CAST+ with ACLED
Two researchers from our Research & Software team spent eight weeks extending ACLED's CAST conflict-forecasting model with Algorithmic Governance. The strongest CAST+ setup cut forecasting error by 24.6% and reached 79.5% directional accuracy in months where conflict levels changed.
The project
ACLED's conflict records from 1997 to 2025 were the basis. Building on ACLED's CAST model, the team handled data preparation, feature engineering, model building, ablation tests, evaluation and technical documentation. Algorithmic Governance supplied the project framework and domain expertise.
The method
Random Forest, LightGBM and CatBoost pipelines were compared on six-month holdout forecasts across 2,205 regions, covering battles, remote violence and violence against civilians. The team also tested whether World Bank economic indicators, holidays and religious-composition data improved the forecasts.
The result
The strongest setup cut error by 24.6% against the baseline and reached 79.5% directional accuracy. More outside data did not automatically help. The best result came from careful use of ACLED's own conflict signal, which gives future forecasting research a foundation to reuse.
On stage & involved
- Data partnerConflict-event data and the existing CAST model
- Framework partnerProject framework and domain collaboration
- Research teamData preparation, feature engineering, modelling, ablation testing, evaluation
24.6% less forecasting error
No press release for this one yet. Writing about it? Reach our press desk via the contact page or grab logos and boilerplates in the media kit.