Forecasting Armed Conflicts Using Machine Learning
Format
SOECS Senior Project
Faculty Mentor Name
Michael Canniff
Faculty Mentor Department
Computer Science
Abstract/Artist Statement
There is a belief that war is too random and sporadic to ever really be predicted, I would like to believe it is patternable and something we can forecast. Research in this field is often made difficult due to the world within which it operates: it is hard to create empirical studies to understand what causes war. There are too many variables, with too many variations, and they are hard to test. However, with enough resources and time, I believe machine learning can become a way for us to create forecasts and furthermore evaluate what we already know about the causes of war. I create a program which uses time-series ML models and indicators to create armed conflict forecasts. The implementation of this project started with doing the research to understand the features, the independent variables, that I was interested in; I then tested the correlation and casual behaviors using statistical tests. I used different validation techniques, for example, a walk-forward cross validation and a k-fold CV test method to account for overfitting and to better understand the data.
This project exposed me to two things in particular: machine learning classification algorithms and data preparation. While my forecasts show relative accuracy, I think there is room for improvement. For one, I think a better algorithm implementation would lead to more complicated analysis, RNN for one, but most importantly, I think feature selection and enrichment is the key to further enhancing accuracy. If we can continue testing forecasting, we can get closer to a point where algorithms have the data and training to accurately predict armed conflict. This will not only bring us closer to a world where we can proactively respond to violence but understand the causes enough to work towards broader and sustaining peace.
Location
School of Engineering & Computer Science
Start Date
7-5-2022 2:30 PM
End Date
7-5-2022 4:00 PM
Forecasting Armed Conflicts Using Machine Learning
School of Engineering & Computer Science
There is a belief that war is too random and sporadic to ever really be predicted, I would like to believe it is patternable and something we can forecast. Research in this field is often made difficult due to the world within which it operates: it is hard to create empirical studies to understand what causes war. There are too many variables, with too many variations, and they are hard to test. However, with enough resources and time, I believe machine learning can become a way for us to create forecasts and furthermore evaluate what we already know about the causes of war. I create a program which uses time-series ML models and indicators to create armed conflict forecasts. The implementation of this project started with doing the research to understand the features, the independent variables, that I was interested in; I then tested the correlation and casual behaviors using statistical tests. I used different validation techniques, for example, a walk-forward cross validation and a k-fold CV test method to account for overfitting and to better understand the data.
This project exposed me to two things in particular: machine learning classification algorithms and data preparation. While my forecasts show relative accuracy, I think there is room for improvement. For one, I think a better algorithm implementation would lead to more complicated analysis, RNN for one, but most importantly, I think feature selection and enrichment is the key to further enhancing accuracy. If we can continue testing forecasting, we can get closer to a point where algorithms have the data and training to accurately predict armed conflict. This will not only bring us closer to a world where we can proactively respond to violence but understand the causes enough to work towards broader and sustaining peace.