KimutaiHub/kenya-political-violence-risk-prediction

Kenya has experienced recurring political violence since 1991. Rich ACLED conflict data and Kenya Census Population & Density Data exists — yet little structured ML work has been applied to predict fatality severity at county level. # Predicting the Severity of Political Violence in Kenya --- A machine learning pipeline predicting the severity of political violence - measured as fatality counts - across Kenya's 47 counties at weekly resolution, using historical conflict data enriched with demographic features. **Team:** Ctrl-Alt-Elite **Contributors:** Kimutai Kevine, Mercy Wambui, Jackline Mwau, Richard Oketch, Winnie Nduva, Reeves Gonah **Date:** June - July 2026 --- ## Elevator Pitch Political violence places communities, election officials, humanitarian organisations, and security planners under significant pressure to make timely decisions with incomplete information. This project demonstrates how historical conflict patterns can be transformed into an early-warning tool that estimates where severe violence is most likely to occur. Rather than attempting to replace expert judgement, the model provides a county-level risk assessment that helps prioritise monitoring, preparedness, and resource allocation before violence escalates. --- ## Overview The project was designed around a practical question: > Given what has happened historically in a county, can we estimate whether violence is likely to become fatal during the coming week, and if so, how severe it may be? Answering this question required combining conflict event records with demographic information, engineering temporal features that capture recent patterns, and evaluating multiple machine learning approaches under realistic forecasting conditions. --- ## Data Sources | Source | Content | Coverage | |---|---|---| | ACLED Africa Aggregated Data | Conflict events and fatalities, filtered to Kenya (16,627 county-week rows) | 1997 – June 2026 | | WorldPop PWD | County-level population, density, area, and population-weighted density | Census snapshots 2000–2020 | WorldPop values are linearly interpolated between census snapshots, backfilled before 2000, and extrapolated after 2020 at 2.2% annual growth (KNBS 2019 inter-censal …