.
When Kushwaha and his team looked at whether their classifications could predict conflict intensity (the number of deaths, duration, geographic spread), they found almost no correlation. Knowing which archetype a conflict belongs to tells you almost nothing about how destructive it will become. “This seems counterintuitive,” Kushwaha says. “You’d think better classification would help prediction. But the data tells us that these are fundamentally different problems.”
Lee frames the discovery more starkly. “Many widely used indicators and datasets may not actually improve our ability to predict how intense conflicts will become, suggesting the need for new approaches rather than more of the same data.”
This isn’t a failure of their algorithm. The research team verified their findings using random forest classifiers (a completely different machine learning approach) and got the same result. The patterns are there. The archetypes are real and reproducible. It’s just that understanding what kind of conflict you have buys you almost no predictive power regarding what damage it will inflict.
The implications ripple outward quietly. Conflict researchers have long assumed that geographical, economic, and demographic factors that shape conflict type would also shape conflict severity. Pour enough resources into better classification systems, the thinking went, and you’d improve early warning. You’d save lives. The data suggests otherwise. Understanding the underlying conditions that create different conflict types does not, in any meaningful way, help forecast their destructive intensity.
Woi Sok Oh, one of the team members at the University of Waterloo, emphasizes what this means practically. “Different kinds of violence emerge in different contexts, and they require different responses.” A major unrest sweeping across urban, densely connected regions demands fundamentally different humanitarian and policy approaches than the localized clan disputes in less developed areas. The classification itself has value.
But predicting impact? That remains elusive. “It is crucial to be aware that many widely used indicators may not actually improve our ability to predict how intense conflicts will become,” Lee cautions. The researchers aren’t saying predictive models are worthless. Random forests still beat random guessing. They’re saying something more unsettling: that the features conflict analysts have focused on for decades might be optimizing for the wrong outcome.
Kushwaha sees this as an opportunity rather than a dead end. “We are examining the limits of what can be predicted and, in doing so, hopefully providing a foundation for future research.” The study reveals as much about absent data as present data. What factors are missing from global datasets? What information could actually improve conflict forecasting? These become the real questions.
On the map, the three conflict archetypes often overlap within the same regions. Around Mogadishu, major unrest from Al-Shabaab operates in close proximity to smaller local conflicts and sporadic events. In the tri-border region where Burundi, Rwanda, and the Democratic Republic of the Congo meet, the same geography contains both large-scale violence that crosses boundaries and smaller conflicts trapped within national borders. This coexistence of different types of violence emerging from similar conditions adds another layer to the mystery.
The researchers have documented their method and dataset publicly, inviting others to build on the work. Their approach integrates multiple types of fine-grained data in ways that could reshape how conflict analysis proceeds. But first, the field might need to recalibrate expectations about what better data alone can achieve.
“We are not just classifying conflicts,” Kushwaha concludes. Perhaps the deeper lesson is that violence, even when sorted into neat archetypes, resists the quantitative prediction we’ve long assumed was possible with enough information. The data has spoken clearly on that point.
If our reporting has informed or inspired you, please consider making a donation. Every contribution, no matter the size, empowers us to continue delivering accurate, engaging, and trustworthy science and medical news. Independent journalism requires time, effort, and resources—your support ensures we can keep uncovering the stories that matter most to you.
Join us in making knowledge accessible and impactful. Thank you for standing with us!