How a Deep Neural Network Contributes to Learning Causal Graph and Forecasting Political Dynamics

Seo Eun Yang (Ohio State University)

Abstract: Nonlinearity has been considerably interested in time series analysis of conflict/opinion dynamics. However, handling unknown nonlinear interactions on time series data is a methodologically challenging task because traditional models such as VAR Granger analysis or B-SVAR models are based on certain pre-assumed linear systems. The difficulty of approximating nonlinear complexity on temporal data can result in poor understanding of the underlying causal relationship between variables and thereby miss the opportunity to get better predictions of political events. In this paper, I examine how deep learning can contribute to learning causal graph and forecasting political dynamics given nonlinearity on time series data. Here, I introduce deep neural networks called Scalable Causal Graph Learning which does not rely on pre-assumed kernel or distribution to detect multiple nonlinear causality (SCGL, Xu et al., 2019). Using SCGL, I reanalyze two sets of time series data previously estimated by B-SVAR models and newly discover nonlinear granger causal relationships between variables: First, I examine the new causal linkages between government evaluations, economic policies, and macroeconomic outcomes using monthly economic and political data for the United Kingdom from 1984 to 2006 (Sattler, Freeman, and Brandt, 2008). Second, I investigate how international public support and online behavior of conflict participants affected the 2012 Gaza Conflicts using the 179 hours of the conflict dynamics (Zeitzoff, 2018). Through modeling nonlinear causality in time series data, I confirm that SCGL model better predict the future periods than those B-SVAR models in two examples.

View Poster in a New Tab