How Decision Trees Can Forecast Future Conflicts?

Imagine a forest… not of trees, but of decision trees. PRIO researchers used these trees to estimate not only whether violent conflict will happen, but also how many casualties it might cause. Instead of a single forecast, the model produces a probability distribution: a kind of "weather report" for conflicts.

Uncertainty comes from two sources: the unpredictable nature of conflicts and gaps in the available data. To handle this, researchers combined various tree‑based classifiers and distributional regressors in a custom auto‑ML system. Each PRIO‑GRID‑month cell (a monthly spatial grid) receives its own probability distribution, giving a clearer picture of local risks.

The results show these models outperform traditional benchmarks up to a year ahead, especially in regions where conflicts have already been observed. Adding smaller models didn’t hurt performance, opening the door to include data with limited spatial coverage.

In practice, turning point predictions into distributions helps policymakers, NGOs, and governments plan more targeted interventions, accounting for the high uncertainty that characterises violent conflicts.

The proposed method relies on ensembles of decision trees, specifically Random Forest and Gradient Boosting, integrated with distributional regressors such as Gradient Boosted Distributional Regression. These models not only predict the binary outcome (conflict yes/no) but also estimate the probability distribution of the response variable, typically the monthly casualty count per geographic cell.

To handle the extreme zero‑inflation of the data—most cells record no casualties—a two‑stage approach was adopted: first a binary classifier to detect conflict presence, then a conditional regression model for casualty counts, using Poisson or Negative Binomial distributions with parameters estimated via maximum likelihood.

The custom auto‑ML automated hyper‑parameter selection, spatial cross‑validation, and the combination of regional models into a spatial ensemble. Integrating smaller regional models, trained on local data, showed that while they do not improve aggregate metrics, they preserve predictive robustness in areas with limited data coverage.

Evaluation metrics include the Continuous Ranked Probability Score (CRPS) and Brier Score, which are more suitable for assessing probabilistic forecasts than point‑error measures. Results indicate that the predictive distribution strategy reduces marginal uncertainty compared to point models, providing a more solid foundation for risk mitigation decisions.