Towards Forecastingting the Probability of Unarmed Revolutionary Destabilization Using Machine Learning Methods

Abstract

The authors provide a broad overview of the main applications for machine learning methods in political sociology. They describe history of the transition from simple regression models to complex machine learning models. The reasons for and benefits of this transition are discussed. The authors identify the main uses of machine learning models in related disciplines and describe how they have been applied to the task of predicting revolutionary episodes. A cohort of other researchers who have tackled the issue of predicting political instability in their own ways, from using multiple regression models to using machine learning as a classifier for tweets during the Arab Spring, is reviewed. An extended description of the main trends in the field of studying predictor behavior in machine learning models is given. Cases of their application and the limitations researchers may face are discussed. The authors describe different statistical approaches to the task of estimating parameters of machine learning models. Using the example of analyzing models built to predict the probability of revolutionary episodes, they discuss ways of ranking model parameters through the estimation of decision trees and changes in the resulting power of models. The authors show how correlated variables can influence the obtained ranking result, why variables can appear in different parts of the ranking under different systems of calculating their importance. The authors also consider the method of determining the boundary after which the model parameters can be considered statistically significant. The authors provide a method of generalized representation of the direction of association of different variables, taking into account their interaction with other predictors, and give an interpretation of the results obtained using Shepley vectors. Among the substantive results of the tests, it is especially worth noting the identification of an exceptionally powerful effect of revolutionary waves in revolutionary events of the 21st century, given that in the 21st century the effect of global revolutionary waves turns out to be stronger than the effect of regional waves. In general, the tests suggest that the following are particularly strong factors that significantly increase the probability of the onset of unarmed revolutionary uprisings in the 21st century (in addition to the effect of revolutionary waves): a high level of political corruption, the effect of inertia (unarmed revolutionary or powerful protest events in the recent past), anomalies in economic growth, high amounts of aid from the United States (the effect of the “iron cage of liberalism” according to Daniel Ritter), the absence of oil rent, high population, high food inflation, middle-income economy, long incumbent duration and an intermediate type of political regime.