Common indicators hurt armed conflict prediction

Are big conflicts different from small or medium size conflicts? To answer this question, we leverage fine-grained conflict data, which we map to climate, geography, infrastructure, economics, raw demographics, and demographic composition in Africa. With an unsupervised learning model, we find three overarching conflict types representing major unrest,'' local conflict,’’ and ``sporadic and spillover events.’’ Major unrest predominantly propagates around densely populated areas with well-developed infrastructure and flat, riparian geography. Local conflicts are in regions of median population density, are diverse socio-economically and geographically, and are often confined within country borders. Finally, sporadic and spillover conflicts remain small, often in low population density areas, with little infrastructure and poor economic conditions. The three types stratify into a hierarchy of factors that highlights population, infrastructure, economics, and geography, respectively, as the most discriminative indicators. Specifying conflict type negatively impacts the predictability of conflict intensity such as fatalities, conflict duration, and other measures of conflict size. The competitive effect is a general consequence of weak statistical dependence. Hence, we develop an empirical and bottom-up methodology to identify conflict types, knowledge of which can hurt predictability and cautions us about the limited utility of commonly available indicators.

💡 Research Summary

The paper investigates whether large‑scale armed conflicts differ fundamentally from smaller or medium‑size ones by combining fine‑grained conflict event data from the Armed Conflict & Event Data (ACLED) project with a suite of 22 background indicators covering climate, geography, infrastructure, economics, raw demographics, and demographic composition across Africa. First, the authors discretize space into a pseudo‑random Voronoi lattice (cell size b km) and time into bins of length a days. Using directed transfer entropy between adjacent cells, they construct a temporal‑causal network that captures how conflict activity in one cell predicts activity in another. By linking events that are temporally adjacent and connected through this network, they define “conflict avalanches” – coherent spatio‑temporal clusters of events – yielding 5,659 avalanches from roughly one million ACLED records (1997‑2024).

For each avalanche they compile the 22 indicators, transform each variable into a ternary code (‑1, 0, +1) based on deviations from the median (33 % and 67 % percentiles), and aggregate these codes to form a “bag‑of‑words” vector that counts how many variables fall below, near, or above the median within each of the six indicator categories. Mutual‑information analysis shows strong intra‑category correlations but weak inter‑category dependence, indicating that each category contributes largely independent information.

Applying unsupervised clustering (e.g., K‑means or Gaussian mixture models) to these vectors, the data consistently collapse into three interpretable clusters:

1. Major unrest – concentrated in densely populated, well‑infrastructured, flat riparian zones; associated with the highest fatality counts and longest conflict durations.
2. Local conflict – occurring in regions of median population density with diverse socio‑economic and geographic characteristics; typically confined within national borders.
3. Sporadic and spillover events – found in low‑density, poorly‑infrastructured, economically deprived areas; generally small in size and short‑lived.

The surprising finding is that labeling an avalanche with its cluster type reduces the predictive performance of models estimating conflict intensity (fatalities, duration, etc.). The authors term this a “competitive effect” and attribute it to the weak statistical dependence among the background indicators: each indicator behaves almost like independent noise, so the addition of a categorical label introduces extra uncertainty rather than useful signal.

Consequently, the study concludes that (i) a data‑driven, bottom‑up methodology can uncover meaningful conflict typologies, but (ii) commonly available background indicators have limited utility for forecasting the magnitude of armed conflict. This cautions policymakers and researchers against over‑reliance on such indicators for conflict risk assessment and underscores the need for more dynamic, interaction‑focused modeling approaches.