Forecasting Crisis Dynamics with Machine Coded Data: A Model of Power, Projection, Influence and Escalation
Principal Investigator: Erik Gartzke, University of California
Years of Award: 2019-2022
Managing Service Agency: Office of Naval Research
Tensions in such disparate settings as missile tests on the Korean Peninsula, Russian incursions in Ukraine, partial state death in Syria and Chinese claims of sovereignty in the South and East China Seas emphasize the challenges posed in seeking to exert influence in world affairs. Which policies or actions are most likely to succeed without being destabilizing or escalatory? What reactions are others likely to implement in response to a given action or policy? How do various political, military, social, economic or historical factors come into play in shaping the duration, intensity or resolution of a crisis?
We propose to produce a prediction model capable of answering these and other related questions and of providing policy makers and researchers with an empirically-grounded theory of foreign policy crisis behavior. We will do so by processing raw data we have collected on the actual actions and reactions of all actors in every interstate crisis in the past one hundred years. Aggregating these data will yield a set of action-response trajectories detailing the evolution of behavior in each crisis. Additional aggregation will generate a “decision tree” populated with probabilities for each action at all nodes in the canonical crisis decision structure. This decision tree constitutes a predictive model of a wide range of behaviors, including the effects of power and restraint, deterrence and escalation strategies and many other crisis dynamics. Different tree structures can be adopted based on aggregation decisions and user priorities. Users will be able to determine what actions and reactions are most likely to occur at any given point in a crisis, based on actual historical data, rather than guesswork or theoretical speculation. The proposed approach will also improve the process of historical inference, assessing how similar a crisis is to other crises and whether current tensions are best interpreted in terms of particular events or assui generis.
Our research pursues an inductive approach, developing insight from an extremely rich dataset of crisis behaviors to be finalized and studied as a core part of this project. To do so, we will take advantage of a variety of “big data” techniques, including machine coding and the use of text as data. The proposed dataset and probabilistic decision tree will contain all relevant behaviors of interest, in sequence and in context, making it possible for the project to infer causal pathways