CAST methodology

Input and output units

The temporal unit of analysis is a four-week rolling period ending on Fridays, meaning the model is trained on these historical four-week periods to produce the forecast. The final period in the data includes the four consecutive weeks leading up to and including the most recent Friday in the ACLED data. The earlier periods extend back in time to January 2018 relative to this most recent period.

The data has a hierarchical spatial structure. The base spatial unit is the first administrative unit level ("ADMIN1," e.g., provinces). These ADMIN1 units are nested within country units, which are in turn nested within the global level. The model produces predictions at each level, which are then reconciled to ensure consistency across all three levels.

The structure of the target variables — or the types of events being forecasted — is also hierarchical, including three base event types as well as their total. These include the “Battles,” “Explosions/remote violence,” and “Violence against civilians” event types from the ACLED dataset. Their total is directly predicted and referred to as organized violence. The outcome is therefore a count of events per four-week period at a given spatial unit of analysis.

Input features

ACLED CAST relies on a variety of indicators to generate its predictions. Many of the indicators are derived from the ACLED dataset, including trends in recent violence, activity of key actor types, spillover violence across borders, and strategic developments like peace agreements, among others. In addition to the indicators derived from the ACLED dataset, CAST uses a variety of other inputs. For example, population estimates from WorldPop, as well as infant mortality rates from the Center for International Earth Science Information Network, act as proxies for relatively static subnational development, while indicators from V-Dem capture variation in political institutions across countries. A time trend variable is also included to capture unobserved time-varying predictors and is essential for capturing seasonality and, for instance, climate effects that vary throughout the year.

Model

The machine learning algorithm underpinning CAST is a light gradient boosted machine (LGBM) with a Tweedie objective function. As a tree-based algorithm, the LGBM flexibly accounts for the non-linear and interactive relationships among our features. Additionally, LGBM provides a strong balance between flexibility and regularization, enabling it to capture complex non-linear and interaction effects while controlling for overfitting. Compared to other tree-based alternatives, the LGBM is extremely efficient to train, making weekly tuning and updates feasible with minimal overhead. Finally, the Tweedie objective addresses the zero-inflated nature of the target event counts.

Hierarchical reconciliation

For each forecasted period, we generate predictions for every spatial unit and outcome in the data, referred to as “base forecasts.” These base forecasts are generated for each ADMIN1-outcome-period combination, each country-outcome-period combination, and globally for each outcome-period (e.g., 8 November to 5 December 2025 for Brazil-Battles). When it comes to aggregating these forecasts, however, there is no guarantee that the lower-level forecasts will correctly sum to the higher-level forecasts. For example, the ADMIN1 forecasts within a given country might not sum to the direct country-level forecast for that country.

Forecasts that do not correctly sum throughout the hierarchy are not coherent. We therefore apply a hierarchical reconciliation algorithm to enforce the aggregation logic inherent in our data. This ensures that the lower-level spatial units coherently sum to higher levels and the individual event types within each spatial unit coherently sum to the total organized violence forecast. More specifically, we apply MinTraceSparse reconciliation. This method has two key elements: 1) It ensures that aggregation constraints are respected (e.g., lower-level spatial units sum to higher levels; individual event types sum to organized violence); and 2) the adjustments made to the base forecast are minimized according to a weighted loss function.

Uncertainty

In addition to the reconciled forecasts, we also produce estimates of high and low scenarios based on the historical uncertainty for each unit. The reconciled forecast may be thought of as our expected estimate of future events, while the high (low) forecasts represent the estimate of the likely upper (lower) bound of likely events given the uncertain nature of conflict dynamics. Our approach draws on conformal inference but makes a few adjustments to better suit our needs.

As is standard in conformal prediction, we first define the expected coverage level, referred to as alpha. If alpha is 0.05, we expect 95% of observed event counts to fall within a forecast interval. We then execute a rolling time-series cross-validation loop to simulate how a tuned model would have performed historically for each unit in the hierarchy. For each iteration, we generate base forecasts and then apply the hierarchical reconciliation algorithm. We store the residuals (observed minus reconciled predicted values) for each step, which results in a distribution of historical forecast errors for each unit-forecast horizon combination. We then sort the residuals and identify upper and lower thresholds from this distribution that ensure the interval range contains 1-alpha historical errors. These values are then added to the point forecasts to generate upper and lower bounds.

We take a few steps to further calibrate the intervals. In some rare situations, the intervals derived from the distribution of residuals might not contain the point forecast. In this case, we identify all negative (positive) values available and select the median negative (positive) value rather than the quantile implied from alpha. If a given case has no negative (positive) values in its distribution of residuals, we make the uncertainty intervals symmetric. As a final means of calibration, we also smooth the intervals across the forecast horizon to reduce noise in the interval widths over time.

Change from the designated moving average

In the ACLED CAST dashboard, all forecasts are presented alongside the moving average for that country or administrative division. The time horizon of the moving average is customizable by the user. This comparison point is provided in order to contextualize the forecasts within the larger scope of a country or administrative division’s conflict environment. For example, if the 12-month moving average is selected, the dashboard will show the selected month’s forecast alongside the 12-month moving average for that country or administrative division. This shows, on average, how many Battles, Explosions/remote violence, or Violence against civilians events occurred in that area in the last 12 months. The time range for the moving average is customizable to allow for dynamic conflict environments; if, for example, a country has a recent spike in conflict events, the user may wish to set the moving average to just the last one to three months.

Summary

ACLED CAST is a hierarchical forecasting model that leverages machine learning to predict organized violence events at multiple spatial levels. The model incorporates a variety of features derived from ACLED data and other sources and employs hierarchical reconciliation to ensure consistency across predictions at different spatial units. It also quantifies high and low forecast scenarios on historical uncertainty for each unit.

All use of ACLED CAST must abide by the ACLED Terms and Conditions. If you wish to reproduce or republish a visual, graphic, or map from ACLED CAST (rather than creating an original image using raw data) for non-commercial purposes, please cite ACLED CAST using the following format:

ACLED, “ACLED Conflict Alert System,” accessed on DD Month Year. https://acleddata.com/conflict-alert-system/ 

Appendix A: Glossary