Anticipating conflict-related violence requires understanding when attacks may occur, where risk is concentrated, and which responses merit consideration. Yet open-source information is often delayed, sparse, and inconsistently categorized, limiting analysis of particular armed groups, institutions, and localities. This proposed dissertation investigates how computational methods tailored to these limitations can support forecasting, spatial risk assessment, and expert-guided exploration of potential policy responses.

To study whether we can predict when attacks will occur, I introduce NTEWS++, which combines large language model-based feature extraction with a temporal forecasting architecture shared across armed groups. It forecasts attack types over one- to six-month horizons, improving predictive performance over the earlier NTEWS ensemble while reducing feature-preparation time from weeks to days. Beyond anticipating when attacks may occur, a complementary question is whether we can identify where violence is more likely. I address this question through two spatial studies using established statistical and machine-learning methods. BHACK assesses relative attack risk to individual schools in Nigeria, while PCORE examines pastoral-conflict risk across four Central African countries and multiple geographic scales. Moving from anticipating violence to informing responses raises a further question: can computational analysis help identify ways to mitigate violence? To investigate this question, PAPER generates candidate policy recommendations that a predictive model assigns to the no-attack class, preserving non-negotiable expert constraints and making necessary departures from expert preferences explicit. Their practical feasibility and effects on violence remain questions for expert review and empirical assessment.

Together, these studies aim to strengthen the evidence available to analysts and policymakers while keeping interpretation and final judgment with domain experts.