We’re thrilled to announce the release of the new fatalities003 conflict prediction model – running on an entirely new production pipeline.

What’s New?

Building on its predecessor (fatalities002), the new model retains the same sub-model structure, with each sub-model corresponding to a counterpart in the previous version. However, several key changes have been made:

• Updates to the set of predictors/indicators that informs the model to reflect changes in data availability and update cycles from source providers. Indicators that are no longer maintained – or are updated either too infrequently or irregularly to meaningfully improve model performance – have been removed. This streamlines the system, reduces noise, and lowers maintenance costs.
• Updated implementation of sub-models, with adjustments to e.g. hyperparameters.
• New ensembling method at the country level: the genetic algorithm used in fatalities002 has been replaced with a simpler unweighted average of sub-model forecasts, consistent with the approach already used for the subnational models. This change has been implemented in preparation for the upcoming launch of a new model governance protocol, which will allow us to systematically evaluate, deploy, and retire sub-models on a monthly basis, drawing on the most recent performance of the models.

Forecasts from fatalities003 are now available via the VIEWS API and in the data download section of this website. Its output mirrors that of fatalities002 – with the exception of the former surrogate models, which will be discontinued when fatalities002 is formally retired in December 2025 (the last forecasts from the fatalities002 model were released last week, along with the first forecasts from the fatalities003 model).

What the Model Forecasts

The fatalities003 model produces monthly forecasts for:

• The number of fatalities from state-based armed conflict (in both logged and non-logged form),
• The probability of observing at least 25 battle-related deaths from state-based armed conflict per country-month, and
• The probability of at least one battle-related death from state-based armed conflict per PRIO-GRID cell and month.

Output variable names and API endpoints remain unchanged. The data dashboard, however, is currently limited to presenting forecasts from fatalities001 and fatalities002 only. It will be upgraded to support the new model output early next year, if not sooner.

Learn More

To learn more about the fatalities003 model, please visit the GitHub repository views-models and navigate to the musically tributed ensembles named “pink_ponyclub” (country-level ensemble) and “skinny_love” (subnational ensemble). The latest run was executed from the development branch.

To consult the model specifications of the fatalities002 model, visit the GitHub repository “viewsforecasting“.

To explore the new model pipeline, visit the GitHub repository views-platform-core.