Locals as Conflict Forecasters: A Study in Cabo Delgado

• In Event: Measurement of Conflict Processes

Abstract

Since the advent of big data, conflict researchers have applied machine learning techniques to build early warning systems with limited results. While these tools successfully identify violent hotspots (Bazzi et al., 2021), early researcher optimism for prediction models based on these big data to detect where but also when violent clashes occur have disappointed. We take a seemingly simplistic approach that has not yet been tested in the conflict literature: to employ locals–instead of conflict experts–as conflict forecasters. We do so under the theory that locals experiencing conflict are the best to predict future attacks, both because local residents exist in the same social and information networks as locally recruited rebels and because they are privy to other locally relevant causal factors such as a soldiers pay day, village celebrations, and the impact of attack timing, among other factors. We test local conflict forecasting accuracy in the context of Cabo Delgado, Mozambique where, since 2017, a violent insurgency group has been attacking villages in northern Mozambique and battling the Mozambique Defence Armed Forces and, since 2021, additional forces from the Southern African Development Community (SADC), and the Rwandan Armed Forces. For over 7 months in 2022-2023, we surveyed 3,382 residents from Cabo Delgado, Mozambique. Half of our respondents were randomly selected to be conflict forecasters, and predicted the likelihood of an attack in their community in the next two weeks, and the other half reported on attacks occurring in the past two weeks. After measuring conflict forecast accuracy, using geographic distance from a conflict report and number of conflict reports to weight accuracy estimates, we find that predictions based on local forecasts far surpassed our baseline estimates that used prior local conflict rates from both past ACLED conflict data and our reporter sample documented conflict. We then review features of accurate predictions such as village location, attack type (e.g., beheading, kidnapping), and the demographic background of members in their reported communication and trust networks and information environments.

Authors

• Carolyn Steinle, UCLA

• Myrl G Marmarelis, USC Information Sciences Institute

• Fred Morstatter, USC Information Sciences Institute