Compared to causal modeling, using Machine Learning techniques to predict violent civil conflict is nascent within the broader conflict literature. In this paper, we run three Machine Learning models – OLS Regression, Ridge Regression, and Random Forest – on a dataset of our own making to predict instances of violent intrastate conflict. We find our Random Forest model to be the most accurate, but at the expense of high false negative occurrences, to which we assign greater weight considering their ramifications. Both our OLS and Ridge models outperform Random Forest in false negatives, and we assess our Ridge model to be an optimal compromise of high accuracy and low false negative scores. We conclude this paper with caveats to our models and remarks about the limitations and analytical trappings of Machine Learning applications to conflict prediction.
This project was conducted for "API 222: Machine Learning and Data Analytics", taught at Harvard Kennedy School during the fall of 2020.
This project's work was split between Liz and myself and so different aspects of the code can be found in our respective working files. Each of us handled half of the data cleaning and gathering. We also split the modelling between the two of us.