| Sumario: | Predicting the emergence of armed conflicts has long been a subject of debate in the social sciences. Numerous studies have shown that structural factors such as state capacity, economic fragility, and institutional inequality influence the risk of conflict. However, the predictive performance of these indicators is often not systematically tested. This study examines the ability of structural fragility indicators to distinguish periods of high-death-toll conflict within a prospective classification framework. The analysis combines structural indicators included in the Fragile States Index with death tolls derived from ACLED event data. The dependent variable is defined as a binary indicator of whether a country experiences at least 25 conflict-related deaths in the following year. To prevent temporal information leakage and obtain a more realistic assessment, a rolling prediction design with an expanding training window was employed instead of the random data-splitting method. The Light Gradient Boosting Machine (LightGBM) algorithm, which has high capacity to capture nonlinear relationships, was preferred for classification. The findings indicate that structural fragility indicators exhibit strong discriminatory performance in identifying country-year observations that exceed the high-fatality threshold. The average ROC-AUC value in the rolling evaluation results is approximately 0.92. However, the prediction performance is not entirely constant over time. Limited weakening in both discriminatory power and probability calibration occurs in some periods. This indicates that structural indicators convey a significant signal of conflict risk, but predictive success is not independent of the temporal context. The study contributes to the conflict prediction literature by demonstrating that variables with strong explanatory power do not necessarily yield stable predictive performance over time.
|
|---|