New study identifies systematic patterns in armed conflicts across Africa – and the limits of forecasting conflict intensity, duration, or fatalities
THE STUDY IN A NUTSHELL
- Three conflict archetypes: The researchers combined more than 20 years of conflict data with information on climate, geography, infrastructure, economics, and demographics to classify conflicts in a reproducible, data-driven way. This analysis consistently identified three conflict archetypes.
- Complement to expert labels: This data-driven approach is designed to complement established expert classifications. While expert assessments provide deep contextual and regional knowledge, the data-driven method helps identify broad patterns across tens of thousands of conflict events worldwide.
- Limited predictive power: Improved classification does not necessarily allow reliable predictions of conflict intensity, duration, or fatalities; the relationships between conflict type and outcomes are statistically weak.
- What’s next? If commonly used indicators and datasets cannot reliably predict the severity of conflicts, it raises the question which key factors are missing from global datasets and which new approaches could improve future models – thereby enabling more targeted humanitarian planning and more efficient allocation of resources.
The language used to describe conflicts naturally reflects assumptions about how different forms of violence emerge and develop. “For instance, we think that ‘civil wars’ are the result of internal strife, and we debate whether wars should be characterized as matters of ‘invasion’ or ‘defense.’ In a similar way, experts also label conflicts to indicate important properties and to make patterns across conflicts comparable for use in systematic analysis, early warning, and policy decisions,” explains Niraj Kushwaha, lead author from the Complexity Science Hub (CSH).
Yet, “existing conflict classifications are largely heuristic, meaning they rely on rules of thumb or expert judgment, which may vary and can be difficult to reproduce,” adds CSH’s Eddie Lee.
NEW CONFLICT LABELS
In the new study published in Royal Society Open Science, the researchers used a machine learning algorithm to identify conflict labels directly from data. By combining more than two decades of fine-grained conflict data from the Armed Conflict Location & Event Data Project with information on climate, geography, infrastructure, economics, and demographics across Africa, they aimed to classify conflicts in a reproducible and empirically grounded way.
“These new labels overlap to some degree with expert labels but are not the same,” says Lee. “Experts have incredible intuition about conflicts and regions that they know well, but there are tens of thousands of conflict events worldwide each year. No single expert can comprehend all this complexity. There is an opening here for a quantitative and automated analysis to support global and transferable expertise,” Lee continues.
THREE TYPES OF CONFLICT
The research team, led by CSH scientists in collaboration with colleagues from the University of Waterloo and Princeton University, discovered that the large data set consistently sorts into three archetypes of conflict that have a characteristic combination of geographic, demographic, infrastructural, and economic properties.
“Our algorithmic method learns what conflict types should be by letting the data speak. And the result is surprisingly simple,” explains Lee. Across thousands of African conflicts that took place between 1997 and the present – the team identified:
- Major unrest: Conflicts such as the Boko Haram insurgency or the Central African Republic civil war belong to this category, characterized by sprawling, long-lasting violence in densely populated, well-connected regions that often cross national borders.
- Local conflict: Conflicts such as the Seleka and anti-Balaka conflict belong to this category, characterized by geographically contained disputes within a single country that typically span months rather than years.
- Sporadic / spillover events: Incidents such as the spillover of the Al-Shabaab insurgency into parts of Somalia belong to this category, characterized by short-lived, low-population flashpoints in remote or underdeveloped areas.
“These three conflict types emerged naturally from the data, again and again, even when we changed the spatial and temporal scale of analysis or data coverage,” says Kushwaha.
A DATA-DRIVEN REALITY CHECK
In addition to what the conflict types reveal, the study also emphasizes what they fail to predict. “As researchers, we hope that understanding the type of conflict will improve forecasts about conflict intensity, duration, or fatalities,” Kushwaha notes.
However, the team found that knowing the type of conflict does not improve predictions of its severity – instead it can worsen predictability. Conflict types and measures such as fatalities, duration, or overall impact show only weak statistical relationships. In other words, while conflicts can be grouped into neat archetypes, these categories do not reliably indicate how destructive a conflict will be.
“This seems counterintuitive,” says Kushwaha. “You’d think better classification would help prediction. But the data tells us that these are fundamentally different problems, serving as a crucial reminder about the limits of existing public datasets and the risks of overconfidence in predictive models.”
DIFFERENT CONFLICT TYPES REQUIRE DIFFERENT RESPONSES
“Part of our contribution is to show how to integrate the many types of fine-grained data that may be relevant to how conflicts start, spread, and evolve,” states Woi Sok Oh at the University of Waterloo.
For policymakers, this provides a sharper lens on where different kinds of violence emerge. Conflicts in densely connected cities look very different from those unfolding in remote border regions – and they require different responses, the researchers say. Understanding these conflict types can thus guide smarter humanitarian planning and better use of resources.
For researchers, this work offers a clear way to categorize conflicts based on real-world patterns and data rather than long-standing heuristic expert labels or intuition. “At the same time, it is crucial – especially when it comes to forecasts – to be aware that 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,” notes Lee.
A NEW MAP OF CONFLICT
Avoiding the reinforcement of existing assumptions, the study offers a new perspective on understanding conflicts. If commonly used indicators and datasets do not reliably forecast conflict severity, which factors are missing from global datasets, and what approaches could improve predictive models in the future?
The study suggests that, despite the political, cultural, and historical complexity of armed conflicts, distinct and reproducible conflict types can be identified from the data. At the same time, it highlights that many widely used predictors provide only limited guidance, underscoring the challenges of forecasting conflict intensity with currently available data.
“We are not just classifying conflicts,” says Kushwaha. “We are examining the limits of what can be predicted and, in doing so, hopefully providing a foundation for future research in this area.”
About the study
The study “Data-driven conflict classification exposes weak predictive indicators” by Niraj Kushwaha, Woi Sok Oh, Shlok Shah, and Edward D. Lee has been published in Royal Society Open Science.
